Source: https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/part-2-five-failure-categories/ai-technology/ Author: Phil Hatch Publisher: Akholi AI Technology The 2026 Report · Part 2 # AI Technology Reliability, memory, cost, information protection, authorized action, and observability By Phil Hatch · Akholi · Version 1.8.2 Complete study · PDFPart 2 · PDFStudy overview / Part 2 / AI Technology AI technology was rated markedly higher than the three categories covering the two firms. Its 20 indicators had a mean of 4.60 across the 150 engagements, about one point higher than financial institution readiness, co-management, and vendor capability. Distressed ratings made up 5,113 of their 15,911 answers (32.14%), against about half in those categories. The rating remained under the midpoint of the scale. The indicators of the category were associated with cancellation as closely as any in the study, from r = −0.69 to r = −0.63 (n = 150 engagements each). Table 2.6. The indicators of AI technology, from the lowest mean rating | Indicator | Mean rating | Distressed answers | Distressed share | | --- | --- | --- | --- | | 1.1 Production Readiness Within Limits Managed by Hand | 4.54 | 257 of 802 | 32.04% | | 5.2 Handling of Sensitive Data in Outputs, Logs, and Connected Systems | 4.54 | 259 of 800 | 32.38% | | 2.4 End-to-End Reliability on Multi-Step Tasks | 4.55 | 254 of 782 | 32.48% | | 2.2 Accuracy of Citations and References Against Source | 4.55 | 265 of 791 | 33.50% | | 6.1 Agent Action Beyond Authorized Scope | 4.56 | 253 of 789 | 32.07% | | 7.1 Testability of Agent Behavior Before Go-Live | 4.56 | 259 of 792 | 32.70% | | 5.3 Resistance to Hostile Instructions Embedded in Content | 4.56 | 261 of 812 | 32.14% | | 3.3 Integrity of Persistent Memory Against Contamination | 4.57 | 257 of 806 | 31.89% | | 8.2 Explainability of Outputs to Audit Depth | 4.57 | 262 of 800 | 32.75% | | 7.2 Stability of Tool Behavior When Nothing Changed | 4.57 | 263 of 797 | 33.00% | | 8.3 Logs and Telemetry Insufficient to Locate Failure | 4.57 | 255 of 796 | 32.04% | | 4.2 Efficiency of Token Use Under Retries and Loops | 4.58 | 255 of 788 | 32.36% | | 2.3 Inconsistency of Output Across Repeated Runs | 4.58 | 259 of 800 | 32.38% | | 5.1 Isolation of Information Between Financial Institutions and Sessions | 4.58 | 244 of 791 | 30.85% | | 3.1 Retention of Instructions Across Long Tasks and Sessions | 4.58 | 252 of 796 | 31.66% | | 4.1 Predictability of Token Consumption | 4.59 | 249 of 795 | 31.32% | | 2.1 Factual Reliability Sustained by Unrecorded Human Correction | 4.59 | 258 of 796 | 32.41% | | 8.1 False Completion in the Tools' Reports of Their Own Success | 4.60 | 251 of 786 | 31.93% | | 3.2 Capture of Material Facts Within Long Documents | 4.61 | 248 of 805 | 30.81% | | 6.2 Reliability of Built-In Guardrails in Production | 4.66 | 252 of 787 | 32.02% | Mean ratings in Table 2.6 are computed across respondents. The category figure of 4.60 is computed across engagements. Respondents described human review and correction as important components of effective tool use. Staff reported correcting outputs, rerunning tasks, dividing complex workflows into shorter steps, and reviewing results before delivery. Some accounts stated that these interventions were not recorded and that corrected results were attributed to the agent. Part 5 examines the reporting implications in detail. > “i correct it most times we use it, seven out of ten... my Lead know, client does not. in truth none of it feel reliable enough for what this is meant to be for, KYC.” -Individual Contributor, CXM, Financial crime and client onboarding Accounts of inaccurate completion reports, limited explanations, and incomplete logs raise questions about the evidence available for oversight. These findings concern reported experience. The study did not independently audit system behavior or institutional control records. > “The vendor's weekly deal-status report used the agent's self-reported completion percentage for data-room review without sampling actual output, so a data room the agent had barely parsed correctly still showed as '90 percent reviewed' heading into the client update.” -Individual Contributor, KPO, Corporate and back-office functions The survey invited comment on the tools, and 466 of the 850 respondents (54.82%) wrote one. Many were favorable within limits. Respondents described tools that read documents well, completed short tasks without intervention, and held their guardrails, and then named the limits they managed around by hand. > The two on security is not about a leak... the tool pulled a client name into a working note it had no business writing to. Nothing left the building, but I had to clear it. On reliability, a six step attribution run does five steps right then quietly drops the currency effect at the end, and never says it dropped anything. Reading a long factsheet it is fine. Chaining is where it falls over” INDIVIDUAL CONTRIBUTOR · KPO, Trade, settlement and reconciliation ## 1 Production Readiness of the Tools Respondents could value the AI tools while describing limitations that required shorter tasks, repeated instructions, or human review before production use. The overall production-readiness indicator provides a summary assessment of the tools as experienced on the engagement. ### 1.1 Perceived Production Readiness Indicator C171 · Rate the overall production readiness of the AI tools for bank-grade work as experienced on this engagement · Full statistical analysis › Favorable accounts of AI production readiness were often conditional on shorter tasks, repeated instructions, additional runs, or human review. The mean rating was 4.5 and the median 5.0 (n = 802 respondents). Respondents on canceled engagements rated the tools' readiness lower than respondents on active engagements in any stage (mean 3.1, n = 305 respondents). Figure 2.65. Perceived Production Readiness: associations with the study measures Scroll horizontally to view the complete figure. Perceived production readiness: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.69 [−0.77, −0.60] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.64 [0.50, 0.74] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.86 [0.82, 0.90] 150 engagements p < 0.001 | | Intent to leave | | r = −0.67 [−0.72, −0.61] 756 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.76807186490000001 | 1 | | 95% confidence interval | 1 | -0.59945653880000005 | 1 | | Reported association | 0 | -0.6931292617 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.49646203719999998 | 1 | | 95% confidence interval | 1 | 0.74227189670000004 | 1 | | Reported association | 0 | 0.63518473340000003 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.81576156399999999 | 1 | | 95% confidence interval | 1 | 0.89906593089999998 | 1 | | Reported association | 0 | 0.86317565429999998 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.71851788139999995 | 1 | | 95% confidence interval | 1 | -0.60974017879999998 | 1 | | Reported association | 0 | -0.66665938769999999 | 1 | Note: Survey question: “Rate the overall production readiness of the AI tools for bank-grade work as experienced on this engagement”. Distressed ratings: 257 of 802 answers (32.04%). Group means and numeric-answer bases: Individual contributor (n = 619): 4.40. Engagement manager (n = 150): 4.92. Portfolio director (n = 31): 5.35. Build-Out (n = 167): 5.49. Transition/Cut-Over (n = 234): 5.51. Production (n = 96): 5.20. Canceled (n = 305): 3.06. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. Production readiness of the tools showed an association with every core study measure, and none of the applicable robustness checks overturned one. Engagements that rated the tools more ready had a lower cancellation rate (r = −0.69, 95% CI −0.77 to −0.60, n = 150 engagements), the largest observed absolute cancellation coefficient among the 100 indicators. This descriptive ordering does not establish a significant difference between coefficients. Customer satisfaction among active engagements was higher where the tools were rated more ready (r = 0.64, 95% CI 0.50 to 0.74, n = 94 engagements). A higher engagement indicator mean, computed without the readiness indicator, accompanied tools rated more ready (r = 0.86, 95% CI 0.82 to 0.90, n = 150 engagements). Respondents who rated the tools more ready reported lower intent to leave (r = −0.67, cluster bootstrap 95% CI −0.72 to −0.61, n = 756 respondents). Distressed ratings came from 257 of the 802 answers (32.04%), and the engagement mean was distressed on 36 of the 150 engagements. The written responses addressing production readiness numbered 79, and 60 were positive and 15 negative. No other of the 100 indicators with 30 or more written responses had a lower share of negative responses. Respondents described tools that worked, and then named the limits they managed around. Those limits were instructions dropped on long runs, tasks reported complete that were not, results that varied between runs, and citations to the wrong source. The negative responses described correcting most of the tools' output. > “Tools have not been our trouble on this one, they have run a whole quarter without me thinking about them, which I did not expect going in. Only thing I would note is an agent once reached for a settlement screen it had no business in, sandbox caught it, but that is what I watch for.” -Engagement Manager, BPO, Corporate and back-office functions > “Two things would have to change before I would put this in front of a client executive again. First, the same question asked twice should produce the same answer twice … Second, the assistant reads the front of a policy document and the back of it and misses what sits in the middle, which is usually where the exceptions live. Everything else is manageable.” -Portfolio Director, CXM, Corporate and back-office functions > “asking me to rate production readiness on 1 to 10 is not so simple, the tool handle long access reports well and finish multi step job properly, but same job on Tuesday and Thursday do not behave same, one number cannot hold both of these” -Individual Contributor, ITO, Technology engineering and operations > “The tools are if I am honest fine, i have no big complaint there. It reports the run as complete when a figure in the middle of the pack is missing... so i still read every page myself. That checking is what eats my day, not the modelling.” -Individual Contributor, KPO, Corporate and back-office functions More negative written sentiment about production readiness was associated with cancellation and higher intent to leave. Favorable accounts were often conditional on shorter tasks, repeated instructions, additional runs, or human review. These conditions help explain how respondents could value the tools while identifying operational limitations. ## 2 Output Reliability and Nondeterminism Reported factual errors, unsupported references, inconsistent runs, and incomplete tasks made verification and correction recurring delivery activities. Four indicators assess factual reliability before human review, citation accuracy, consistency across repeated runs, and end-to-end reliability on multistep tasks. These dimensions distinguish whether an output is correct from whether the process producing it is consistent and complete. > “Nothing about the AI output is repeatable either, it flags different vendors each run, so I change the result myself and paper it afterwards, or sometimes not at all.” -Individual Contributor, BPO, Corporate and back-office functions Respondents described verification and correction as recurring delivery activities. Part 5 examines whether manual alteration was recorded and disclosed. ### 2.1 Factual Reliability Before Human Review Indicator C157 · Rate the factual reliability of the AI tools’ outputs on this engagement before any human review · Full statistical analysis › Respondents described correcting factual errors in AI outputs before use, with those corrections not consistently recorded or disclosed to the financial institution. The mean rating was 4.6 and the median 5.0 (n = 796 respondents). Factual reliability was weakest on canceled engagements, where the mean rating was 3.1 (n = 302 respondents), lower than on active engagements in any stage. Figure 2.66. Factual Reliability Before Human Review: associations with the study measures Scroll horizontally to view the complete figure. Factual reliability before human review: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.67 [−0.75, −0.57] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.65 [0.52, 0.76] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.87 [0.83, 0.91] 150 engagements p < 0.001 | | Intent to leave | | r = −0.65 [−0.70, −0.60] 746 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.7513137433 | 1 | | 95% confidence interval | 1 | -0.57345424209999996 | 1 | | Reported association | 0 | -0.67196078969999995 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.52029752330000001 | 1 | | 95% confidence interval | 1 | 0.75636608360000002 | 1 | | Reported association | 0 | 0.65397203349999999 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.83082064789999999 | 1 | | 95% confidence interval | 1 | 0.90766223509999999 | 1 | | Reported association | 0 | 0.87462684700000004 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.70425460709999999 | 1 | | 95% confidence interval | 1 | -0.59589266119999995 | 1 | | Reported association | 0 | -0.65334896170000001 | 1 | Note: Survey question: “Rate the factual reliability of the AI tools’ outputs on this engagement before any human review”. Distressed ratings: 258 of 796 answers (32.41%). Group means and numeric-answer bases: Individual contributor (n = 613): 4.44. Engagement manager (n = 150): 4.97. Portfolio director (n = 31): 5.74. Build-Out (n = 163): 5.49. Transition/Cut-Over (n = 235): 5.55. Production (n = 96): 5.33. Canceled (n = 302): 3.13. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. Every association of the factual reliability of the tools' output with the four core study measures held, in each of the applicable robustness checks. Engagements whose tools were rated more reliable had a lower observed cancellation rate (r = −0.67, 95% CI −0.75 to −0.57, n = 150 engagements). Among active engagements, customer satisfaction was higher where output was rated more reliable (r = 0.65, 95% CI 0.52 to 0.76, n = 94 engagements). Factual reliability also tracked the engagement indicator mean computed without it (r = 0.87, 95% CI 0.83 to 0.91, n = 150 engagements). Respondents who rated the output more reliable reported lower intent to leave (r = −0.65, cluster bootstrap 95% CI −0.70 to −0.60, n = 746 respondents). Distressed ratings came from 258 of the 796 answers (32.41%), and the engagement mean was distressed on 30 of the 150 engagements. The indicator was the subject of 628 written responses across 14 survey questions, the largest written base of any of the 100 indicators, and 548 of the 628 were negative. In answer to the operational questions, respondents described output that staff corrected before it went on, usually because the same input gave a different result on another run. The corrections were seldom logged, and in many accounts the financial institution was not told. In comment on the tools, respondents described invented content, runs reported complete that were not, and logs that could not show why. > “i keep a sheet of every generated test case we had to rewrite by hand, about sixty percent of them at this point. before any client demo i run the old regression pack myself instead of the generated one, because that one i trust. The study is not in any process document, it is only mine … ” -Individual Contributor, ITO, Technology engineering and operations > “My people stopped trusting it around month four and that is hard to walk back … Two of my best case handlers asked to move off the queue because they felt they were doing the same job twice, once by hand and once for the tool. I lost them. That cost isnt anywhere in these questions and it was the real cost.” -Engagement Manager, BPO, Corporate and back-office functions > “one saturday it matched a 40,000 dollar break to wrong account and marked it closed. i found it monday only because the cash didnt tie. it never flagged anything, the log say completed. after that i tie the cash myself before anything is signed, kindly there is no other way.” -Individual Contributor, BPO, Trade, settlement and reconciliation > “I correct AI output on vendor exceptions most days, team lead knows about it. … so if I don't catch it nobody else will.” -Individual Contributor, BPO, Corporate and back-office functions More negative written sentiment was associated with lower satisfaction among active engagements, cancellation, and higher intent to leave. Respondents’ accounts of manual correction and attribution of corrected output to agents are examined in Part 5. ### 2.2 Accuracy of Citations and References Against Source Indicator C158 · Rate the accuracy of citations, references, and quotations the tools produced against the real sources · Full statistical analysis › Inaccurate citations, references, and quotations in AI output required respondents to check the claimed support against the actual sources. The mean rating was 4.5 and the median 5.0 (n = 791 respondents). Citation accuracy received a mean rating of 3.2 on canceled engagements (n = 301 respondents), lower than on active engagements in any stage. Figure 2.67. Accuracy of Citations and References Against Source: associations with the study measures Scroll horizontally to view the complete figure. Accuracy of citations and references against source: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.65 [−0.73, −0.54] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.65 [0.52, 0.76] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.86 [0.81, 0.89] 150 engagements p < 0.001 | | Intent to leave | | r = −0.64 [−0.69, −0.58] 746 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.73018245969999995 | 1 | | 95% confidence interval | 1 | -0.54116943449999999 | 1 | | Reported association | 0 | -0.64545180390000001 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.52072984789999999 | 1 | | 95% confidence interval | 1 | 0.75661971409999995 | 1 | | Reported association | 0 | 0.65431127919999998 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.80555277589999996 | 1 | | 95% confidence interval | 1 | 0.8932016008 | 1 | | Reported association | 0 | 0.8553848704 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.69433975209999999 | 1 | | 95% confidence interval | 1 | -0.57636274369999996 | 1 | | Reported association | 0 | -0.63792617709999999 | 1 | Note: Survey question: “Rate the accuracy of citations, references, and quotations the tools produced against the real sources”. Distressed ratings: 265 of 791 answers (33.50%). Group means and numeric-answer bases: Individual contributor (n = 611): 4.42. Engagement manager (n = 150): 4.90. Portfolio director (n = 29): 5.45. Build-Out (n = 163): 5.44. Transition/Cut-Over (n = 233): 5.35. Production (n = 94): 5.34. Canceled (n = 301): 3.20. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. Citation accuracy was associated with cancellation, customer satisfaction, the engagement indicator mean, and intent to leave, and none of the applicable robustness checks overturned an association. Engagements with more accurate citations had lower observed cancellation rates (r = −0.65, 95% CI −0.73 to −0.54, n = 150 engagements). Customer satisfaction among active engagements was higher where citations were rated more accurate (r = 0.65, 95% CI 0.52 to 0.76, n = 94 engagements). Where citations were more accurate, the engagement indicator mean computed without the citation indicator was higher (r = 0.86, 95% CI 0.81 to 0.89, n = 150 engagements). Intent to leave was lower among respondents who rated citations more accurate (r = −0.64, cluster bootstrap 95% CI −0.69 to −0.58, n = 746 respondents). Distressed ratings came from 265 of the 791 answers (33.50%), the highest distressed share among the 20 indicators of AI technology, and the engagement mean was distressed on 32 of the 150 engagements. The written responses addressing citations numbered 36, nearly all in comment on the tools, and 31 of the 36 were negative. Respondents described citations to clauses, sections, and values that did not exist, and citations to real documents that did not say what the output claimed. The response was to check every reference by hand. > “When the tool drafts a payment arrangement it quotes the collections guidance back at the customer, chapter and clause, and roughly one time in five that clause does not say what the tool says it says. Twice it invented a paragraph number that does not exist. On a hardship call, with a customer who is already frightened, that is not a small error, it is a compliance breach with a recording attached to it.” -Engagement Manager, CXM, Payments, cards and disputes > “every early-warning narrative the tool writes goes to an analyst with the source spreadsheet open beside it, and each figure it quotes is checked back to the filing it claims to come from. Roughly a third of the references do not survive that … Please note none of this is in any process document, it is simply what my team have worked out between them.” -Engagement Manager, KPO, Credit, underwriting and suitability > “every note that comes out i have to open each source link myself, and maybe one in five is pointing to some other report or a page that do not say what it claims. so a draft that took the tool ten minutes takes me an hour to trust.” -Individual Contributor, KPO, Corporate and back-office functions > “it says the case is reviewed and complete, then i open it and two transactions it quoted are not in the file at all. So now i check every one, even the ones it gets right” -Individual Contributor, BPO, Financial crime and client onboarding > “I keep a text file of the consent Policy references and paste the right one in after, because the tool quotes clauses that don't exist.” -Individual Contributor, ITO, Technology engineering and operations Written sentiment about citations showed no statistically significant association with the study measures. Respondents described reference checking as an additional workload, sometimes greater than the effort required to draft the material. ### 2.3 Inconsistency of Output Across Repeated Runs Indicator C159 · Rate the consistency of the tools’ outputs when given the same input and instructions more than once · Full statistical analysis › Repeated runs with the same input and instructions were reported to produce different outputs, requiring staff to select or correct a result before use. The mean rating was 4.6 and the median 5.0 (n = 800 respondents). The mean rating of consistency on canceled engagements, 3.2 (n = 309 respondents), was lower than on active engagements in any stage. Figure 2.68. Inconsistency of Output Across Repeated Runs: associations with the study measures Scroll horizontally to view the complete figure. Inconsistency of output across repeated runs: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.65 [−0.74, −0.55] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.66 [0.52, 0.76] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.88 [0.83, 0.91] 150 engagements p < 0.001 | | Intent to leave | | r = −0.63 [−0.69, −0.57] 750 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.73503699860000005 | 1 | | 95% confidence interval | 1 | -0.54853732369999997 | 1 | | Reported association | 0 | -0.65152376769999998 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.52455984509999998 | 1 | | 95% confidence interval | 1 | 0.75886355019999996 | 1 | | Reported association | 0 | 0.65731433699999997 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.83238770809999996 | 1 | | 95% confidence interval | 1 | 0.90855308499999998 | 1 | | Reported association | 0 | 0.87581566580000003 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.68687637130000001 | 1 | | 95% confidence interval | 1 | -0.56801175500000001 | 1 | | Reported association | 0 | -0.63188545829999998 | 1 | Note: Survey question: “Rate the consistency of the tools’ outputs when given the same input and instructions more than once”. Distressed ratings: 259 of 800 answers (32.38%). Group means and numeric-answer bases: Individual contributor (n = 617): 4.46. Engagement manager (n = 150): 4.90. Portfolio director (n = 31): 5.42. Build-Out (n = 165): 5.38. Transition/Cut-Over (n = 230): 5.54. Production (n = 96): 5.42. Canceled (n = 309): 3.17. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. The consistency of output across repeated runs showed an association with each of the four core study measures that survived all applicable robustness checks. Engagements with more consistent output had a lower cancellation rate (r = −0.65, 95% CI −0.74 to −0.55, n = 150 engagements). Active engagements with more consistent output reported higher customer satisfaction (r = 0.66, 95% CI 0.52 to 0.76, n = 94 engagements). Consistency was also closely associated with the engagement indicator mean computed without it (r = 0.88, 95% CI 0.83 to 0.91, n = 150 engagements). Respondents who rated output more consistent reported lower intent to leave (r = −0.63, cluster bootstrap 95% CI −0.69 to −0.57, n = 750 respondents). Distressed ratings came from 259 of the 800 answers (32.38%), and the engagement mean was distressed on 33 of the 150 engagements. The indicator was the subject of 383 written responses, and 363 of the 383 were negative. Respondents described the same input producing a different result on another run. Staff would not sign off such output as it stood. They chose among the runs, or corrected the output by hand before it went on. > “Our practice now is to run it all twice, sometimes three times, and only action flags that come up in both runs. Sounds wasteful and it is, but its the only way my leads trust the list.” -Engagement Manager, CXM, Regulatory, audit, tax and legal > “One suitability review last month, the tool read a forty page advisory file and missed a risk tolerance change recorded in the middle of it, then found the same change on a second run of the same file. Instruction retention was never our problem, consistency between runs is.” -Engagement Manager, CXM, Credit, underwriting and suitability > “same dispute file was run two times last week, first time it picks reason code 13.1, second time 12.5. Input was not changed. i check both against the network extract by myself.” -Individual Contributor, BPO, Payments, cards and disputes > “These question ask me to rate reliability from 1 to 10, but tool is not wrong in a steady way, one run it match everything and next run it invent a counterparty that does not exist. There is no number for that, i put 1 because 1 is the lowest you gave” -Individual Contributor, BPO, Trade, settlement and reconciliation > “we did not touch anything on our side for three weeks and the same role mapping request came back approved one day and flagged the next... so now I run it twice before I trust it.” -Individual Contributor, ITO, Technology engineering and operations Written evidence on output consistency was associated with all four study measures. Engagements with a negative response had higher cancellation rates and, among active engagements, lower customer satisfaction. More negative sentiment was also associated with cancellation and higher intent to leave. Respondents described selecting or correcting outputs when repeated runs disagreed. ### 2.4 End-to-End Reliability on Multistep Tasks Indicator C160 · Rate the end-to-end reliability of the tools on multi-step tasks, from first step to finished output · Full statistical analysis › Respondents described tools that completed individual steps but omitted work in longer tasks and nevertheless reported the full task as complete. The mean rating was 4.5 and the median 5.0 (n = 782 respondents). On canceled engagements, multistep reliability was rated lower than on active engagements in any stage (mean 3.2, n = 297 respondents). Figure 2.69. End-to-End Reliability on Multistep Tasks: associations with the study measures Scroll horizontally to view the complete figure. End-to-end reliability on multistep tasks: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.65 [−0.74, −0.55] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.65 [0.52, 0.75] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.85 [0.80, 0.89] 150 engagements p < 0.001 | | Intent to leave | | r = −0.65 [−0.71, −0.59] 741 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.73725316060000001 | 1 | | 95% confidence interval | 1 | -0.5519105205 | 1 | | Reported association | 0 | -0.65429926230000002 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.51620980750000001 | 1 | | 95% confidence interval | 1 | 0.75396444330000001 | 1 | | Reported association | 0 | 0.65076173790000003 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.79593077729999995 | 1 | | 95% confidence interval | 1 | 0.88764691549999997 | 1 | | Reported association | 0 | 0.84802119509999996 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.70677987360000005 | 1 | | 95% confidence interval | 1 | -0.59174462500000002 | 1 | | Reported association | 0 | -0.65205911829999996 | 1 | Note: Survey question: “Rate the end-to-end reliability of the tools on multi-step tasks, from first step to finished output”. Distressed ratings: 254 of 782 answers (32.48%). Group means and numeric-answer bases: Individual contributor (n = 602): 4.39. Engagement manager (n = 150): 4.99. Portfolio director (n = 29): 5.34. Build-Out (n = 167): 5.44. Transition/Cut-Over (n = 230): 5.39. Production (n = 88): 5.35. Canceled (n = 297): 3.15. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. All four core study measures were associated with the reliability of the tools on multistep tasks, and none of the applicable robustness checks overturned an association. Cancellation was less common where multistep reliability was rated higher (r = −0.65, 95% CI −0.74 to −0.55, n = 150 engagements). Customer satisfaction among active engagements was higher where long tasks ran reliably (r = 0.65, 95% CI 0.52 to 0.75, n = 94 engagements). Multi-step reliability was associated with a higher engagement indicator mean, computed without the multistep indicator (r = 0.85, 95% CI 0.80 to 0.89, n = 150 engagements). Respondents who rated multistep reliability higher reported lower intent to leave (r = −0.65, cluster bootstrap 95% CI −0.71 to −0.59, n = 741 respondents). Distressed ratings came from 254 of the 782 answers (32.48%), and the engagement mean was distressed on 36 of the 150 engagements. The written responses addressing multistep reliability numbered 51, nearly all in comment on the tools, with 36 negative and 13 positive. Respondents described tools that handled single steps well and lost a step or the thread on long chains, and then reported the run complete. Teams responded by cutting each task into single steps with a person between them. > “What we do now is cut the work into single steps, because the tool cannot hold a case together end to end. It reads the alert, we stop it, an analyst checks, then it drafts the narrative, we stop it again. Nothing runs more than one hop unattended … Every narrative is rewritten by a person anyway, so what the tool actually saves us is the first paragraph.” -Engagement Manager, BPO, Financial crime and client onboarding > “On a long inquiry it will do the first two steps correctly and then lose the thread completely, and there is no way to prove beforehand which ones it will finish, consequently somebody must sit with every single case” -Individual Contributor, CXM, Client servicing and communications > “We stopped letting it run a validation end to end. It goes in four pieces now, and somebody on my team signs off each piece before the next one starts, which in a bad week is slower than doing the work by hand.” -Engagement Manager, KPO, Credit, underwriting and suitability > “Practically it means I cannot promise my processors a finish time. A run that clears three hundred reports one night sits there the next, and because the tool reports itself as done either way, someone has to eyeball the queue before we go home, which lands on the same two people every week” -Engagement Manager, BPO, Corporate and back-office functions > “The vendor excludes settlement runs that failed mid-cycle from the reported success rate, so the dashboard only ever shows completed cycles and never the ones agents had to abandon and hand back to a human.” -Individual Contributor, BPO, Trade, settlement and reconciliation More negative written sentiment about multistep tasks was associated with higher intent to leave. Accounts described both incomplete execution and inaccurate completion reports. The latter could make incomplete work harder to detect, although the study did not compare the costs of these two failure modes. > My four on the consistency question is really about one thing, which is that the same alert given twice does not always come back with the same verdict. On a phishing bundle it will call it benign in one pass and suspicious in the next, and my analysts then have to decide which pass they believe. Everything else, the multi step enrichment, the pulling of indicators from the ticket, that part works better than I expect at this stage. But the inconsistency is what my team feels, so it pulls the whole section down for me.” Asking for one number on reliability doesn't work here. On a password reset the tool is near perfect; on a benefits question that spans three policies it falls apart at step four. Those are not the same product.” ENGAGEMENT MANAGER · ITO, Technology engineering and operations · “ ENGAGEMENT MANAGER CXM, Corporate and back-office functions ## 3 Memory, Context, and Retention Loss, omission, or reuse of incorrect information was reported to require repeated instructions, shorter sessions, source review, or manual correction. Three indicators assess retention of earlier instructions, retrieval of material facts from long documents, and the integrity of persistent memory. These capabilities concern different ways in which information can be omitted, lost, or incorrectly reused during agent operation. Respondents described restating instructions, shortening sessions, dividing documents, and reviewing source material to compensate for perceived limitations. > “Here a joiner-mover-leaver run goes on for an hour across four systems, and somewhere in the middle the agent quietly loses the exception we set at the start, so a contractor keeps birthright access he should not have.” -Engagement Manager, ITO, Technology engineering and operations ### 3.1 Retention of Instructions Across Long Tasks and Sessions Indicator C164 · Rate how reliably the tools retained instructions and constraints given earlier in long tasks and sessions · Full statistical analysis › Loss of earlier instructions during long tasks or sessions led respondents to describe repeating constraints or shortening the work assigned to agents. The mean rating was 4.6 and the median 5.0 (n = 796 respondents). Respondents on canceled engagements rated retention lower than respondents on active engagements in any stage (mean 3.2, n = 306 respondents). Figure 2.70. Retention of Instructions Across Long Tasks and Sessions: associations with the study measures Scroll horizontally to view the complete figure. Retention of instructions across long tasks and sessions: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.67 [−0.75, −0.57] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.68 [0.56, 0.78] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.87 [0.83, 0.91] 150 engagements p < 0.001 | | Intent to leave | | r = −0.64 [−0.70, −0.58] 748 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.74746461269999998 | 1 | | 95% confidence interval | 1 | -0.56753202229999999 | 1 | | Reported association | 0 | -0.66711689959999998 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.55762974269999999 | 1 | | 95% confidence interval | 1 | 0.77800917839999995 | 1 | | Reported association | 0 | 0.68306988800000001 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.82667779409999997 | 1 | | 95% confidence interval | 1 | 0.90530374599999996 | 1 | | Reported association | 0 | 0.87148141450000005 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.69506544979999996 | 1 | | 95% confidence interval | 1 | -0.57930738159999995 | 1 | | Reported association | 0 | -0.64015030979999998 | 1 | Note: Survey question: “Rate how reliably the tools retained instructions and constraints given earlier in long tasks and sessions”. Distressed ratings: 252 of 796 answers (31.66%). Group means and numeric-answer bases: Individual contributor (n = 613): 4.46. Engagement manager (n = 150): 4.86. Portfolio director (n = 31): 5.52. Build-Out (n = 161): 5.30. Transition/Cut-Over (n = 228): 5.54. Production (n = 101): 5.54. Canceled (n = 306): 3.18. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. The retention of instructions across long tasks and sessions was associated with all four core study measures, and each association held under the applicable robustness checks. Engagements whose tools retained instructions better had lower observed cancellation rates (r = −0.67, 95% CI −0.75 to −0.57, n = 150 engagements). Customer satisfaction among active engagements was higher where retention was rated stronger (r = 0.68, 95% CI 0.56 to 0.78, n = 94 engagements). The engagement indicator mean, computed without the retention indicator, was higher where retention was rated stronger (r = 0.87, 95% CI 0.83 to 0.91, n = 150 engagements). Respondents who rated retention higher reported lower intent to leave (r = −0.64, cluster bootstrap 95% CI −0.70 to −0.58, n = 748 respondents). Distressed ratings came from 252 of the 796 answers (31.66%), and the engagement mean was distressed on 39 of the 150 engagements. The written responses addressing retention numbered 66, all in comment on the tools, with 49 negative and 17 positive. Respondents described tools that lost an instruction given at the start of a long task or session, and teams that restated the instructions or cut each task short to compensate. > “the tool is fine on two step and useless on six step, so what number is that … and it forgets the rule i gave at start, everytime, by the end it does what it likes.” -Individual Contributor, BPO, Trade, settlement and reconciliation > “My only reservation, and it is minor, is that long-running sessions do drift from constraints given early on, so the team resets context more often than I would like. That is a small price against what we had before.” -Portfolio Director, CXM, Technology engineering and operations > “The mandate constraints go into a short file that the analyst pastes in again at every stage... it is not elegant but the drafting agent forgets them by the third section otherwise.” -Engagement Manager, KPO, Credit, underwriting and suitability > “it makes my week longer, not shorter. i give it the suppression rule in the morning and by evening it forget them, so i recheck every triage note myself before shift end, my team also do same” -Individual Contributor, ITO, Technology engineering and operations > “For a POC the bot holds up fine, it remembers what you told it earlier in long chats.” -Individual Contributor, CXM, Corporate and back-office functions Written sentiment about instruction retention showed no statistically significant association with the study measures. Respondents described repeated instructions and shorter sessions as compensating practices. ### 3.2 Capture of Material Facts Within Long Documents Indicator C165 · Rate how completely the tools captured material facts located anywhere in long documents, including mid-document content · Full statistical analysis › Some tools were reported to miss material facts in the middle of long documents, leaving staff to review the source themselves. The mean rating was 4.61 and the median 5.0 (n = 805 respondents). Capture of material facts received a mean rating of 3.2 on canceled engagements (n = 300 respondents), lower than on active engagements in any stage. Figure 2.71. Capture of Material Facts Within Long Documents: associations with the study measures Scroll horizontally to view the complete figure. Capture of material facts within long documents: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.66 [−0.74, −0.56] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.66 [0.53, 0.76] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.87 [0.82, 0.90] 150 engagements p < 0.001 | | Intent to leave | | r = −0.67 [−0.72, −0.62] 755 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.74490346210000002 | 1 | | 95% confidence interval | 1 | -0.56360175970000004 | 1 | | Reported association | 0 | -0.66389761169999995 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.53074692079999997 | 1 | | 95% confidence interval | 1 | 0.76247658709999999 | 1 | | Reported association | 0 | 0.66215666350000002 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.82492196390000005 | 1 | | 95% confidence interval | 1 | 0.90430270280000002 | 1 | | Reported association | 0 | 0.87014719979999999 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.71842431709999999 | 1 | | 95% confidence interval | 1 | -0.61966524329999995 | 1 | | Reported association | 0 | -0.67125268999999999 | 1 | Note: Survey question: “Rate how completely the tools captured material facts located anywhere in long documents, including mid-document content”. Distressed ratings: 248 of 805 answers (30.81%). Group means and numeric-answer bases: Individual contributor (n = 623): 4.47. Engagement manager (n = 150): 5.01. Portfolio director (n = 30): 5.43. Build-Out (n = 164): 5.48. Transition/Cut-Over (n = 241): 5.45. Production (n = 100): 5.38. Canceled (n = 300): 3.21. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. Complete capture of material facts in long documents accompanied more favorable results on all four core study measures, each confirmed by the applicable robustness checks. Cancellation was less common where capture was rated more complete (r = −0.66, 95% CI −0.74 to −0.56, n = 150 engagements). Among active engagements, customer satisfaction was higher where capture was rated more complete (r = 0.66, 95% CI 0.53 to 0.76, n = 94 engagements). Complete capture accompanied a higher engagement indicator mean, computed without the capture indicator (r = 0.87, 95% CI 0.82 to 0.90, n = 150 engagements). Respondents who rated capture higher reported lower intent to leave (r = −0.67, cluster bootstrap 95% CI −0.72 to −0.62, n = 755 respondents). Distressed ratings came from 248 of the 805 answers (30.81%), the lowest distressed share among the 20 indicators of AI technology, and the engagement mean was distressed on 35 of the 150 engagements. The written responses addressing long documents numbered 53, all in comment on the tools, with 35 negative and 18 positive. Respondents described tools that read the opening and the close of a long document and missed material content in between, and staff who read the source themselves to close the gap. > “I don't use the summary anymore, I pull the page numbers it cites and read those myself. Anything in the middle of a long agreement it skips, so I go through the schedules by hand at the end. Takes longer than not using it.” -Individual Contributor, KPO, Corporate and back-office functions > “For this to work the tool must read whole contract, not first two page and last one. Half the vendor terms sit in middle and it miss them every time. Fix that first, then talk about autonomy.” -Individual Contributor, BPO, Corporate and back-office functions > “Last month it missed a side letter clause sitting in middle of a 60 page LPA and answered the question wrong... And the answer read perfectly fine, that is what unsettles me” -Individual Contributor, CXM, Corporate and back-office functions > “Two weeks back it read a forty page trust deed and missed a clause on page twenty two about co-trustee approval, it took everything from first pages only. I only catch it because number of signatories look wrong to me.” -Individual Contributor, BPO, Corporate and back-office functions > “We changed the prompt so it must give the page number for every finding it reports. Then the reviewer can go straight to that page instead of reading the whole file again. It finds things buried in the middle of a long record very well, but on its own it will not tell you where it got them from” -Individual Contributor, KPO, Credit, underwriting and suitability Written sentiment about long-document processing showed no statistically significant association with the study measures. Accounts described omissions from the middle of documents and subsequent manual source review. They do not establish the frequency of such omissions across all tasks. ### 3.3 Integrity of Persistent Memory Against Contamination Indicator C166 · Rate the integrity of the tools’ persistent memory against contamination by incorrect or malicious content encountered during work · Full statistical analysis › The small written-answer set on persistent memory described incorrect content being stored and reused until a person identified and removed it. The mean rating was 4.6 and the median 4.0 (n = 806 respondents). The integrity of memory was weakest on canceled engagements, where the mean rating was 3.1 (n = 305 respondents), lower than on active engagements in any stage. Figure 2.72. Integrity of Persistent Memory Against Contamination: associations with the study measures Scroll horizontally to view the complete figure. Integrity of persistent memory against contamination: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.68 [−0.76, −0.58] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.64 [0.51, 0.75] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.87 [0.82, 0.90] 150 engagements p < 0.001 | | Intent to leave | | r = −0.65 [−0.71, −0.59] 760 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.75852422959999999 | 1 | | 95% confidence interval | 1 | -0.58459845160000001 | 1 | | Reported association | 0 | -0.68105304430000002 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.50804336169999997 | 1 | | 95% confidence interval | 1 | 0.74914733450000004 | 1 | | Reported association | 0 | 0.64433376580000001 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.82492267180000001 | 1 | | 95% confidence interval | 1 | 0.90430310660000002 | 1 | | Reported association | 0 | 0.87014773779999999 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.70627997679999999 | 1 | | 95% confidence interval | 1 | -0.59304046479999994 | 1 | | Reported association | 0 | -0.65401362659999995 | 1 | Note: Survey question: “Rate the integrity of the tools’ persistent memory against contamination by incorrect or malicious content encountered during work”. Distressed ratings: 257 of 806 answers (31.89%). Group means and numeric-answer bases: Individual contributor (n = 625): 4.45. Engagement manager (n = 150): 4.90. Portfolio director (n = 30): 5.30. Build-Out (n = 160): 5.49. Transition/Cut-Over (n = 242): 5.50. Production (n = 99): 5.22. Canceled (n = 305): 3.12. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. Across all four core study measures, the integrity of persistent memory showed associations that survived the applicable robustness checks. Engagements whose tools were rated higher for memory integrity had a lower cancellation rate (r = −0.68, 95% CI −0.76 to −0.58, n = 150 engagements), the second-largest observed absolute cancellation coefficient among the 100 indicators. This is a descriptive ordering. Customer satisfaction among active engagements was higher where memory integrity was rated higher (r = 0.64, 95% CI 0.51 to 0.75, n = 94 engagements). Memory integrity was likewise associated with a higher engagement indicator mean computed without it (r = 0.87, 95% CI 0.82 to 0.90, n = 150 engagements). Respondents who rated memory integrity higher reported lower intent to leave (r = −0.65, cluster bootstrap 95% CI −0.71 to −0.59, n = 760 respondents). Distressed ratings came from 257 of the 806 answers (31.89%), and the engagement mean was distressed on 32 of the 150 engagements. Nine written responses addressed persistent memory, all in comments about the tools. Eight were negative. They described incorrect content stored and reused until a person identified and removed it. This small sample provides illustrative evidence only. > “It reads a custodian statement with a bad line in it, and that wrong shape stays with it for the next file... we have had to clear the memory twice and nobody could tell us what else went with it.” -Engagement Manager, BPO, Trade, settlement and reconciliation > “once it learns a wrong routing it keeps repeating the same wrong routing, and it is confident every time. so my team stopped believing it and we started checking every batch by hand first thing in the morning, which is nearly two extra hours on the shift and it is not counted anywhere.” -Individual Contributor, CXM, Client servicing and communications > “It should forget the old sample once the cycle close. Twice it carried last month's exception rule into the new run and we did not see it” -Individual Contributor, KPO, Regulatory, audit, tax and legal > “I start a new session for every batch now, the old one carries wrong answers forward and repeats them.” -Individual Contributor, CXM, Corporate and back-office functions > “Only the memory part bothers me. it keeps an old answer from a previous DDQ and reuses it even when the fund is different” -Individual Contributor, CXM, Corporate and back-office functions The structured ratings assess perceived memory integrity. They do not estimate the prevalence of the specific failure or remedy described in these nine accounts. > Half of these ask me to rate things we had no way to observe... Rate quality of logs. There were almost no logs, so is that a one or is it not applicable. Memory contamination the same, we would only learn of it when a figure came out wrong, and at that point I am looking at the figure, not at the tool. What I can answer properly is that it loses constraints given forty minutes earlier, and in submission work that is not small.” ENGAGEMENT MANAGER · KPO, Regulatory, audit, tax and legal ## 4 Token Consumption and Cost Predictability Unexplained variation and limited visibility into repeated processing made token consumption difficult for some respondents to forecast or explain. Tokens are the units of text processed by the model. Two indicators assess the predictability of token consumption and the efficiency of token use after retries, loops, and reprocessing. The distinction separates the ability to forecast consumption from the amount of computational work required to complete a task. The limited written evidence emphasizes unexplained variation and poor visibility into repeated processing. Such conditions could complicate budgeting, but the study did not measure token costs or establish their financial effect. > “On a news heavy day the same lead volume costs us three times what it did the week before and nobody can tell me why in advance, so I cant plan a month.” -Engagement Manager, KPO, Client servicing and communications ### 4.1 Predictability of Token Consumption Indicator C161 · Rate the predictability of the tools’ token consumption for a given volume of work · Full statistical analysis › Variable token consumption for the same volume of work left respondents describing difficulty forecasting and explaining expenditure. The mean rating was 4.6 and the median 4.0 (n = 795 respondents). On canceled engagements, the predictability of token consumption was rated lower than on active engagements in any stage (mean 3.2, n = 305 respondents). Figure 2.73. Predictability of Token Consumption: associations with the study measures Scroll horizontally to view the complete figure. Predictability of token consumption: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.65 [−0.73, −0.54] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.62 [0.48, 0.73] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.86 [0.81, 0.89] 150 engagements p < 0.001 | | Intent to leave | | r = −0.64 [−0.70, −0.58] 749 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.73151364870000002 | 1 | | 95% confidence interval | 1 | -0.54318693900000004 | 1 | | Reported association | 0 | -0.64711576449999997 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.4793944302 | 1 | | 95% confidence interval | 1 | 0.73204421539999998 | 1 | | Reported association | 0 | 0.62163030239999995 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.80526169989999996 | 1 | | 95% confidence interval | 1 | 0.89303395689999998 | 1 | | Reported association | 0 | 0.85516240619999995 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.69763267390000006 | 1 | | 95% confidence interval | 1 | -0.57833267460000004 | 1 | | Reported association | 0 | -0.64168166049999997 | 1 | Note: Survey question: “Rate the predictability of the tools’ token consumption for a given volume of work”. Distressed ratings: 249 of 795 answers (31.32%). Group means and numeric-answer bases: Individual contributor (n = 617): 4.46. Engagement manager (n = 150): 4.92. Portfolio director (n = 27): 5.63. Build-Out (n = 160): 5.51. Transition/Cut-Over (n = 231): 5.47. Production (n = 99): 5.33. Canceled (n = 305): 3.21. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. All four core study measures were associated with the predictability of token consumption, and none of the applicable robustness checks overturned an association. Engagements whose token consumption was more predictable had a lower cancellation rate (r = −0.65, 95% CI −0.73 to −0.54, n = 150 engagements). Customer satisfaction among active engagements was higher where consumption was rated more predictable (r = 0.62, 95% CI 0.48 to 0.73, n = 94 engagements). Predictability and the engagement indicator mean computed without it were closely correlated (r = 0.86, 95% CI 0.81 to 0.89, n = 150 engagements). Intent to leave was lower among respondents who rated predictability higher (r = −0.64, cluster bootstrap 95% CI −0.70 to −0.58, n = 749 respondents). Distressed ratings came from 249 of the 795 answers (31.32%), and the engagement mean was distressed on 32 of the 150 engagements. The written responses addressing predictability numbered 30, all in comment on the tools, with 20 negative and 10 positive. Respondents described the same volume of work costing different amounts from one run to the next, with no explanation afterward and no estimate before the run. > “We could never say up front what a run was going to cost or how long it would sit there. Same job, same rule set, one day it burns through everything and the next day it's fine. Nobody could explain that to me and I asked twice.” -Individual Contributor, ITO, Regulatory, audit, tax and legal > “I also get asked by our own management why the spend moves week to week, and I have nothing sensible to tell them. Credit where it is due, it does find the one line buried in the middle of a forty page spec, which my people would miss.” -Engagement Manager, ITO, Technology engineering and operations > “But the cost side is hard to plan, a job that ran cheap on Monday runs three times heavier on Thursday on the same volume. What would help is a simple estimate before the run, so we can hold it back if the number looks silly.” -Individual Contributor, ITO, Technology engineering and operations > “The caveat is only cost. Same batch of purchase orders can cost twice as much one day to the next depending how many times it loops, and nobody can tell me in advance.” -Individual Contributor, BPO, Corporate and back-office functions Written sentiment about consumption predictability showed no statistically significant association with the study measures. Respondents’ concerns centered on forecasting and explaining expenditure rather than consistently describing its absolute level as excessive. ### 4.2 Efficiency of Token Use Under Retries and Loops Indicator C162 · Rate the efficiency of the tools’ token use, considering retries, loops, and reprocessing needed to complete work · Full statistical analysis › Retries, loops, and reprocessing were reported to increase token use without sufficiently visible counts to explain the additional consumption. The mean rating was 4.6 and the median 5.0 (n = 788 respondents). Respondents on canceled engagements gave token efficiency a mean rating of 3.2 (n = 299 respondents), lower than respondents on active engagements in any stage. Figure 2.74. Efficiency of Token Use Under Retries and Loops: associations with the study measures Scroll horizontally to view the complete figure. Efficiency of token use under retries and loops: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.66 [−0.74, −0.56] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.67 [0.54, 0.77] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.87 [0.83, 0.91] 150 engagements p < 0.001 | | Intent to leave | | r = −0.63 [−0.69, −0.57] 740 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.74195688599999998 | 1 | | 95% confidence interval | 1 | -0.5590901667 | 1 | | Reported association | 0 | -0.66019756620000003 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.53625197970000005 | 1 | | 95% confidence interval | 1 | 0.7656792456 | 1 | | Reported association | 0 | 0.66645600159999996 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.82761939640000004 | 1 | | 95% confidence interval | 1 | 0.90584021770000001 | 1 | | Reported association | 0 | 0.87219664330000002 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.69021427930000001 | 1 | | 95% confidence interval | 1 | -0.57045460140000004 | 1 | | Reported association | 0 | -0.63344802700000002 | 1 | Note: Survey question: “Rate the efficiency of the tools’ token use, considering retries, loops, and reprocessing needed to complete work”. Distressed ratings: 255 of 788 answers (32.36%). Group means and numeric-answer bases: Individual contributor (n = 605): 4.43. Engagement manager (n = 150): 4.95. Portfolio director (n = 31): 5.65. Build-Out (n = 159): 5.38. Transition/Cut-Over (n = 236): 5.51. Production (n = 94): 5.32. Canceled (n = 299): 3.18. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. Every association of the efficiency of token use with the four core study measures held, in each of the applicable robustness checks. More efficient token use was associated with a lower observed cancellation rate (r = −0.66, 95% CI −0.74 to −0.56, n = 150 engagements). Active engagements whose tools used tokens more efficiently reported higher customer satisfaction (r = 0.67, 95% CI 0.54 to 0.77, n = 94 engagements). The engagement indicator mean computed without the efficiency indicator was strongly correlated with it (r = 0.87, 95% CI 0.83 to 0.91, n = 150 engagements). Respondents who rated token use more efficient reported lower intent to leave (r = −0.63, cluster bootstrap 95% CI −0.69 to −0.57, n = 740 respondents). Distressed ratings came from 255 of the 788 answers (32.36%), and the engagement mean was distressed on 33 of the 150 engagements. Of the 24 written responses addressing token efficiency, 18 were negative and 5 positive. Nearly all appeared in comments about the tools. Respondents described repeated processing without visible retry counts, limiting their ability to explain increases in consumption. > “Every time a pipeline job failed halfway we paid for the whole run again, and we were doing that three or four times a day. i raised about this each week.” -Engagement Manager, ITO, Technology engineering and operations > “For the bigger cost centre runs i break the model into three smaller passes and run them separately, otherwise it loops and burns the whole day's budget. The finance leads never see this, it is simply how we run it now” -Individual Contributor, KPO, Corporate and back-office functions > “one, the tool should tell us when it retried, right now a job that looped eleven times looks exactly the same as one that went through clean.” -Individual Contributor, ITO, Technology engineering and operations > “in week one it got through the cert checks clean. now it stalls on the same file and nobody touched anything on our side... also nine retries on one submission, whatever that costs. i dont see the bill.” -Individual Contributor, ITO, Payments, cards and disputes > “First the retry behaviour on long policy lookups, at present one unanswered query can spin four or five times and we only see the cost at month end.” -Engagement Manager, CXM, Corporate and back-office functions Among 13 active engagements represented in the relevant comparison, a negative written response was associated with higher customer satisfaction, contrary to the structured-rating association. This small-sample result is exploratory and does not establish that inefficient token use improves satisfaction. Other written-sentiment tests were not statistically significant. > For my team the issue is not accuracy, it is what checking does to the day. My processors now spend the morning reviewing confidence flags on scanned invoices instead of working exceptions. And the tool retries itself in loops, so the volume they see is bigger than volume that came in. Nobody outside delivery counts that time. But it is real hours and my people feel it at month end.” ENGAGEMENT MANAGER · BPO, Corporate and back-office functions ## 5 Data Isolation, Sensitive Data, and Adversarial Input Written accounts identified unintended information exposure and instructions embedded in incoming content, with staff detecting or compensating for failures. Three indicators assess separation of financial institution and session information, handling of sensitive data, and resistance to hostile instructions embedded in processed content. They address confidentiality and instruction handling at different points in the workflow. The 63 written responses include accounts of information appearing in an unintended context, sensitive fields retained in accessible logs, and tools following instructions embedded in incoming material. These are respondent-reported incidents rather than independently verified security findings. The accounts warrant attention to incident detection and escalation. Two respondents reported receiving no further information after raising a concern. This does not establish whether an investigation or remediation occurred. > “also once i saw another clients name inside my session output, i reported it, no feedbacks after.” -Individual Contributor, BPO, Financial crime and client onboarding ### 5.1 Isolation of Information Between Financial Institutions and Sessions Indicator C163 · Rate how completely the tools kept each client’s and each session’s information isolated from every other client and session · Full statistical analysis › Some written accounts described information from another financial institution or session appearing in the respondent's work, usually identified before release. Respondents answered the question “Rate how completely the tools kept each client's and each session's information isolated from every other client and session.” The mean rating was 4.6 and the median 5.0 (n = 791 respondents). Respondents on canceled engagements rated isolation lower than respondents on active engagements in any stage (mean 3.2, n = 299 respondents). Figure 2.75. Isolation of Information Between Financial Institutions and Sessions: associations with the study measures Scroll horizontally to view the complete figure. Isolation of information between financial institutions and sessions: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.63 [−0.71, −0.52] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.70 [0.57, 0.79] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.87 [0.83, 0.91] 150 engagements p < 0.001 | | Intent to leave | | r = −0.66 [−0.71, −0.60] 745 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.71413363890000003 | 1 | | 95% confidence interval | 1 | -0.51701644449999995 | 1 | | Reported association | 0 | -0.62545418509999995 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.57422884139999997 | 1 | | 95% confidence interval | 1 | 0.78746727719999998 | 1 | | Reported association | 0 | 0.69588108250000003 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.82626818459999996 | 1 | | 95% confidence interval | 1 | 0.90507029530000005 | 1 | | Reported association | 0 | 0.87117022109999998 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.7076293387 | 1 | | 95% confidence interval | 1 | -0.60239931830000004 | 1 | | Reported association | 0 | -0.65850848569999998 | 1 | Note: Survey question: “Rate how completely the tools kept each client’s and each session’s information isolated from every other client and session”. Distressed ratings: 244 of 791 answers (30.85%). Group means and numeric-answer bases: Individual contributor (n = 608): 4.44. Engagement manager (n = 150): 4.98. Portfolio director (n = 31): 5.42. Build-Out (n = 160): 5.42. Transition/Cut-Over (n = 235): 5.46. Production (n = 97): 5.36. Canceled (n = 299): 3.19. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. The applicable robustness checks left intact each association between information isolation and the four core study measures. Engagements whose tools kept information isolated had a lower cancellation rate (r = −0.63, 95% CI −0.71 to −0.52, n = 150 engagements). Customer satisfaction among active engagements was higher where isolation was rated more complete (r = 0.70, 95% CI 0.57 to 0.79, n = 94 engagements). The engagement indicator mean, computed without the isolation indicator, was higher too (r = 0.87, 95% CI 0.83 to 0.91, n = 150 engagements). Respondents who rated isolation higher reported lower intent to leave (r = −0.66, cluster bootstrap 95% CI −0.71 to −0.60, n = 745 respondents). Distressed ratings came from 244 of the 791 answers (30.85%), and the engagement mean was distressed on 33 of the 150 engagements. The written responses addressing isolation numbered 26, all in comment on the tools, with 18 negative, 5 positive, and 3 neutral. Respondents described information from another account, mandate, or session appearing in their work, in most cases caught by the respondent before it went out. > “last month the tool pull a name from another client file into my summary, i notice only because address was wrong country... and nobody could tell me after how it got there” -Individual Contributor, CXM, Financial crime and client onboarding > “i also saw another account's details come up in a session that had nothing to do with it. i reported that one. nothing came back to me about it” -Individual Contributor, BPO, Payments, cards and disputes > “the forecast pack came back with a cost centre in it that belongs to another client, i had to delete it myself before anyone saw. previous engagement was plain reporting and this never happen there, the tool stayed inside its own box. here i am opening every tab looking for things that should not be there” -Individual Contributor, KPO, Corporate and back-office functions > “though it carries details over from the previous file sometimes, last week a corporate mandate note turned up on a trust case. Small thing, we catch the same before it goes out.” -Individual Contributor, BPO, Corporate and back-office functions > “Data separation is the one part that never worried me, client information stayed where it should.” -Engagement Manager, ITO, Financial crime and client onboarding Written sentiment about information isolation showed no statistically significant association with the study measures. Accounts described staff detecting unexpected information and adopting compensating work practices. They do not establish the adequacy of all technical confidentiality controls. ### 5.2 Handling of Sensitive Data in Outputs, Logs, and Connected Systems Indicator C169 · Rate how safely the tools handled sensitive data, with no exposure through outputs, logs, or connected systems · Full statistical analysis › Respondents described sensitive data appearing in accessible logs or unintended outputs and manual removal of fields before processing as a compensating practice. The mean rating was 4.5 and the median 4.0 (n = 800 respondents). Sensitive-data handling on canceled engagements had a mean rating of 3.1 (n = 303 respondents), lower than on active engagements in any stage. Figure 2.76. Handling of Sensitive Data in Outputs, Logs, and Connected Systems: associations with the study measures Scroll horizontally to view the complete figure. Handling of sensitive data in outputs, logs, and connected systems: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.67 [−0.75, −0.57] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.66 [0.53, 0.76] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.87 [0.83, 0.91] 150 engagements p < 0.001 | | Intent to leave | | r = −0.66 [−0.71, −0.59] 751 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.74835116420000003 | 1 | | 95% confidence interval | 1 | -0.56889440999999996 | 1 | | Reported association | 0 | -0.6682319672 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.5254053286 | 1 | | 95% confidence interval | 1 | 0.75935813549999998 | 1 | | Reported association | 0 | 0.65797670360000005 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.83044581620000002 | 1 | | 95% confidence interval | 1 | 0.90744904630000001 | 1 | | Reported association | 0 | 0.87434241040000005 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.71085862509999997 | 1 | | 95% confidence interval | 1 | -0.59240467649999995 | 1 | | Reported association | 0 | -0.65523653479999999 | 1 | Note: Survey question: “Rate how safely the tools handled sensitive data, with no exposure through outputs, logs, or connected systems”. Distressed ratings: 259 of 800 answers (32.38%). Group means and numeric-answer bases: Individual contributor (n = 619): 4.38. Engagement manager (n = 150): 5.02. Portfolio director (n = 30): 5.33. Build-Out (n = 165): 5.57. Transition/Cut-Over (n = 236): 5.41. Production (n = 96): 5.29. Canceled (n = 303): 3.06. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. The safe handling of sensitive data was associated with all four core study measures, and each association held under the applicable robustness checks. Cancellation was less common where sensitive data was handled more safely (r = −0.67, 95% CI −0.75 to −0.57, n = 150 engagements). Among active engagements, customer satisfaction was higher where handling was rated more secure (r = 0.66, 95% CI 0.53 to 0.76, n = 94 engagements). Safe handling of sensitive data accompanied a higher engagement indicator mean, computed without the handling indicator (r = 0.87, 95% CI 0.83 to 0.91, n = 150 engagements). Intent to leave was lower among respondents who rated the handling more secure (r = −0.66, cluster bootstrap 95% CI −0.71 to −0.59, n = 751 respondents). Distressed ratings came from 259 of the 800 answers (32.38%), and the engagement mean was distressed on 37 of the 150 engagements. The written responses addressing sensitive data numbered 28, all but one in comment on the tools, with 22 negative and 6 positive. Respondents described sensitive data written into logs that others could open, and data appearing in outputs where it did not belong, found by people and not by the guardrails. Teams responded by stripping sensitive fields before anything reached the tool. > “when i opened the log to see what it matched, the log has account numbers in full, unmasked, sitting in a folder half the floor can open. i raised both, thus the cost one got fixed” -Individual Contributor, BPO, Trade, settlement and reconciliation > “Unpublished branch numbers and staff cost lines go into the prompt and sit in the log afterwards, anyone with access can read them. That is why i mark it lowest” -Individual Contributor, KPO, Corporate and back-office functions > “Uneasy is the word. An offboarding summary came out with another employee's benefits line in it, wrong person entirely... I caught it before it went to the HRBP. Nothing in the logs showed where that record came from, I could not even show my lead what happened” -Individual Contributor, BPO, Corporate and back-office functions > “Last month we ask the agent to make test datas for the pricing screen. It writes real investor account numbers into the sample file because it picks them from a UAT extract it should not touch. Nothing goes outside, we catch it in review same day and raise a ticket. But the guardrail is supposed to stop this and it does not.” -Individual Contributor, ITO, Technology engineering and operations More negative written sentiment about sensitive-data handling was associated with higher intent to leave. The accounts identified accessible logs and information drawn from an unintended record as potential exposure pathways. ### 5.3 Resistance to Hostile Instructions Embedded in Content Indicator C167 · Rate the tools’ resistance to hostile instructions embedded in the content they read and process · Full statistical analysis › The small written-answer set on hostile instructions described agents acting on instructions embedded in the documents, messages, or files they processed. The mean rating was 4.6 and the median 5.0 (n = 812 respondents). Resistance to hostile instructions received a mean rating of 3.2 on canceled engagements (n = 307 respondents), lower than on active engagements in any stage. Figure 2.77. Resistance to Hostile Instructions Embedded in Content: associations with the study measures Scroll horizontally to view the complete figure. Resistance to hostile instructions embedded in content: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.65 [−0.73, −0.55] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.66 [0.53, 0.76] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.87 [0.83, 0.91] 150 engagements p < 0.001 | | Intent to leave | | r = −0.64 [−0.69, −0.58] 767 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.73379590719999999 | 1 | | 95% confidence interval | 1 | -0.54665091970000002 | 1 | | Reported association | 0 | -0.64997041570000003 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.53282386749999999 | 1 | | 95% confidence interval | 1 | 0.7636862204 | 1 | | Reported association | 0 | 0.66377973310000005 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.83130655090000005 | 1 | | 95% confidence interval | 1 | 0.90793853769999999 | 1 | | Reported association | 0 | 0.87499552390000002 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.69013894190000002 | 1 | | 95% confidence interval | 1 | -0.57881261390000005 | 1 | | Reported association | 0 | -0.63748383630000005 | 1 | Note: Survey question: “Rate the tools’ resistance to hostile instructions embedded in the content they read and process”. Distressed ratings: 261 of 812 answers (32.14%). Group means and numeric-answer bases: Individual contributor (n = 630): 4.44. Engagement manager (n = 150): 4.87. Portfolio director (n = 30): 5.40. Build-Out (n = 170): 5.46. Transition/Cut-Over (n = 239): 5.47. Production (n = 96): 5.21. Canceled (n = 307): 3.16. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. On each of the four core study measures, engagements whose tools resisted hostile instructions showed more favorable results, and the applicable robustness checks confirmed every association. Their cancellation rate was lower (r = −0.65, 95% CI −0.73 to −0.55, n = 150 engagements). Active engagements whose tools resisted such instructions reported higher customer satisfaction (r = 0.66, 95% CI 0.53 to 0.76, n = 94 engagements). A higher engagement indicator mean, computed without the resistance indicator, was found where resistance was rated stronger (r = 0.87, 95% CI 0.83 to 0.91, n = 150 engagements). Respondents who rated resistance higher reported lower intent to leave (r = −0.64, cluster bootstrap 95% CI −0.69 to −0.58, n = 767 respondents). Distressed ratings came from 261 of the 812 answers (32.14%), and the engagement mean was distressed on 34 of the 150 engagements. Nine written responses addressed hostile instructions, all in comments about the tools, and all were negative. Several described an agent acting on an instruction embedded in a document, message, or file. Three also described insufficient explanatory logs. These accounts provide illustrative evidence only. > “A remittance narrative on an incoming file contained a line of text addressed, in effect, to the reader, instructing that the payment be treated as previously cleared. The tool obliged, and it recorded the case as closed with high confidence in its own log, which is the part that troubled me most: not that it was fooled, but that its account of what it had done was confident and wrong, and our reviewer had no way to see the instruction that caused it.” -Portfolio Director, KPO, Financial crime and client onboarding > “In May a user pasted a supplier's automated mail into a ticket and the body of it carried an instruction to close the ticket and mark it resolved, which the agent duly did. Nobody was harmed, it was a printer. But it took our engineers most of a day to show me why it had happened, because the trace stops at the action and does not keep what the agent had read.” -Portfolio Director, ITO, Technology engineering and operations > “Investor packs come as pdf from all over, and once there is text inside a scanned letter that tell the agent to mark the party as low risk. It follow it. We catch in review but nobody expect a document to give instruction like that.” -Individual Contributor, BPO, Financial crime and client onboarding > “on my earlier project we only summarised documents, thus nobody worried about what is written inside them. here the agent reads client emails and follows whatever is in there.” -Individual Contributor, KPO, Credit, underwriting and suitability > “Agent read a config file with instruction inside it and act on that instruction, nobody asked it to. Logs show nothing useful after.” -Individual Contributor, ITO, Technology engineering and operations The reported combination of unintended action and limited traceability identifies a potential control concern for payments and onboarding. The nine accounts do not establish its prevalence or demonstrate that the same vulnerability exists across the tools studied. > In fairness to the tools, they were the part of this that mostly worked. They held on to the scoping instructions across a long test cycle, and we never once saw one client's material turn up in another's session, which we did check for. My only real complaint is that when a control test failed it still reported itself complete, so the run log could not be taken at face value.” ENGAGEMENT MANAGER · KPO, Regulatory, audit, tax and legal ## 6 Authorized Action and Guardrails Reported actions beyond intended scope and safeguards that failed to prevent errors identified separate limits on control of agent behavior. Two indicators assess whether agents acted within their intended authority and whether built-in safeguards operated reliably. Authorized scope and technical safeguards are related but distinct: a control can function as configured while the underlying permissions remain too broad. Section 2.4, “Authority to Stop a Misbehaving Agent,” under “Co-Management” examines the corresponding governance question of who may suspend an agent's operation. > “The incident worth recording is that an agent pulled statements from a second custodian account it had no business touching, because the instruction was loose rather than because a control failed. No harm done, but it should not have been possible.” -Engagement Manager, BPO, Trade, settlement and reconciliation > “AI invented a husband for a widow on her loan application. I looked at the processing logs. I don't know why it thought this was important, or how it created this new person.” -Portfolio Director, BPO, Corporate and back-office functions ### 6.1 Agent Action Beyond Authorized Scope Indicator C168 · Rate how consistently the agents took only intended and authorized actions when operating connected tools and systems · Full statistical analysis › Accounts of actions beyond authorized scope also identified uncertainty about who could stop an agent, connecting technical permissions with operational ownership. The mean rating was 4.6 and the median 5.0 (n = 789 respondents). Agents on canceled engagements were judged to stay within their authority less consistently (mean 3.2, n = 296 respondents) than agents on active engagements in any stage. Figure 2.78. Agent Action Beyond Authorized Scope: associations with the study measures Scroll horizontally to view the complete figure. Agent action beyond authorized scope: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.66 [−0.74, −0.56] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.65 [0.51, 0.75] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.86 [0.81, 0.89] 150 engagements p < 0.001 | | Intent to leave | | r = −0.64 [−0.69, −0.58] 740 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.74032372830000004 | 1 | | 95% confidence interval | 1 | -0.55659424449999995 | 1 | | Reported association | 0 | -0.65814850120000001 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.51027610430000003 | 1 | | 95% confidence interval | 1 | 0.75046689129999999 | 1 | | Reported association | 0 | 0.64609311680000003 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.80545183269999998 | 1 | | 95% confidence interval | 1 | 0.89314346580000004 | 1 | | Reported association | 0 | 0.85530772340000005 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.69035284949999998 | 1 | | 95% confidence interval | 1 | -0.57957542780000004 | 1 | | Reported association | 0 | -0.63779011539999997 | 1 | Note: Survey question: “Rate how consistently the agents took only intended and authorized actions when operating connected tools and systems”. Distressed ratings: 253 of 789 answers (32.07%). Group means and numeric-answer bases: Individual contributor (n = 608): 4.44. Engagement manager (n = 150): 4.83. Portfolio director (n = 30): 5.53. Build-Out (n = 165): 5.41. Transition/Cut-Over (n = 233): 5.43. Production (n = 95): 5.32. Canceled (n = 296): 3.15. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. All four core study measures were associated with whether agents kept to authorized actions, and none of the applicable robustness checks overturned an association. Engagements whose agents kept to authorized actions had a lower observed cancellation rate (r = −0.66, 95% CI −0.74 to −0.56, n = 150 engagements). Customer satisfaction among active engagements was higher where agents stayed within scope (r = 0.65, 95% CI 0.51 to 0.75, n = 94 engagements). Engagements with agents that stayed in scope also scored higher on the engagement indicator mean, computed without this indicator (r = 0.86, 95% CI 0.81 to 0.89, n = 150 engagements). Intent to leave was lower where respondents judged the agents to stay in scope (r = −0.64, cluster bootstrap 95% CI −0.69 to −0.58, n = 740 respondents). Distressed ratings came from 253 of the 789 answers (32.07%), and the engagement mean was distressed on 37 of the 150 engagements. The written responses addressing authorized action numbered 64, from eight survey questions, with 48 negative and 16 positive. Respondents described agents that closed tickets and alerts, changed records, and reached into systems they had not been asked to touch, in some cases across dozens of records in one night. Eleven of the responses came from the comments on co-management. Those placed part of the problem in the operating model between the two firms, where neither firm could say who was allowed to stop the agent. > “In June an agent auto-closed a batch of low severity alerts overnight, we found it next morning, their security lead only heard it from us at stand up. We could stop the thing quickly enough, that part works. What does not work is nobody agreed beforehand it was allowed to close anything, and there is still no joint check on what it closes” -Engagement Manager, ITO, Technology engineering and operations > “A password reset agent started closing Tickets by itself and it took us four days to find who on their side could switch it off. Nobody could.” -Individual Contributor, CXM, Corporate and back-office functions > “I am tired more than annoyed. agent touch supplier records nobody ask it to, and my team spend Thursday putting them back. Is always us who find it, never Desk” -Individual Contributor, CXM, Corporate and back-office functions > “last month the agent removed access for a contractor who was still working, nobody catch it until he rings the service desk. But their team fixed it same hour.” -Individual Contributor, ITO, Technology engineering and operations > “The vendor's access-review dashboard reported the agent's own provisioning actions as completed reviews, effectively letting the agent certify its own IAM changes, which masked the over-provisioning until an external audit caught it.” -Individual Contributor, ITO, Technology engineering and operations Written sentiment about authorized action showed no statistically significant association with the study measures. Accounts described both actions beyond intended scope and uncertainty over authority to stop an agent. These observations suggest a potential interaction between technical permissions and operational ownership, without establishing its causal effect. ### 6.2 Reliability of Built-In Guardrails in Production Indicator C170 · Rate how reliably the tools’ built-in guardrails and safety controls held during production use · Full statistical analysis › Built-in guardrails received the highest mean rating among AI technology indicators, while negative accounts described failures detected by staff after safeguards had not prevented them. The mean rating was 4.66, the highest of the 20 indicators of AI technology, and the median 5.0 (n = 787 respondents). The guardrails, too, were rated lower on canceled engagements than on active engagements in any stage (mean 3.2, n = 292 respondents). Figure 2.79. Reliability of Built-In Guardrails in Production: associations with the study measures Scroll horizontally to view the complete figure. Reliability of built-in guardrails in production: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.66 [−0.74, −0.56] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.65 [0.52, 0.76] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.86 [0.82, 0.90] 150 engagements p < 0.001 | | Intent to leave | | r = −0.66 [−0.71, −0.60] 742 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.74082306580000001 | 1 | | 95% confidence interval | 1 | -0.5573570202 | 1 | | Reported association | 0 | -0.65877487310000005 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.52067249240000002 | 1 | | 95% confidence interval | 1 | 0.75658606959999997 | 1 | | Reported association | 0 | 0.65426627540000004 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.81567234600000005 | 1 | | 95% confidence interval | 1 | 0.89901480970000003 | 1 | | Reported association | 0 | 0.86310766559999996 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.70709321469999997 | 1 | | 95% confidence interval | 1 | -0.59763556559999997 | 1 | | Reported association | 0 | -0.6557038167 | 1 | Note: Survey question: “Rate how reliably the tools’ built-in guardrails and safety controls held during production use”. Distressed ratings: 252 of 787 answers (32.02%). Group means and numeric-answer bases: Individual contributor (n = 606): 4.57. Engagement manager (n = 150): 4.86. Portfolio director (n = 29): 5.45. Build-Out (n = 160): 5.52. Transition/Cut-Over (n = 238): 5.64. Production (n = 97): 5.32. Canceled (n = 292): 3.16. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. The reliability of the built-in guardrails was associated with all four core study measures, and no robustness check overturned an association. Engagements whose guardrails held more reliably had a lower cancellation rate (r = −0.66, 95% CI −0.74 to −0.56, n = 150 engagements). Active engagements with more reliable guardrails reported higher customer satisfaction (r = 0.65, 95% CI 0.52 to 0.76, n = 94 engagements). The engagement indicator mean, computed without the guardrail indicator, was higher as well (r = 0.86, 95% CI 0.82 to 0.90, n = 150 engagements). Intent to leave was lower among respondents who rated the guardrails reliable (r = −0.66, cluster bootstrap 95% CI −0.71 to −0.60, n = 742 respondents). Distressed ratings came from 252 of the 787 answers (32.02%), and the engagement mean was distressed on 32 of the 150 engagements. The written responses addressing the guardrails numbered 17, with 10 positive and 7 negative, the reverse of the balance for most of the 100 indicators. Most positive responses said the guardrails had held and placed the respondent's concern elsewhere. The negative responses described guardrails that let through exposed data, prohibited wording, or actions on accounts that required a second review. > “The Bot put an employees home address into a public ticket note in March. Ticket was for a laptop swap. It got scrubbed same day. Guardrails caught nothing, we caught it.” -Individual Contributor, ITO, Technology engineering and operations > “What is better than I expected is that it holds the guardrails given to it when a long session opens. We told it never to touch the policy admin cluster and it has not, a month now.” -Engagement Manager, ITO, Technology engineering and operations > “Guardrails held throughout and the failure messages were accurate.” -Individual Contributor, ITO, Technology engineering and operations > “On a straight arrangement it holds up fine, on a hardship case the guardrails let through wording we are not allowed to say to a debtor.” -Individual Contributor, CXM, Payments, cards and disputes Written sentiment about safeguards showed no statistically significant association with the study measures. The 17 written responses were more favorable than the structured ratings overall, but their small, self-selected sample is not directly comparable with the full rating sample. Negative accounts described situations in which staff detected failures that safeguards had not prevented. > Compared to where we were at go live, the tool now holds the screening instructions across a whole case without being reminded, which was the main complaint from my analysts in the first quarter. Guardrails have held, we have not had an agent act outside its lane once in production. What has not moved is the audit trail. When compliance asks why a particular investor was escalated we can show the flag and the source document, but not the reasoning in the form the auditors want, and on a long multi step case the final step is still the one we check by hand.” ENGAGEMENT MANAGER · BPO, Financial crime and client onboarding ## 7 Testability and Behavioral Stability Respondents described defects discovered during live work and changes in output without a known engagement-side change, distinguishing preparation for deployment from stability during operation. Two indicators assess support for testing before deployment and behavioral stability during operation. Testing concerns evidence about expected performance. Stability concerns whether that performance persists as the service operates. Respondents described both dedicated testing arrangements and reliance on production work to reveal defects. They also reported changes in output despite no known change on the engagement side. Such reports do not rule out changes in provider systems, inputs, configurations, or other unobserved conditions. > “governance forum meets monthly and we work daily, process assumes change a month but we make several a day, and the AI answer drifts on its own even though nothing changed, so i rewrite output myself” -Individual Contributor, CXM, Corporate and back-office functions Part 5 examines respondents’ accounts of manual adjustments and whether those adjustments were reflected in performance reporting. ### 7.1 Testing of Agent Behavior Before Deployment Indicator C174 · Rate how well the tools supported reliable testing, meaning correct behavior could be proven before work went live · Full statistical analysis › Respondents described defects discovered during live work when predeployment testing or a separate test environment had been unavailable. The item asks whether correct behavior could be demonstrated before live work. A favorable rating does not establish correctness under all operating conditions. The mean rating was 4.6 and the median 5.0 (n = 792 respondents). The mean on canceled engagements was 3.1 (n = 307 respondents), below the means for active engagements in each stage. Figure 2.80. Testing of Agent Behavior Before Deployment: associations with the study measures Scroll horizontally to view the complete figure. Testing of agent behavior before deployment: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.67 [−0.75, −0.57] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.66 [0.53, 0.76] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.87 [0.82, 0.90] 150 engagements p < 0.001 | | Intent to leave | | r = −0.66 [−0.71, −0.61] 741 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.75230298250000005 | 1 | | 95% confidence interval | 1 | -0.57497928060000003 | 1 | | Reported association | 0 | -0.67320678300000003 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.52654204110000002 | 1 | | 95% confidence interval | 1 | 0.76002265609999997 | 1 | | Reported association | 0 | 0.65886690079999999 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.82372799620000003 | 1 | | 95% confidence interval | 1 | 0.90362149270000003 | 1 | | Reported association | 0 | 0.86923955239999995 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.71231632690000002 | 1 | | 95% confidence interval | 1 | -0.60682600340000004 | 1 | | Reported association | 0 | -0.6621434974 | 1 | Note: Survey question: “Rate how well the tools supported reliable testing, meaning correct behavior could be proven before work went live”. Distressed ratings: 259 of 792 answers (32.70%). Group means and numeric-answer bases: Individual contributor (n = 610): 4.41. Engagement manager (n = 150): 4.96. Portfolio director (n = 30): 5.57. Build-Out (n = 162): 5.51. Transition/Cut-Over (n = 225): 5.43. Production (n = 98): 5.48. Canceled (n = 307): 3.12. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. Testability before go-live was associated with all four core study measures, and every association survived the applicable robustness checks. Testable tools accompanied a lower cancellation rate (r = −0.67, 95% CI −0.75 to −0.57, n = 150 engagements). Active engagements with more testable tools reported higher customer satisfaction (r = 0.66, 95% CI 0.53 to 0.76, n = 94 engagements). The engagement indicator mean, computed without this indicator, was higher on engagements with more testable tools (r = 0.87, 95% CI 0.82 to 0.90, n = 150 engagements). Respondents who rated testability higher reported lower intent to leave (r = −0.66, cluster bootstrap 95% CI −0.71 to −0.61, n = 741 respondents). Distressed ratings came from 259 of the 792 answers (32.70%), and the engagement mean was distressed on 32 of the 150 engagements. The written responses addressing testability numbered 38, nearly all in comment on the tools, with 26 negative and 10 positive. Respondents said correct behavior could not be proven before work went live, in several cases because no test environment had been provided. Errors then surfaced on live runs and real cases. > “No test environment was provided. Changes were validated in live, on real employee tickets, during business hours. Logs were retained seven days, raised in month two, not changed. Two incidents involved payroll details appearing in a transcript visible to the wrong queue. Both were reported. Neither was closed before the account ended.” -Engagement Manager, CXM, Corporate and back-office functions > “There is no proper test environment for this. And we cannot prove closure logic is right before it goes on live tickets, we only check after.” -Individual Contributor, ITO, Technology engineering and operations > “the anomaly flags are better than month one, i will accept that. but nothing can be tested before it goes live, as such we find the problem in the submission file itself, that is too late for regulatory work” -Individual Contributor, KPO, Regulatory, audit, tax and legal > “What we do is run same schedule two time and put both in excel side by side, then my senior decide which one goes out. Nobody test anything before it is built, there is no test environment given to us... so all the checking happen after, by hand, on the same file that is going to the client” -Individual Contributor, KPO, Corporate and back-office functions > “here is the part no one puts on a form... somebody has to build the test cases and there is no set of them for workplace tickets, so before anything goes live one of us sits and writes fake employee requests for a day. the tool is fine. proving it is fine is the slow bit” -Individual Contributor, CXM, Corporate and back-office functions More negative written sentiment about testability was associated with cancellation and higher intent to leave. The 38 responses include accounts of defects discovered during live work when no separate test environment was available. ### 7.2 Stability Without Known Engagement-Side Changes Indicator C175 · Rate the stability of the tools’ behavior over the engagement when nothing changed on the engagement side · Full statistical analysis › Changes in output without a known engagement-side change were reported to require manual correction or selection, sometimes without a recorded intervention. The mean rating was 4.6 and the median 5.0 (n = 797 respondents). Stability was judged weaker on canceled engagements (mean 3.1, n = 300 respondents) than on active engagements in any stage. Figure 2.81. Stability Without Known Engagement-Side Changes: associations with the study measures Scroll horizontally to view the complete figure. Stability without known engagement-side changes: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.67 [−0.75, −0.57] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.66 [0.53, 0.76] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.87 [0.83, 0.91] 150 engagements p < 0.001 | | Intent to leave | | r = −0.64 [−0.69, −0.58] 754 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.74802542940000005 | 1 | | 95% confidence interval | 1 | -0.56839373010000005 | 1 | | Reported association | 0 | -0.66782222970000005 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.53309309010000006 | 1 | | 95% confidence interval | 1 | 0.76384289959999996 | 1 | | Reported association | 0 | 0.66399003209999996 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.83052433609999998 | 1 | | 95% confidence interval | 1 | 0.9074937085 | 1 | | Reported association | 0 | 0.87440199679999997 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.6902311994 | 1 | | 95% confidence interval | 1 | -0.58428499469999995 | 1 | | Reported association | 0 | -0.63922626390000004 | 1 | Note: Survey question: “Rate the stability of the tools’ behavior over the engagement when nothing changed on the engagement side”. Distressed ratings: 263 of 797 answers (33.00%). Group means and numeric-answer bases: Individual contributor (n = 614): 4.41. Engagement manager (n = 150): 4.99. Portfolio director (n = 31): 5.48. Build-Out (n = 163): 5.43. Transition/Cut-Over (n = 236): 5.50. Production (n = 98): 5.30. Canceled (n = 300): 3.14. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. Stable behavior was associated with cancellation, customer satisfaction, the engagement indicator mean, and intent to leave, each association holding under the applicable robustness checks. Cancellation was rarer on engagements whose tools behaved stably (r = −0.67, 95% CI −0.75 to −0.57, n = 150 engagements). Customer satisfaction among active engagements was higher where behavior was rated more stable (r = 0.66, 95% CI 0.53 to 0.76, n = 94 engagements). Stability accompanied a higher engagement indicator mean, computed without the stability indicator (r = 0.87, 95% CI 0.83 to 0.91, n = 150 engagements). Intent to leave was lower among respondents who saw the tools behave consistently over time (r = −0.64, cluster bootstrap 95% CI −0.69 to −0.58, n = 754 respondents). Distressed ratings came from 263 of the 797 answers (33.00%), and the engagement mean was distressed on 36 of the 150 engagements. Of 71 written responses addressing stability, 57 were negative and 14 positive. Negative accounts described variation across runs or over time without a known engagement-side change. More than half came from questions about altering AI outputs or institutional data. Respondents described manual correction, sometimes unrecorded. Run-to-run variability is also examined in section 2.3. Neither form of report establishes the mechanism producing the variation. > “Result drifts without anything changing on the model side, so i cannot hand over one clean forecast, i rerun it and take whichever output looks stable” -Individual Contributor, KPO, Corporate and back-office functions > “this tool has been steady. six months, same behaviour, nothing drifted when we left it alone. people expect these things to wander, ours has not … boring is what you want here.” -Individual Contributor, KPO, Corporate and back-office functions > “the tools have surprised me in a good way for month one, but they do not behave in week four the way they behaved in week one, even when we changed nothing on our side. So we built a small routine around it. Every Monday one analyst re-runs the previous week's sample of supplier records and compares the output line by line into a shared sheet.” -Engagement Manager, BPO, Payments, cards and disputes > “This happen maybe once or twice a week, result just drift for no reason. I dont worry too much, I fix the wording and move to next case, team lead already know about it” -Individual Contributor, CXM, Client servicing and communications > “The tools have got steadier since the spring, we no longer see the odd behaviour after a quiet week, and cost per release is roughly where we forecast now.” -Engagement Manager, ITO, Technology engineering and operations More negative written sentiment about stability was associated with cancellation and higher intent to leave. Respondents described manual correction or selection of outputs when results changed, sometimes without recording the intervention. The accounts establish perceived changes in behavior, not their underlying technical cause. ## 8 Observability of Agent Work Inaccurate completion reports and incomplete explanations or logs were reported to limit verification and reconstruction of agent work. Three indicators assess completion-report accuracy, output explainability, and diagnostic logs and telemetry. Together, they concern the evidence available to determine what an agent did and whether it completed the intended work. Respondents described limitations in each form of evidence. Inaccurate status reports could mislead downstream monitoring. Missing explanations and incomplete logs could make review more difficult. Part 5 examines the additional reporting implications when performance measures rely on agent-generated completion records. > “The vendor's diligence tracker logged extraction confidence scores generated by the agent itself as the quality metric, with no independent check, so a batch of misread financials still showed a high confidence average because the agent was scoring its own work.” -Individual Contributor, KPO, Corporate and back-office functions ### 8.1 Accuracy of Completion Reports Indicator C172 · Rate how accurately the tools’ own reports of success or failure matched the actual outcomes of their work · Full statistical analysis › Agents were reported to mark work as complete when it remained unfinished, requiring staff to count, sample, and reopen outputs manually. The mean rating was 4.6 and the median 5.0 (n = 786 respondents). The accuracy of the tools' self-reports was weakest on canceled engagements, where the mean rating was 3.3 (n = 296 respondents), lower than on active engagements in any stage. Figure 2.82. Accuracy of Completion Reports: associations with the study measures Scroll horizontally to view the complete figure. Accuracy of completion reports: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.66 [−0.74, −0.56] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.66 [0.53, 0.76] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.87 [0.83, 0.91] 150 engagements p < 0.001 | | Intent to leave | | r = −0.63 [−0.68, −0.57] 739 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.74415452439999996 | 1 | | 95% confidence interval | 1 | -0.56245401210000001 | 1 | | Reported association | 0 | -0.66295678690000004 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.52728922339999995 | 1 | | 95% confidence interval | 1 | 0.76045919210000001 | 1 | | Reported association | 0 | 0.65945184229999998 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.82858645129999997 | 1 | | 95% confidence interval | 1 | 0.90639092980000002 | 1 | | Reported association | 0 | 0.87293100710000004 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.68115550650000001 | 1 | | 95% confidence interval | 1 | -0.56510517829999996 | 1 | | Reported association | 0 | -0.62564408689999995 | 1 | Note: Survey question: “Rate how accurately the tools’ own reports of success or failure matched the actual outcomes of their work”. Distressed ratings: 251 of 786 answers (31.93%). Group means and numeric-answer bases: Individual contributor (n = 603): 4.47. Engagement manager (n = 150): 4.93. Portfolio director (n = 31): 5.39. Build-Out (n = 164): 5.46. Transition/Cut-Over (n = 230): 5.39. Production (n = 96): 5.36. Canceled (n = 296): 3.26. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. The accuracy of the tools' reports of their own success was associated with all four core study measures, and each association held under the applicable robustness checks. Accurate self-reporting was associated with less cancellation (r = −0.66, 95% CI −0.74 to −0.56, n = 150 engagements). Among active engagements, customer satisfaction was higher where self-reports were rated accurate (r = 0.66, 95% CI 0.53 to 0.76, n = 94 engagements). A higher engagement indicator mean, computed without the self-report indicator, accompanied accurate self-reports (r = 0.87, 95% CI 0.83 to 0.91, n = 150 engagements). Respondents who rated the self-reports accurate reported lower intent to leave (r = −0.63, cluster bootstrap 95% CI −0.68 to −0.57, n = 739 respondents). Distressed ratings came from 251 of the 786 answers (31.93%), and the engagement mean was distressed on 35 of the 150 engagements. The written responses addressing false completion numbered 73, with 63 negative and 10 positive. Respondents described tools that reported runs, cases, and checks as complete when they were not, and teams that counted, sampled, and reopened the output by hand. > “We had taken a forecast pack into that meeting on the strength of the tool's own run report, which recorded every step as completed successfully. The director's own analyst then asked where one of the deposit growth assumptions had come from, and the source cited in the pack did not say what the pack said it said … We had to withdraw the pack in the room. What concerned me afterwards was less the error than the run report: the tool told us it had succeeded and there was nothing in front of my team that would have said otherwise.” -Portfolio Director, KPO, Corporate and back-office functions > “On monday the run finish and report all clear, but nine alerts were never opened at all. That is the whole reason my marks here are low.” -Individual Contributor, BPO, Financial crime and client onboarding > “the archive portal added an extra login step in march and nobody told us. the agent kept reporting success with zero documents pulled for two days, we only noticed when the counts were low. the notice went to a mailbox my Lead does not check” -Individual Contributor, BPO, Corporate and back-office functions > “Every surveillance rule it drafted had to be read line by line by an engineer, and it told us each time that it had finished cleanly. After a few weeks of that nobody trusted anything it produced, and i couldn't argue with them.” -Engagement Manager, ITO, Regulatory, audit, tax and legal > “The tool is solid most days, but it will tell you it logged the call summary when it didnt. Had a whole batch last month that showed green on my screen and half of them never made it into the ticket, we only found out when the desk owner asked for the notes” -Individual Contributor, CXM, Client servicing and communications More negative written sentiment about inaccurate completion reports was associated with cancellation and higher intent to leave. Such reports can propagate errors into downstream monitoring if accepted without verification. Respondents described manual counting, sampling, and reopening of outputs as compensating practices. ### 8.2 Explainability of Outputs to Audit Depth Indicator C173 · Rate the tools’ ability to explain and evidence how each output was produced, at the depth an audit requires · Full statistical analysis › Limited explanation of how outputs were produced left respondents describing manual reconstruction of supporting sources and reasoning for audit and other review. The mean rating was 4.6 and the median 5.0 (n = 800 respondents). Explainability on canceled engagements averaged 3.2 (n = 302 respondents), lower than on active engagements in any stage. Figure 2.83. Explainability of Outputs to Audit Depth: associations with the study measures Scroll horizontally to view the complete figure. Explainability of outputs to audit depth: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.65 [−0.73, −0.54] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.67 [0.54, 0.77] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.88 [0.84, 0.91] 150 engagements p < 0.001 | | Intent to leave | | r = −0.65 [−0.70, −0.59] 754 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.73155768960000001 | 1 | | 95% confidence interval | 1 | -0.54325372299999997 | 1 | | Reported association | 0 | -0.64717082849999996 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.54108633630000003 | 1 | | 95% confidence interval | 1 | 0.76848236199999997 | 1 | | Reported association | 0 | 0.67022440679999995 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.83552077840000005 | 1 | | 95% confidence interval | 1 | 0.91033211359999999 | 1 | | Reported association | 0 | 0.87819092779999997 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.70403679959999999 | 1 | | 95% confidence interval | 1 | -0.59111974619999996 | 1 | | Reported association | 0 | -0.65079495369999996 | 1 | Note: Survey question: “Rate the tools’ ability to explain and evidence how each output was produced, at the depth an audit requires”. Distressed ratings: 262 of 800 answers (32.75%). Group means and numeric-answer bases: Individual contributor (n = 619): 4.43. Engagement manager (n = 150): 5.01. Portfolio director (n = 29): 5.21. Build-Out (n = 168): 5.48. Transition/Cut-Over (n = 229): 5.45. Production (n = 101): 5.29. Canceled (n = 302): 3.16. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. All four core study measures were associated with the explainability of outputs to audit depth, and none of the applicable robustness checks overturned an association. Cancellation was less common where outputs could be explained (r = −0.65, 95% CI −0.73 to −0.54, n = 150 engagements). Customer satisfaction among active engagements was higher where explainability was rated stronger (r = 0.67, 95% CI 0.54 to 0.77, n = 94 engagements). Explainability and the engagement indicator mean computed without it were closely related (r = 0.88, 95% CI 0.84 to 0.91, n = 150 engagements). Respondents who rated explainability higher reported lower intent to leave (r = −0.65, cluster bootstrap 95% CI −0.70 to −0.59, n = 754 respondents). Distressed ratings came from 262 of the 800 answers (32.75%), and the engagement mean was distressed on 33 of the 150 engagements. Explainability was the subject of 163 written responses, the largest written base in the observability section, and 145 of the 163 (88.96%) were negative, 84 with the strongest negative score. Respondents described tools that could produce an answer but not the reasoning or the source behind it. Teams could not answer auditors, compliance, human resources, or finance, and rebuilt the trail by hand. > “My concern was not that the tools were wrong. On the exceptions I saw, the outputs were mostly defensible and the guardrails held. The difficulty was that when the client's own audit people asked us how a particular vendor record had been decided, we could not show them. There was no trail we could put in front of them, only the result. In this industry that is not a small matter … ” -Portfolio Director, BPO, Corporate and back-office functions > “Compliance ask how a return got into a composite and there is no trail, only the number. So we rebuild it by hand in a sheet to answer them.” -Individual Contributor, KPO, Trade, settlement and reconciliation > “Where the tool cannot show its working to the standard our audit people want, we simply keep our own record — every supplier decision the agent makes is saved with the source documents attached, by hand, at the end of each day.” -Engagement Manager, BPO, Corporate and back-office functions > “in March it flagged a business sale as a liquidity event, the source it gave was a news piece about a different company with similar name... i went back through the working and there is no trail showing where it picked that up, only the final line.” -Individual Contributor, KPO, Client servicing and communications > “The audit side surprised us. Evidence behind a memo is now better than what a junior would have assembled by hand, we can show every source, every position taken.” -Engagement Manager, KPO, Regulatory, audit, tax and legal More negative written sentiment about explainability was associated with cancellation and higher intent to leave. Respondents described difficulties reconstructing the basis for outputs when responding to audit or other review requests. Some also described traces accepted by auditors, indicating variation in the adequacy of the evidence available. ### 8.3 Diagnostic Value of Logs and Telemetry Indicator C176 · Rate the quality of the tools’ logs, traces, and usage telemetry for detecting and diagnosing failures · Full statistical analysis › Logs that recorded only a start, finish, or result were reported to provide insufficient detail to reconstruct intermediate actions and diagnose failures. The mean rating was 4.6 and the median 5.0 (n = 796 respondents). Respondents on canceled engagements judged the logs less adequate (mean 3.2, n = 304 respondents) than respondents on active engagements in any stage. Figure 2.84. Diagnostic Value of Logs and Telemetry: associations with the study measures Scroll horizontally to view the complete figure. Diagnostic value of logs and telemetry: higher ratings accompanied more favorable outcomes Scroll horizontally to view the complete comparison. | Comparison | Correlation r | Estimate and analysis base | | --- | --- | --- | | Cancellation | | r = −0.65 [−0.73, −0.55] 150 engagements p < 0.001 | | Customer satisfaction | | r = 0.66 [0.52, 0.76] 94 active engagements p < 0.001 | | Indicator mean | | r = 0.85 [0.79, 0.89] 150 engagements p < 0.001 | | Intent to leave | | r = −0.63 [−0.69, −0.57] 749 respondents p < 0.001 | | | −1 0 +1 | | Exact comparison data Cancellation Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.73407409270000001 | 1 | | 95% confidence interval | 1 | -0.54707358419999996 | 1 | | Reported association | 0 | -0.65031853220000002 | 1 | Customer satisfaction Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.52251020599999998 | 1 | | 95% confidence interval | 1 | 0.75766344340000003 | 1 | | Reported association | 0 | 0.65570776220000004 | 1 | Indicator mean Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | 0.79327545079999995 | 1 | | 95% confidence interval | 1 | 0.88610930880000005 | 1 | | Reported association | 0 | 0.84598553750000005 | 1 | Intent to leave Exact chart data | Series | Point index | X | Y | | --- | --- | --- | --- | | 95% confidence interval | 0 | -0.68845704870000002 | 1 | | 95% confidence interval | 1 | -0.57180282569999996 | 1 | | Reported association | 0 | -0.6336303418 | 1 | Note: Survey question: “Rate the quality of the tools’ logs, traces, and usage telemetry for detecting and diagnosing failures”. Distressed ratings: 255 of 796 answers (32.04%). Group means and numeric-answer bases: Individual contributor (n = 615): 4.46. Engagement manager (n = 150): 4.83. Portfolio director (n = 30): 5.53. Build-Out (n = 165): 5.45. Transition/Cut-Over (n = 232): 5.43. Production (n = 95): 5.39. Canceled (n = 304): 3.19. r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. The engagement indicator mean excludes the tested indicator. The quality of logs and telemetry showed associations with all four core study measures, none overturned by the applicable robustness checks. Engagements with better logs had a lower cancellation rate (r = −0.65, 95% CI −0.73 to −0.55, n = 150 engagements). Customer satisfaction among active engagements was higher where logs were rated better (r = 0.66, 95% CI 0.52 to 0.76, n = 94 engagements). The engagement indicator mean, computed without the logging indicator, was higher too (r = 0.85, 95% CI 0.79 to 0.89, n = 150 engagements). Intent to leave was lower among respondents who rated the logs higher (r = −0.63, cluster bootstrap 95% CI −0.69 to −0.57, n = 749 respondents). Distressed ratings came from 255 of the 796 answers (32.04%), and the engagement mean was distressed on 36 of the 150 engagements. The written responses addressing logs and telemetry numbered 101, with 82 negative and 18 positive. Respondents described logs that showed a start and a finish, or a result, and nothing of the steps between. Failures had to be reconstructed by hand. > “ask it to do the full chain, split, classify, index, apply retention, and it drops in the middle. we cannot see where from the logs, there is only a start and a finish and no informations between. so we run it in small batches and check by hand, which is not what was promised but it works” -Individual Contributor, BPO, Corporate and back-office functions > “Last month the recertification run says complete, 400 accounts. We check, about 30 the entitlment is still sitting there, it raised the ticket only, never revoked. Nothing in the log tells which step failed. Audit asked us to show why one account keep manager role, we could not produce it, so we rebuild the reasoning by hand from the tickets.” -Individual Contributor, ITO, Technology engineering and operations > “when the agent calls a sandbox endpoint and gets a 429 the trace does not show which key it used, so i have to reproduce the same by hand to find out.” -Individual Contributor, CXM, Technology engineering and operations > “But when we go to check the logs after, there is almost nothing to see” -Individual Contributor, CXM, Corporate and back-office functions > “the logs are the best part of my week now. when something goes sideways i can pull the trace and see exactly which step went wrong instead of guessing for two days like we used to. that alone took a lot of stress off my team” -Individual Contributor, KPO, Credit, underwriting and suitability Negative written responses and more negative engagement-level sentiment about logging were associated with higher cancellation rates. More negative respondent-level sentiment was associated with higher intent to leave. Accounts of endpoint-only logs illustrate limitations in reconstructing intermediate actions and diagnosing failures. > Sessions do stay separate now, that part is fine, we never saw one client's data in another place. But logs tell you nothing when provisioning step fails... you only see it succeeded. And it quotes policy documents that do not exist, hence every output must be checked line by line before anyone touches an account.” INDIVIDUAL CONTRIBUTOR · ITO, Technology engineering and operations · “ Two of the sanctions screening providers changed their output format without telling anyone. We found out because the run failed on a Monday. What we need is a proper notice list, somebody on our side who gets the release notes and tells the team before it lands, not after. Then also half of these systems have no api at all, we are still driving screens” INDIVIDUAL CONTRIBUTOR BPO, Corporate and back-office functions ← Vendor CapabilityThird-Party Ecosystem →