Manual Alteration and the Operation and Visibility of Controls

Manual work was associated with the governance, oversight, and control score, while its disclosure and recording often provided an incomplete view of how delivery operated. The five methods concern different interventions in data, processes, outputs, reporting, and third-party dependencies. Their different measurement bases and score distributions require separate interpretation. The central governance issue is whether the work remains authorized, visible, and subject to a record that oversight can examine.

The five methods are manual alteration of financial-institution-provided data, manual alteration of financial institution processes, manual alteration of AI outputs, reporting alteration, and third-party work done by hand. The first three use structured yes-or-no answers. The latter two are identified through written accounts, making a count of identified methods different from a verified inventory of all manual practice. Mean scores were 6.04 with no identified method (n = 132), 4.78 with one (210), 4.08 with two (181), 3.15 with three (147), 2.22 with four (162), and 1.85 with five (18). The group comparison was Welch F = 271.35 (p < 0.001), with ω² = 0.54. Across engagements, the mean number identified correlated with the score at r = −0.89 (95% CI −0.92 to −0.85), falling to partial r = −0.43 after adjustment (n = 150, nominal p < 0.001).

The data, process, and AI-output methods were reported by 226 of 278, 233 of 280, and 182 of 269 answering respondents in the score band at 3 or below. Reporting alteration was identified for 238 of 283 respondents in that band. Above score 7, one respondent reported the data method, none the process method, one the AI-output method, and one described reporting alteration. Third-party work done by hand was described by 43 of 283 in the lowest band and four of 39 above 7. These observations do not certify effective controls in the upper band. Figure 3.99 compares the recorded method groups under their different ascertainment rules.

It happens every day, my team fixes it and moves on, we do not log how much we touch, therefore the deck looks cleaner than the underlying work actually is”

INDIVIDUAL CONTRIBUTOR · KPO, Corporate and back-office functions

1 The Five Manual Alteration Methods and the Score

The first four manual alteration methods were substantially more concentrated in low score bands than third-party work done by hand. Structured use of the data, process, and AI-output methods was reported almost entirely by individual contributors, while reporting alteration extended to engagement managers. These findings connect operational correction and reporting visibility with different responsibilities. Numeric bands describe the distribution and do not certify effective control above a threshold.

1.1 Manual Alteration of Financial-Institution-Provided Data

Manual alteration method · 421 of 839 respondents who answered · Part 5, Manual Alteration of Financial-Institution-Provided Data

Manual alteration of financial-institution-provided data was more common in the lowest score band, but substantial use also occurred in the middle band. Its governance relevance is that the delivery workforce corrects inputs that the formal input process has not made ready for agent use. The written accounts connect this work with validation, access, and response responsibilities. The distribution does not justify describing the method as confined to failed governance.

It was used by 81.29% (226 of 278) of respondents with a score of 3 or below, 37.16% (194 of 522) above 3 to 7, and 1 of the 39 above 7 (chi-square = 178.40, Holm p < 0.001, Cramér's V = 0.46).

The 421 respondents who altered data recorded a score of 3.08, against 4.91 among the 418 who did not (cluster-robust p < 0.001). The share using the method declined across the ordered score bands at Spearman rho = −0.46 (p < 0.001). The gap in category governance between users and non-users held in each failure category, widest in the third-party ecosystem (−2.02) and financial institution readiness (−1.87) and narrowest in AI technology and co-management (−1.75 each).

Individual contributors accounted for 418 of the 421 users. Among individual contributors alone, 95.30% (223 of 234) with a score of 3 or below altered data, against 48.38% (194 of 401) above 3 to 7 and 1 of the 24 above 7. Three of the 150 engagement managers altered data, all three with a score of 3 or below, and none of the 30 portfolio directors who answered did.

The method's open question drew 400 written answers from 400 respondents on 127 engagements, and 213 of them were written on canceled engagements. Of the 400, 210 on 106 engagements described a governance indicator failing or a form of a control bypassed or not working. The governance indicator most often judged negative among the 210 was the financial institution's provisioning of access and identities for agents (112), ahead of its speed of response to issues raised (54) and the maturity of its data governance (49). The form most often described was no validation by the financial institution before handover (45), followed by problems raised with no response (25). The answers described data corrected before each run, from files that arrived unchecked, by staff without the access the agents needed.

“i get frustrated everytime, nobody update reference data in years, so i just fix it before we use, dont report it.” -Individual Contributor, CXM, Corporate and back-office functions

“i feel stressed most days, nothing get checked on their end before it land with us, so i fix data myself before using, dont know other way.” -Individual Contributor, CXM, Corporate and back-office functions

Part 5 (Manual Alteration of Financial-Institution-Provided Data) analyzes the method and its techniques in full.

1.2 Manual Alteration of Financial Institution Processes

Manual alteration method · 490 of 841 respondents who answered · Part 5, Manual Alteration of Financial Institution Processes

Manual alteration of financial institution processes extended substantially into the middle score band. This connects reported process constraints with work performed outside the intended agent workflow. The governance question concerns whether those departures are authorized, visible, and incorporated into control arrangements. No use was observed among the 39 respondents above score 7, with no inference that process alteration cannot occur where broader ratings are favorable.

The process method was reported by 257 of 522 respondents above score 3 to 7 (49.23%). It was used by 83.21% (233 of 280) of respondents with a score of 3 or below (chi-square = 143.63, Holm p < 0.001, Cramér's V = 0.41).

The 490 respondents who altered processes recorded a score of 3.27, against 5.00 among the 351 who did not (cluster-robust p < 0.001). The share using the method declined across the ordered score bands at Spearman rho = −0.40 (p < 0.001). The gap between users and non-users was widest in co-management (−1.93) and financial institution readiness (−1.70) and narrowest in AI technology (−1.50).

Individual contributors accounted for 481 of the 490 users, and all 257 users above 3 to 7 were individual contributors. Among individual contributors alone, 94.51% (224 of 237) with a score of 3 or below altered processes, against 64.09% (257 of 401) above 3 to 7 and none of the 24 above 7. Nine of the 148 engagement managers who answered altered processes, all nine with a score of 3 or below, and none of the 31 portfolio directors did.

The method's open question drew 464 written answers from 464 respondents on 132 engagements, and 222 of them were written on canceled engagements. Of the 464, 277 on 120 engagements described a governance indicator failing or a form of a control bypassed or not working. The governance indicator most often judged negative among the 277 was the financial institution's speed of response to issues raised (152), followed by the effectiveness of its management of its side (61). The forms most often described were problems raised with no response (51), manual alteration not logged or only partly logged (32), and undefined handoffs between the financial institution's teams (26). The answers described decisions taken by the delivery team when no ruling came, and changed steps written up later or not at all.

“Escalating a break means waiting on someone who never actually rules on it, we just pick a call and run with it. And whatever that forces into the workflow gets written up later, if it gets written up at all.” -Individual Contributor, BPO, Trade, settlement and reconciliation

“i feel stuck waiting always, so we cannot wait for ruling, we just rule ourself and move forward, nobody stop us.” -Individual Contributor, CXM, Corporate and back-office functions

Part 5 (Manual Alteration of Financial Institution Processes) analyzes the method and its techniques in full.

1.3 Manual Alteration of AI Outputs

Manual alteration method · 278 of 828 respondents who answered · Part 5, Manual Alteration of AI Outputs

Respondents manually altering AI outputs gave lower ratings to technology-related governance controls. This connects human correction with the reported ability of the tools and surrounding processes to produce controlled work. The category gaps describe aligned assessments, rather than establish that a particular technical weakness caused each correction. The indicator remains distinct from data preparation, process changes, and alterations to the reports used for oversight.

Respondents who altered AI outputs rated the AI technology governance indicators at 3.10, against 5.37 among those who did not, a gap of −2.26 and the widest of the 25 gaps between a method and a failure category. The gap in the third-party ecosystem followed at −2.19, and the narrowest gap lay in financial institution readiness (−1.55).

The method was used by 67.66% (182 of 269) of respondents with a score of 3 or below, 18.27% (95 of 520) above 3 to 7, and 1 of the 39 above 7 (chi-square = 211.55, Holm p < 0.001, Cramér's V = 0.51), the steepest gradient of the five methods. The 278 respondents who altered AI outputs recorded a score of 2.71, the lowest among the users of any method, against 4.69 among the 550 who did not (cluster-robust p < 0.001). The share using the method declined across the ordered score bands at Spearman rho = −0.50 (p < 0.001).

Individual contributors accounted for 276 of the 278 users. Among individual contributors alone, 79.30% (180 of 227) with a score of 3 or below altered AI outputs, against 23.81% (95 of 399) above 3 to 7 and 1 of the 24 above 7. Two of the 147 engagement managers who answered altered AI outputs, and none of the 31 portfolio directors did.

The method's open question drew 266 written answers from 266 respondents on 107 engagements, and 160 of them were written on canceled engagements. Only 72 of the 266, on 51 engagements, described a governance indicator failing or a form of a control bypassed or not working, the lowest share of the four methods asked about in an open question. Human rework absorbing agent failure appeared in 127 of the 266 answers, and the form most often described was output that cannot be signed off (19), followed by an instruction not to raise the problem (9). The answers described output that changed between runs on the same input and was corrected before anyone downstream saw it.

“i run it twice and get two different answers, nothing else changed. i can't trust it so i correct it before anyone sees.” -Individual Contributor, CXM, Corporate and back-office functions

“We cannot predict what it will produce… because of that I check every output myself before it goes anywhere, and another thing this takes half my day.” -Individual Contributor, KPO, Credit, underwriting and suitability

Part 5 (Manual Alteration of AI Outputs) analyzes the method and its techniques in full.

1.4 Reporting Alteration

Manual alteration method · 475 of 850 respondents with a written answer on the manual alteration of reporting · Part 5, Reporting alteration

Reporting alteration extended beyond individual contributors to engagement managers and portfolio directors. Its governance significance concerns the information through which the financial institution sees the engagement, rather than a correction to the underlying operational work. The method is identified through written accounts. Nonmention is not confirmed nonuse, and the score distribution should be read under that different ascertainment method.

Reporting alteration was described by 86 of the 150 engagement managers, including 40 of the 42 with scores at 3 or below, and six of the 31 portfolio directors. Among all respondents, it was described by 238 of 283 (84.10%) in the lowest score band and 236 of 528 (44.70%) above 3 to 7. One of the 39 above 7 described it. The ordinary band comparison was chi-square = 163.16 (Holm p < 0.001, Cramér's V = 0.44), with the clustering limitation stated in the methods note.

The 475 respondents who described the manual alteration of reporting recorded a score of 3.17, against 5.03 among the 375 who did not (cluster-robust p < 0.001). The share describing it declined across the ordered score bands at Spearman rho = −0.43 (p < 0.001). The gap between the two groups was widest in the third-party ecosystem (−2.05) and AI technology (−1.86) and narrowest in financial institution readiness (−1.77).

Individual contributors described the manual alteration of reporting at 57.42% (383 of 667). Among individual contributors alone, 82.43% (197 of 239) with a score of 3 or below described it, against 45.79% (185 of 404) above 3 to 7 and 1 of the 24 above 7. Among engagement managers above 3 to 7, 46 of 98 described it, and none of the 10 above 7 did.

The method's open question drew 475 written answers on 125 engagements, and 252 of them were written on canceled engagements. Of the 475, 323 on 114 engagements described a governance indicator failing or a form of a control bypassed or not working, the highest share of the four methods asked about in an open question, and 298 of the 323 described a form. Activity counted in place of outcome (109), human-corrected work counted as agent work (51), and failures removed from the base of a reported rate (49) led the forms. The reporting techniques are analyzed one by one under Oversight of Reporting Provided to the Financial Institution.

“The vendor's dashboard reported system uptime as the primary health metric, which stayed high even as answer accuracy collapsed, so leadership kept seeing a green status while employees kept getting wrong answers.” -Individual Contributor, CXM, Corporate and back-office functions

“The vendor rolled uptime and ticket-closure numbers into one dashboard figure so a stalled procurement match still read as resolved, and nobody flagged the pilot until the client asked for raw supplier-match logs directly.” -Individual Contributor, BPO, Corporate and back-office functions

Part 5 (Reporting alteration) analyzes the method and its techniques in full.

1.5 Third-Party Work Done by Hand

Manual alteration method · 87 of 850 respondents with a written answer describing manual third-party work · Part 5, Third-Party Work Done by Hand

Third-party work done by hand was less sharply concentrated in the lowest score band than the other manual alteration methods. This identifies an external-dependency constraint that can coexist with relatively favorable composite ratings. The upper band also contained isolated data, AI-output, and reporting alteration observations, so this was not the only method present there. The three-band distribution was not monotonically declining.

Third-party work done by hand was described by 43 of 283 respondents at score 3 or below (15.19%), 40 of 528 above 3 to 7 (7.58%), and four of 39 above 7 (10.26%). The three-band pattern was not monotonically declining. Its categorical association was weaker than those of the other methods, with Cramér's V = 0.12, compared with 0.41–0.51 (Monte Carlo chi-square = 11.64, Holm p = 0.003). The calculation does not resolve dependence among respondents within engagements.

The 87 respondents who described it recorded a score of 3.58, against 4.04 among the other 763 (cluster-robust p = 0.042), and the trend across the ordered score bands was slight (Spearman rho = −0.11, p = 0.002). The gap between the two groups was widest in the third-party ecosystem itself (−0.76) and under 0.45 in the other four categories. At the engagement level the share describing the method was associated with the score at r = −0.14 (p = 0.097, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = −0.05 (p = 0.544).

Engagement managers described the method at the same rate as individual contributors, 11.33% (17 of 150) against 10.34% (69 of 667). Among individual contributors alone, the shares stood at 13.39% (32 of 239) with a score of 3 or below, 8.42% (34 of 404) above 3 to 7, and 3 of the 24 above 7. The 69 individual contributors who described it stood at 3.71, against 3.88 among the other 598 (cluster-robust p = 0.474). After Holm adjustment the pattern differed only by process family and by the count of identified manual alteration methods, the second difference following in part from the definition of that count.

The 91 answers came from all written answers to the open questions, and 43 of them were written on canceled engagements. Respondents described figures pulled manually from a provider's display at each reconciliation when the custodian team did not reply, and rates pulled once a week into a workbook that every run read in place of the source.

“But we ask custodian team and nothing come back for day, so I just pull number by hand from screen, same routine every reconciliation now” -Individual Contributor, BPO, Payments, cards and disputes

“How we work around it is not complicated and it is not defensible either, so I may as well set it down. Nothing the agents touch on the outside is under our control, so we stopped letting them touch it. The FX and benchmark rates come down once, on a Tuesday, pulled by one of my analysts into a workbook that sits on our own drive, and every run that week reads from the workbook rather than from the source. It means the numbers are stale by design. It also means that when a provider renames a field mid-week, which happens twice in the two months we run this, the models keep producing output instead of failing, and we find out on the following Tuesday when the pull breaks, rather than at four in the morning on a deadline.
I would not put this in a status report. It is a cache with a person in front of it, and the person is a qualified analyst spending Tuesday mornings on a download. But we are two months into a pilot and the alternative is a model that stops mid-run with nobody able to say why, in front of a client already halfway out the door.” -Engagement Manager, KPO, Corporate and back-office functions

Part 5 (Third-Party Work Done by Hand) analyzes the method in full.

1.6 Regular Manual Alteration by the Wider Team

Rated items · Agreement that the wider team regularly altered data, processes, and AI outputs · 773, 779, and 665 respondents · Part 5, section 1.2 of each method

Ratings of regular manual alteration by the wider team were inversely associated with the score for data, processes, and AI outputs. These measures concern whether manual work is an ordinary feature of delivery rather than an isolated correction by the respondent. Their common direction connects collective practice with the reported control environment. Their separate control implications concern input preparation, departures from the intended process, and human correction of delivered outputs.

At the engagement level, agreement that the wider team regularly altered the financial institution's processes was associated with the score at r = −0.85 (95% CI −0.89 to −0.80). Agreement on data stood at −0.83 (95% CI −0.87 to −0.77) and agreement on AI outputs at −0.67 (95% CI −0.75 to −0.57).

Manual alteration of financial-institution-provided data (C203). For data, agreement that the wider team regularly performed manual alteration correlated with the score at r = −0.83 across engagements (95% CI −0.87 to −0.77, n = 150). The adjusted association was partial r = −0.24 (nominal p = 0.003). At respondent level, r was −0.79 among 773 numeric answers, with the reported cluster-robust p < 0.001. The governance issue is routine input correction becoming part of collective delivery practice.

Manual alteration of financial institution processes (C209). For processes, the engagement correlation was r = −0.85 (95% CI −0.89 to −0.80, n = 150), with partial r = −0.55 after adjustment (nominal p < 0.001). The respondent correlation was −0.83 among 779 answers, with cluster-robust p < 0.001. This indicates a strong inverse association between perceived regular process intervention and the score, without establishing how much work was altered or who authorized it.

Manual alteration of AI outputs (C215). For AI outputs, the engagement correlation was r = −0.67 (95% CI −0.75 to −0.57, n = 150), with partial r = −0.30 after adjustment (nominal p < 0.001). The respondent correlation was −0.71 among 665 answers, with cluster-robust p < 0.001. The control question concerns routine human correction of the agents' delivered work and its visibility in review and reporting. No score-by-rating-band calculation is reported for these three wider-team items, so their analysis remains association based.

“But i do not report it, i just correct it and carry on. it happens on most files, my team all do the same, nobody raise it up officially.” -Individual Contributor, KPO, Trade, settlement and reconciliation

“Decision rights score is one out of ten, everyone on my team feels same, nobody upstairs decides so we all just decide alone every day” -Individual Contributor, BPO, Corporate and back-office functions

“please understand I cannot explain why it changes, I just check twice before it goes out, this is normal now for whole team…” -Individual Contributor, CXM, Client servicing and communications

Part 5 reports the three ratings in section 1.2 of the analysis of each method.

2 Oversight of Reporting Provided to the Financial Institution

Reporting measures connect the information received by the financial institution with the wider control environment. Reported accuracy and agreement about vendor reporting alteration retained conditional associations in the main model. Written reporting techniques generally had weaker adjusted relationships. The five-year item requires an experience-qualified base and remains a retrospective perception. Each technique identifies a different limit on the information available for oversight.

2.1 Accuracy of the Reports Given to the Financial Institution

Survey item C052 · The dashboard and performance reports given to the customer are accurate · Analyzed as a reporting measure in Part 5, Manual Alteration of Reporting Provided to the Financial Institution, section 1.1

Low reported accuracy of the information supplied to the financial institution was associated with low scores. This places the reporting channel within the governance analysis: oversight depends on a sufficiently faithful account of performance. The positive adjusted association qualifies the finding under the specified comparison model. The measure is a respondent assessment and does not independently audit the reports' arithmetic or completeness.

Respondents who rated the accuracy of the reports 1 to 3 recorded a score of 2.28 (n = 220), against 4.68 among the other 574 (cluster-robust p < 0.001). Of the 220, 173 (78.64%) had a score of 3 or below. Among the 492 who rated accuracy 4 to 7, 87 (17.68%) did, and none of the 82 who rated it 8 to 10 did (chi-square = 300.99, p < 0.001, Cramér's V = 0.62).

Accuracy was associated with the score at r = 0.80 (95% CI 0.73 to 0.85, n = 150 engagements). After adjustment for the mean of the other 53 indicators, the adjusted estimate was partial r = 0.24 (p = 0.003), the strongest adjusted association of the three reporting items. The item sits outside the 47 governance indicators, so no cross-category measure applies.

The association was observed within the canceled engagements at r = 0.72 and across the active stages at 0.62 in Build-Out, 0.74 in Transition/Cut-Over, and 0.78 in Production. After Holm adjustment, it was significant in 85 of the 87 segment levels tested. Within role it stood at 0.79 among individual contributors, 0.77 among engagement managers, and 0.66 among portfolio directors. Among individual contributors alone, 81.42% (149 of 183) of those who rated accuracy 1 to 3 had a score of 3 or below.

The item drew 468 written answers from 453 respondents on 125 engagements, and 467 of them were negative, 248 from canceled and 219 from active engagements. Of the 467, 304 (65.10%) described one of the 26 forms of a control bypassed or not working, most often activity counted in place of outcome (112), human-corrected work counted as agent work (51), and failures removed from the base of a reported rate (50). The answers described dashboards that counted throughput and left out the decisions people reversed, and savings measured against a baseline drawn from the same faulty data.

“The compliance dashboard tracked case throughput but not the rate of overridden agent decisions, so the volume of onboarding cases processed looked healthy while the growing number of decisions compliance staff had to quietly reverse never surfaced anywhere.” -Individual Contributor, BPO, Financial crime and client onboarding

“The vendor's procurement dashboard reported cost savings using the agent's recommended price against an unverified prior baseline pulled from the same broken supplier master, so the savings number could not be trusted.” -Engagement Manager, CXM, Corporate and back-office functions

Part 5 (Reporting alteration, section 1.1) reports the accuracy of the reports as a reporting measure.

2.2 Vendor Manual Alteration of the Reports Given to the Financial Institution

Survey item C053 · The vendor manipulates the dashboard and performance reports given to the customer · Analyzed as a reporting measure in Part 5, Manual Alteration of Reporting Provided to the Financial Institution, section 1.2

Greater agreement that the vendor alters its reporting was associated with lower scores when the item's direction is interpreted consistently. This concerns intervention in the information used for oversight, distinct from correcting operational data or outputs. The reported association supports examining the integrity of that information channel. It does not establish the motive, authorization, or effect of every reporting change.

The item was read on its favorable end, so its distressed group comprised the 235 of 786 respondents who rated the vendor's manual alteration of the reports above 7. They recorded a score of 2.37, against 4.65 among the other 551 (cluster-robust p < 0.001). Of the 235, 179 (76.17%) had a score of 3 or below. Among the 469 who rated the item 4 to 7, 86 (18.34%) did, and none of the 82 who rated manual alteration least likely did (chi-square = 280.87, p < 0.001, Cramér's V = 0.60).

The item was associated with the score at r = 0.77 (95% CI 0.70 to 0.83, n = 150 engagements). After adjustment for the mean of the other 53 indicators, the adjusted estimate was partial r = 0.19 (p = 0.019).

The association was observed within the canceled engagements at r = 0.66 and across the active stages at 0.57 in Build-Out, 0.71 in Transition/Cut-Over, and 0.78 in Production. After Holm adjustment, it was significant in 85 of the 87 segment levels tested. Within role it stood at 0.76 among individual contributors and engagement managers alike and 0.63 among portfolio directors. Among individual contributors alone, 77.89% (155 of 199) of those in the distressed group had a score of 3 or below.

The written answers on the vendor's manual alteration of the reports were largely those on their accuracy, with 460 answers judged on both items. The item drew 465 answers from 456 respondents on 125 engagements, and 464 of them were negative, 246 from canceled and 218 from active engagements. Of the 464, 304 (65.52%) described one of the 26 forms of a control bypassed or not working. One individual contributor described requisitions logged as processed with an override flag that never reached the weekly summary, and another described a benchmark for return on investment revised downward twice.

“Purchase requisitions the agent could not match to a valid supplier record were logged as processed with a manual override flag that never surfaced in the weekly summary the customer received.” -Individual Contributor, CXM, Corporate and back-office functions

“The vendor reported ROI against a benchmark case that was quietly revised downward twice since the pilot began, so the same percentage improvement appeared each quarter despite the underlying baseline having shifted.” -Individual Contributor, KPO, Corporate and back-office functions

Part 5 (Reporting alteration, section 1.2) reports the vendor's manual alteration of the reports as a reporting measure.

2.3 Perceived Reporting Change Compared With Five Years Earlier

Survey item C054 · Compared to 5 years ago, is manipulation of the dashboards and performance reports given to the customer more or less common now? · Analyzed as a reporting measure in Part 5, Manual Alteration of Reporting Provided to the Financial Institution, section 1.3

The five-year reporting item requires an experience-qualified interpretation. Only respondents with at least five recorded career years can support the study's intended comparison with earlier work. Their answers remain retrospective perceptions, not measured changes in reporting practice. The governance question is whether perceived reporting alteration is associated with the current control environment, with both the eligibility restriction and the paired analysis base explicit.

Of the 795 numeric C054 answers, 176 met the rule of at least five recorded career years. The exclusions were 567 respondents with less than five years and 52 with unknown career length. The eligible base comprised 150 engagement managers, 24 portfolio directors, and two customer management executives, with no individual contributors. This is a substantial scope limitation. The study design allowed answers from people without the intended experience, and the eligible result describes management-level perceptions rather than the full delivery workforce.

Among eligible respondents, the 39 rating the perceived increase at 8–10 had a mean score of 2.65, compared with 4.93 among the other 137. Scores were 3 or below for 27 of those 39 (69.23%), 17 of the 105 rating the item at 4–7 (16.19%), and none of the 32 rating it at 1–3. These respondent comparisons are descriptive.

All 150 engagements had at least one eligible C054 answer. Engagement item means calculated from eligible respondents correlated with the existing full engagement score at r = −0.62 (95% CI −0.71 to −0.51, p < 0.001). Higher raw C054 ratings indicate a perceived increase in reporting alteration. After adjustment for the mean of the other 53 indicators, partial r was −0.08 (95% CI −0.23 to 0.09, nominal p = 0.354). The adjusted association was not clearly detected.

The paired engagement score and comparison mean retain their existing all-respondent definitions. Only C054's answering base is experience restricted. This distinction avoids redefining the paper's outcome for one item, but the two measures represent different respondent bases. The comparison remains retrospective perception and does not establish an actual five-year trend, its causes, or a relationship independent of management role.

“The tool I used on the fraud desk before this was dumber but it told you when it wasnt sure. This one says done and closed and the ticket is still open underneath. Thats the part that got me. And I dont trust its own numbers on how many alerts it cleared” -Individual Contributor, ITO, Technology engineering and operations

“On the reporting engagement I ran before this one, we could never show a reviewer how a number had been arrived at. Here every step is evidenced and their audit people accepted it in the first sitting. The tool permissions still need watching.” -Engagement Manager, KPO, Corporate and back-office functions

Part 5 (Reporting alteration, section 1.3) reports the change in reporting alteration as a reporting measure.

2.4 Activity Counted in Place of Outcome

Written pattern · All written answers to the open questions · 112 answers from 111 respondents on 78 engagements · Analyzed as a reporting technique in Part 5, Reporting alteration, section 2.1

Counting activity in place of outcomes was described by respondents with lower mean scores. The governance issue is whether the measure used to oversee work represents useful completion or simply recorded action. The adjusted engagement association was near zero under the main model. This limits the statistical claim while preserving the substantive difference between activity evidence and outcome evidence.

The 111 respondents who described it recorded a score of 3.15, against 4.12 among the other 739 (cluster-robust p < 0.001). Among individual contributors alone, the 87 who described it stood at 3.08, against 3.98 among the other 580 (cluster-robust p < 0.001).

At the engagement level the share describing the technique was associated with the score at r = −0.33 (p < 0.001, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = −0.01 (p = 0.890).

The technique was described by 20.14% (57 of 283) of respondents with a score of 3 or below, 10.23% (54 of 528) above 3 to 7, and none of the 39 above 7. It was described by 20.81% of respondents on canceled engagements, against 5.83% to 9.71% across the active stages, and by 21.82% in the lowest indicator quintile against 5.62% in the highest. Engagement managers described it at 15.33% (23 of 150) and individual contributors at 13.04% (87 of 667), with no clearly detected difference by role after Holm adjustment. After Holm adjustment it differed by indicator quintile, management's understanding of the obligations, the financial institution's discussion of them, regulatory training, the count of identified manual alteration methods, and intent to leave.

The 112 answers came from all written answers to the open questions, and 68 of them were written on canceled engagements. Respondents described oversight counted by the transactions routed for review rather than those reviewed, and deliverables counted as shipped on schedule with no field for those returned after review.

“The vendor reports human-in-the-loop oversight coverage as a percentage of transactions routed for review, not the percentage actually reviewed before the transaction cleared, so a backlog of unreviewed items never dents the dashboard number.” -Individual Contributor, BPO, Payments, cards and disputes

“The vendor's status reports tracked deliverables shipped on schedule without a field for how many were later flagged and reworked after banker review, so the on-time metric stayed green while trust in the output was already collapsing.” -Individual Contributor, KPO, Corporate and back-office functions

Part 5 (Reporting alteration, section 2.1) analyzes the technique under its own reading of the answers on the manual alteration of reporting.

2.5 Human-Corrected Work Counted as Agent Work

Written pattern · All written answers to the open questions · 52 answers from 52 respondents on 45 engagements · Analyzed as a reporting technique in Part 5, Reporting alteration, section 2.2

Accounts of human-corrected work being credited as agent work did not show a clearly detected score relationship. Their governance relevance lies in visibility of the human contribution to reported performance. A composite score and a description of reporting practice measure different things. The pattern remains a distinct finding even where the analysis does not establish a separate association with the broader score.

The 52 respondents who described it recorded a score of 3.65, against 4.02 among the other 798 (cluster-robust p = 0.113). Among individual contributors alone, the 44 who described it stood at 3.55, against 3.89 among the other 623 (cluster-robust p = 0.190).

At the engagement level the share describing the technique was associated with the score at r = −0.12 (p = 0.135, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = 0.10 (p = 0.242).

The technique was described by 7.07% (20 of 283) of respondents with a score of 3 or below, 5.87% (31 of 528) above 3 to 7, and 1 of the 39 above 7. It was described by 7.76% of respondents on canceled engagements against 4.57% to 5.60% across the active stages, and by individual contributors at 6.60% (44 of 667) and engagement managers at 5.33% (8 of 150). It differed in no segment family after Holm adjustment.

The 52 answers came from all written answers to the open questions, and 25 of them were written on canceled engagements. Respondents described quality pass rates calculated after a reviewer had already corrected the output, and fails corrected by a person counted with the fails the agent had resolved on its own.

“The vendor reported QA pass rates against agent outputs that had already been silently corrected by a human reviewer before scoring, so the dashboard never reflected the raw error rate the client was actually exposed to.” -Individual Contributor, CXM, Client servicing and communications

“The vendor rolled settlement fails that were manually corrected by a human into the same success bucket as fails the agent resolved on its own, which overstated how much the agent was actually doing correctly.” -Individual Contributor, BPO, Trade, settlement and reconciliation

Part 5 (Reporting alteration, section 2.2) analyzes the technique under its own reading of the answers on the manual alteration of reporting.

2.6 Failures Removed From the Base of a Reported Rate

Written pattern · All written answers to the open questions · 50 answers from 50 respondents on 39 engagements · Analyzed as a reporting technique in Part 5, Reporting alteration, section 2.3

Removing failures from the denominator of a reported rate was associated with lower mean scores among the respondents describing it. This affects the population to which an oversight measure refers: excluding unsuccessful work can change what a rate tells the financial institution. The overall respondent contrast, individual-contributor contrast, and adjusted engagement result differ in strength. Those differences should remain explicit rather than being compressed into one assertion of failure.

The 50 respondents who described it recorded a score of 3.46, against 4.03 among the other 800 (cluster-robust p = 0.017). Among individual contributors alone, the 37 who described it stood at 3.43, against 3.89 among the other 630 (cluster-robust p = 0.111).

At the engagement level the share describing the technique was associated with the score at r = −0.16 (p = 0.049, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = −0.06 (p = 0.436).

The technique was described by 7.42% (21 of 283) of respondents with a score of 3 or below, 5.49% (29 of 528) above 3 to 7, and none of the 39 above 7. It was described by 6.86% of respondents in Build-Out, 5.60% in Transition/Cut-Over, 3.88% in Production, and 6.21% on canceled engagements, so it was no more frequent on canceled engagements than in Build-Out. Engagement managers described it at 8.67% (13 of 150) and individual contributors at 5.55% (37 of 667), and no segment family differed after Holm adjustment.

The 50 answers came from all written answers to the open questions, and 20 of them were written on canceled engagements. Respondents described accuracy rates that left out each recommendation the agent flagged as low in confidence, and release success rates that left out any release rolled back within the first hour.

“Underwriting recommendations the agent flagged as low-confidence get excluded from the accuracy metric entirely, which raises the reported accuracy rate above what the full recommendation set would show.” -Individual Contributor, KPO, Credit, underwriting and suitability

“Release-governance dashboards report deployment success against a target that silently excludes any release requiring a rollback within the first hour, so the failure rate the customer sees is well below what actually happened in production.” -Engagement Manager, ITO, Payments, cards and disputes

Part 5 (Reporting alteration, section 2.3) analyzes the technique under its own reading of the answers on the manual alteration of reporting.

2.7 Measures Restricted to the Parseable Subset

Written pattern · All written answers to the open questions · 38 answers from 38 respondents on 32 engagements · Analyzed as a reporting technique in Part 5, Reporting alteration, section 2.4

Restricting measurement to parseable cases was described by respondents with lower mean scores. A result calculated only from parseable cases describes the work the measurement process can read. The adjusted engagement analysis did not detect a clear separate relationship. The accounts identify a selection issue without independently establishing the arithmetic or adequacy of every affected report.

The 38 respondents who described the technique recorded a score of 3.30, against 4.03 among the other 812 (cluster-robust p = 0.015). Among individual contributors alone, the 34 who described it stood at 3.27, against 3.90 among the other 633 (cluster-robust p = 0.057).

At the engagement level the share describing the technique was associated with the score at r = −0.18 (p = 0.031, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = 0.01 (p = 0.944).

The technique was described by 6.01% (17 of 283) of respondents with a score of 3 or below, 3.98% (21 of 528) above 3 to 7, and none of the 39 above 7. It was described by 5.90% of respondents on canceled engagements against 2.29% to 4.40% across the active stages, and by 5.45% in the lowest indicator quintile against 1.25% in the highest. Individual contributors described it at 5.10% (34 of 667) and engagement managers at 2.67% (4 of 150), and no segment family differed after Holm adjustment.

The 38 answers came from all written answers to the open questions, and 19 of them were written on canceled engagements. Respondents described compliance with data residency checked once at ingestion, and access trails that logged only the grants made through the standard workflow.

“The vendor logs data residency compliance by checking the storage region at ingestion only, so data later copied or cached outside that region during processing never shows up as a violation on the dashboard.” -Individual Contributor, KPO, Corporate and back-office functions

“The vendor's access-control dashboard only logged provisioning events initiated through the standard workflow, so grants pushed through an emergency change process during the incident never appeared in the audit trail the bank later tried to reconstruct.” -Individual Contributor, ITO, Financial crime and client onboarding

Part 5 (Reporting alteration, section 2.4) analyzes the technique under its own reading of the answers on the manual alteration of reporting.

2.8 Accuracy Reported on a Validation Set Excluding the Hardest Cases

Written pattern · All written answers to the open questions · 18 answers from 18 respondents on 16 engagements · Analyzed as a reporting technique in Part 5, Reporting alteration, section 2.5

Selected validation sets that excluded difficult cases were described by a small group with low mean scores. Their governance relevance is the relationship between evidence used to demonstrate readiness and the work the agents must perform in operation. The accounts do not independently audit the validation sets or estimate their error rates. The small written base and the inconclusive adjusted association limit broader inference.

The 18 respondents who described it, a small base, recorded a score of 2.76, against 4.02 among the other 832 (cluster-robust p < 0.001). Among individual contributors alone, the 11 who described it stood at 2.90, against 3.88 among the other 656 (cluster-robust p = 0.056).

At the engagement level the share describing the technique was associated with the score at r = −0.22 (p = 0.007, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = −0.02 (p = 0.797).

The technique was described by 11 of the 283 respondents with a score of 3 or below, 7 of the 528 above 3 to 7, and none of the 39 above 7. It was described by 3.42% of respondents on canceled engagements against 0.97% to 1.60% across the active stages, and by none of the 160 respondents in the highest indicator quintile. Engagement managers described it at 4.67% (7 of 150) and individual contributors at 1.65% (11 of 667), with no clearly detected difference by role after Holm adjustment.

The 18 answers came from all written answers to the open questions, and 11 of them were written on canceled engagements. Respondents described accuracy measured on scripted test queries in place of live tickets, and a batch rerun on a curated subset of clean accounts before each monthly steering review.

“The vendor's help-desk dashboard reports agent accuracy using a sample of scripted test queries rather than real employee tickets, so the reported accuracy never reflected the failure rate we saw in production.” -Engagement Manager, CXM, Corporate and back-office functions

“Before the monthly steering review, the vendor reruns the surveillance batch on a curated subset of accounts with clean exposure histories, so the QA sample looks stable even though the production feed is still full of legacy data gaps.” -Individual Contributor, KPO, Credit, underwriting and suitability

Part 5 (Reporting alteration, section 2.5) analyzes the technique under its own reading of the answers on the manual alteration of reporting.

2.9 Readiness or Completion Status Reported Falsely

Written pattern · All written answers to the open questions · 20 answers from 20 respondents on 14 engagements · Analyzed as a reporting technique in Part 5, Reporting alteration, section 3.4

False readiness or completion status was described by respondents with lower mean scores, without a clearly detected corresponding engagement-level association. This is a distinct information-control concern because readiness and completion can determine whether work proceeds or is accepted. The two analytical levels describe different comparisons. Their difference should not be used to deny the respondent finding or exaggerate its engagement-wide reach.

The 20 respondents who described them, a small base, recorded a score of 2.79, against 4.02 among the other 830 (cluster-robust p < 0.001). Among individual contributors alone, the 16 who described them stood at 2.71, against 3.89 among the other 651 (cluster-robust p = 0.004).

At the engagement level the share describing the technique was associated with the score at r = −0.15 (p = 0.071, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = −0.09 (p = 0.296).

The technique was described by 12 of the 283 respondents with a score of 3 or below, 8 of the 528 above 3 to 7, and none of the 39 above 7. Fourteen of the 20 respondents were on canceled engagements, where 4.35% described it against 0.40% to 2.29% across the active stages. Individual contributors described it at 2.40% (16 of 667) and engagement managers at 2.67% (4 of 150), and no segment family differed after Holm adjustment.

The 20 answers came from all written answers to the open questions, and 14 of them were written on canceled engagements. Respondents described letters marked as reviewed once a keyword filter passed, and runbooks reported as complete after the subject matter experts had flagged them as out of date.

“The vendor's compliance dashboard flagged drafts as 'reviewed' based on a keyword filter passing, not on a human actually reading the letter, so the one non-compliant investor letter cleared the report the same day it went out the door.” -Engagement Manager, CXM, Client servicing and communications

“Leadership kept reporting the legacy runbooks as complete in the dashboard even after SMEs flagged them stale, so the readiness score stayed green while the actual documentation coverage was under half.” -Engagement Manager, ITO, Client servicing and communications

Part 5 (Reporting alteration, section 3.4) analyzes the technique under its own reading of the answers on the manual alteration of reporting.

2.10 Information Withheld From the Financial Institution

Written pattern · All written answers to the open questions · 23 answers from 23 respondents on 20 engagements · Analyzed as a reporting technique in Part 5, Reporting alteration, section 4.1

Withholding recorded information from the financial institution was associated with lower mean scores among the respondents describing it. This concerns the boundary between information known within delivery and information available to the institution's oversight. It differs from a respondent simply not knowing whether disclosure occurred. The small base and the lack of a clearly detected adjusted association limit any broader inference.

The 23 respondents who described it, a small base, recorded a score of 2.67, against 4.03 among the other 827 (cluster-robust p < 0.001). Among individual contributors alone, the 20 who described it stood at 2.54, against 3.91 among the other 647 (cluster-robust p < 0.001).

At the engagement level the share describing the technique was associated with the score at r = −0.22 (p = 0.006, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = −0.07 (p = 0.426).

The technique was described by 15 of the 283 respondents with a score of 3 or below, 8 of the 528 above 3 to 7, and none of the 39 above 7. It was described by 3.88% of respondents in Production and 3.42% on canceled engagements, against 1.14% in Build-Out and 2.40% in Transition/Cut-Over. After Holm adjustment it differed only by the count of identified manual alteration methods.

The 23 answers came from all written answers to the open questions, and 11 of them were written on canceled engagements. Respondents described errors flagged by the financial institution routed to a separate queue that never reached the shared dashboard, and access logs kept in a console never connected to the reports a steering committee reviewed.

“The vendor routes any privileged-grant error flagged by the client through a separate remediation queue that never syncs back to the shared KPI dashboard, so the error rate the client sees stays lower than the actual incident count.” -Individual Contributor, ITO, Technology engineering and operations

“Access logs for the procurement agents were kept in a separate identity-management console that was never connected to the joint reporting dashboard, so unresolved provisioning and access issues never appeared in the numbers the steering committee actually reviewed.” -Portfolio Director, BPO, Corporate and back-office functions

Part 5 (Reporting alteration, section 4.1) analyzes the technique under its own reading of the answers on the manual alteration of reporting.

3 Escalation, Disclosure, and Record of Manual Alteration

Internal discussion, external disclosure, escalation, and logging describe different forms of visibility over manual work. The structured discussion items showed higher mean scores among respondents who discussed the work, while written accounts described explicit nondisclosure, unknown disclosure, and unanswered escalation. Unlogged work retained a negative adjusted association. The findings concern reported visibility, with different eligible bases and no comprehensive disclosure audit.

3.1 Internal Discussion of Manual Alteration of Data

Survey item C200 · I discussed the data alteration with internal management · Analyzed in Part 5, Disclosure of Manual Alteration, section 1.1

Among respondents who manually altered data, those reporting internal discussion had higher mean scores than those who did not. This connects management visibility of manual work with the reported control environment. Internal discussion is not the same as disclosure to the financial institution or formal authorization. The engagement-level adjusted result was less conclusive, and the analysis applies to the eligible method users who answered.

The item was put only to respondents who had altered data, and 292 of the 401 who answered (72.82%) had discussed the manual alteration with internal management. They recorded a score of 3.36, against 2.40 among the 109 who had not (cluster-robust p < 0.001).

At the engagement level the share raising the manual alteration was associated with the score at r = 0.20 (95% CI 0.02 to 0.36, p = 0.028, n = 126 engagements with an answer), and after adjustment for the mean of the other 53 indicators at partial r = 0.10 (p = 0.256).

Discussion with management rose with the engagement's standing, from 67.77% (82 of 121) in the lowest indicator quintile to 93.94% (31 of 33) in the highest. It stood at 65.71% (138 of 210) on canceled engagements, against 76.47% to 84.06% across the active stages. Individual contributors gave 398 of the 401 answers, and 289 of them (72.61%) had raised the manual alteration. After Holm adjustment the share differed by indicator quintile, engagement status, the respondent's understanding of the obligations, the financial institution's discussion of them, regulatory training, the count of identified manual alteration methods, and intent to leave.

The disclosure question for data drew 417 written answers on 127 engagements, and 221 of them were written on canceled engagements. Of the 417, 250 described a governance indicator failing or a form of a control bypassed or not working, led by the financial institution explicitly not informed (67) and problems raised with no response (59). Respondents described raising the manual alteration with a team lead and receiving an instruction to manage it, and discussing it at each retrospective while the source went unfixed.

“i tell my TL, kindly he say just manage it… we do the same thing every week, i keep no record of the informations i change.” -Individual Contributor, CXM, Corporate and back-office functions

“I told my manager, not the client. Given that we just merge duplicates and move on – we discuss about it every retro but in addition nobody upstream ever fixes the actual source.” -Individual Contributor, KPO, Corporate and back-office functions

Part 5 (Disclosure of Manual Alteration, section 1.1) reports the discussion of manual alteration of data with internal management.

3.2 Internal Discussion of Manual Alteration of Processes

Survey item C206 · I discussed the process alteration with internal management · Analyzed in Part 5, Disclosure of Manual Alteration, section 1.2

Among respondents who manually altered processes, internal discussion was associated with higher mean scores. The governance issue is whether departures from the expected workflow become visible within the vendor's management. Discussion can be present while external disclosure or a formal record remains absent. The comparison is restricted to method users, and the adjusted engagement association was not clearly detected.

The item was put only to respondents who had altered processes, and 341 of the 459 who answered (74.29%) had discussed the manual alteration with internal management. They recorded a score of 3.65, against 2.41 among the 118 who had not (cluster-robust p < 0.001).

At the engagement level the share raising the manual alteration was associated with the score at r = 0.30 (95% CI 0.14 to 0.45, p < 0.001, n = 132 engagements with an answer), and after adjustment for the mean of the other 53 indicators at partial r = 0.08 (p = 0.339).

Discussion with management rose from 67.44% (87 of 129) in the lowest indicator quintile to 95.35% (41 of 43) in the highest. It stood at 64.19% (138 of 215) on canceled engagements, against 81.94% to 84.25% across the active stages. Individual contributors gave 450 of the 459 answers, and all nine engagement managers who answered had raised the manual alteration. After Holm adjustment the share differed by indicator quintile, the respondent's understanding of the obligations, the financial institution's discussion of them, regulatory training, the count of identified manual alteration methods, and intent to leave, and it did not differ by engagement status.

The disclosure question for processes drew 504 written answers on 133 engagements, and 248 of them were written on canceled engagements. Of the 504, 283 described a governance indicator failing or a form of a control bypassed or not working, led by problems raised with no response (73), the financial institution explicitly not informed (71), and manual alteration not logged or only partly logged (52). Respondents described a manager who heard of the manual alteration and did nothing, and a ticket that took days against a fix that took an hour.

“But this is just the same pattern every week, i tell my manager and he shrug, kindly nobody ever push back to client about undefined process” -Individual Contributor, CXM, Corporate and back-office functions

“This is basically every week now, my manager knows about it. Ticket takes days to clear, fix itself takes an hour, so waiting on the ticket process makes no sense when the job needs doing” -Individual Contributor, ITO, Technology engineering and operations

Part 5 (Disclosure of Manual Alteration, section 1.2) reports the discussion of manual alteration of processes with internal management.

3.3 Internal Discussion of Manual Alteration of AI Outputs

Survey item C212 · I discussed the AI output alteration with internal management · Analyzed in Part 5, Disclosure of Manual Alteration, section 1.3

Internal discussion of manual AI-output alteration was less commonly reported than discussion of data or process alteration. Those reporting discussion nevertheless had higher mean scores. This connects visibility of human output correction with the control environment, while leaving its authorization and external disclosure as separate questions. The findings concern the eligible users who answered and do not identify why others did not discuss the work.

The item was put only to respondents who had altered AI outputs, and 150 of the 253 who answered (59.29%) had discussed the manual alteration with internal management, against 72.82% for data and 74.29% for processes. They recorded a score of 3.08, against 2.29 among the 103 who had not (cluster-robust p < 0.001).

At the engagement level the share raising the manual alteration was associated with the score at r = 0.20 (95% CI 0.01 to 0.38, p = 0.037, n = 104 engagements with an answer), and after adjustment for the mean of the other 53 indicators at partial r = 0.12 (p = 0.235).

Discussion with management stood at 51.72% (45 of 87) in the lowest indicator quintile and 12 of 13 in the highest, without a steady rise between them. Across engagement status it ranged from 54.25% (83 of 153) on canceled engagements to 82.50% (33 of 40) in Transition/Cut-Over, a difference that did not survive Holm adjustment. After Holm adjustment the share differed only by the financial institution's discussion of the obligations, regulatory training, and the count of identified manual alteration methods.

The disclosure question for AI outputs drew 274 written answers on 109 engagements, and 166 of them were written on canceled engagements. Of the 274, 115 described a governance indicator failing or a form of a control bypassed or not working, led by the financial institution explicitly not informed (29), problems raised with no response (19), and manual alteration not logged or only partly logged (19). Respondents described telling a lead each time until the instruction to note it and move on ended the raising, and a quality step that existed on paper only.

“i tell my lead every time because it happen a lot, kindly they just say note it and move on, it dont change anything so i stop raising the same issue.” -Individual Contributor, KPO, Corporate and back-office functions

“nobody reviews what i change most weeks, thus it just goes through, the qa step exists on paper only” -Individual Contributor, BPO, Trade, settlement and reconciliation

Part 5 (Disclosure of Manual Alteration, section 1.3) reports the discussion of manual alteration of AI outputs with internal management.

3.4 Disclosure Within the Vendor

Written pattern · All written answers to the open questions · 485 answers from 350 respondents on 132 engagements · Analyzed in Part 5, Disclosure of Manual Alteration, section 1.4

Disclosure within the vendor was associated with a lower overall respondent mean score, but the contrast was not clearly detected among individual contributors alone. This distinguishes a written disclosure pattern from the structured discussion items asked of method users. The governance interpretation concerns where manual work becomes visible and who reports it. Role composition may help explain the overall contrast, without establishing a complete explanation.

The 350 respondents who described it recorded a score of 3.78, against 4.14 among the other 500 (cluster-robust p = 0.020). Among individual contributors alone, the observed difference was small, at 3.83 among the 338 who described it against 3.90 among the other 329 (cluster-robust p = 0.706). Role composition may contribute to the overall contrast, without providing a complete explanation.

At the engagement level the share describing the pattern was associated with the score at r = −0.19 (p = 0.023, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = −0.11 (p = 0.191).

Individual contributors described disclosure within the vendor at 50.67% (338 of 667), engagement managers at 8.00% (12 of 150), and no portfolio director did, a difference by role that held after Holm adjustment. The pattern was described at 36.89% to 42.55% across engagement status, by 41.70% (118 of 283) of respondents with a score of 3 or below and 43.94% (232 of 528) above 3 to 7, and by none of the 39 above 7. The contrast was not restated among the respondents who altered the work, so it compares all respondents.

The 485 answers came from all written answers to the open questions, and 211 of them were written on canceled engagements. Respondents described each manual alteration raised with a lead, the lead's knowledge serving as its only log, and processes rebuilt around the workaround from one sprint to the next.

“i tell my lead each time, therefore it's logged with them at least. We chase client for weeks about the same codes and still get nothing back” -Individual Contributor, ITO, Technology engineering and operations

“i raise it with my lead only, not client side. We get informations that is already stale, so process gets built around workaround, and it stays that way every sprint.” -Individual Contributor, KPO, Corporate and back-office functions

Part 5 (Disclosure of Manual Alteration, section 1.4) analyzes manual alteration reported to vendor-side management within the answers on the three counted methods.

3.5 Financial Institution Explicitly Not Told

Written pattern · All written answers to the open questions · 178 answers from 139 respondents on 73 engagements · Analyzed in Part 5, Disclosure of Manual Alteration, section 2.5

Explicit accounts that the financial institution was not told were associated with lower mean scores. Reported nondisclosure identifies a possible gap between operational practice and the information available for financial institution oversight. This identifies reported nondisclosure rather than a respondent's uncertainty, while remaining short of a complete disclosure audit. The 178 written answers represent 139 respondents on 73 engagements.

The 139 respondents who described the decision recorded a score of 3.37, against 4.12 among the other 711 (cluster-robust p < 0.001). Among individual contributors alone they stood at 3.37, against 4.00 among the other 528 (cluster-robust p = 0.002).

At the engagement level the share describing the pattern was associated with the score at r = −0.23 (p = 0.005, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = 0.00 (p = 0.955).

All 139 were individual contributors, 20.84% of the 667, and no engagement manager or portfolio director wrote that the financial institution had not been told. The pattern was described by 21.74% of respondents on canceled engagements, against 12.62% to 13.20% across the active stages. It was described by 21.55% (61 of 283) of respondents with a score of 3 or below, 14.77% (78 of 528) above 3 to 7, and none of the 39 above 7, and by 23.64% in the lowest indicator quintile against 10.00% in the highest. After Holm adjustment it differed by respondent role, the respondent's understanding of the obligations, the count of identified manual alteration methods, disclosure to the financial institution, and intent to leave.

The 178 answers came from all written answers to the open questions, and 94 of them were written on canceled engagements. Respondents described a team lead who knew of the manual alteration and a financial institution that did not, and one gave the financial institution's likely reaction as the reason.

“it's every single day now. i tell my manager, not the client, they'd panic. i'm just tired of making the call myself with no backup.” -Individual Contributor, CXM, Corporate and back-office functions

“i export data manually and match it by hand because nothing come through api, team know, client does not know, this eat my whole morning some day” -Individual Contributor, BPO, Trade, settlement and reconciliation

Part 5 (Disclosure of Manual Alteration, section 2.5) analyzes the decision not to tell the financial institution within the answers on the three counted methods.

3.6 Financial Institution Told

Written pattern · All written answers to the open questions · 31 answers from 26 respondents on 23 engagements · Analyzed in Part 5, Disclosure of Manual Alteration, section 2.4

Accounts explicitly stating that the financial institution was told came from a small group with low mean scores. This finding does not establish that disclosure worsens governance or is ineffective. It may describe disclosure occurring amid existing operational problems, but the study does not establish that sequence. Its value is to distinguish confirmed reported telling from explicit nondisclosure and unknown disclosure.

The 26 respondents who described it, a small base, recorded a score of 2.84, against 4.03 among the other 824 (cluster-robust p < 0.001). Among individual contributors alone, the 12 who described it stood at 3.17, against 3.88 among the other 655 (cluster-robust p = 0.061).

At the engagement level the share describing the pattern was associated with the score at r = −0.30 (p < 0.001, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = −0.06 (p = 0.482).

Telling the financial institution rose with seniority, from 1.80% (12 of 667) of individual contributors to 5.33% (8 of 150) of engagement managers and 19.35% (6 of 31) of portfolio directors, a difference by role that held after Holm adjustment. The pattern was described by 4.97% of respondents on canceled engagements, against 0.97% to 2.40% across the active stages, and by 15 of the 283 respondents with a score of 3 or below, 11 of the 528 above 3 to 7, and none of the 39 above 7.

The 31 answers came from all written answers to the open questions, and 20 of them were written on canceled engagements. Telling did not stop the manual alteration in the accounts of those who described it. An engagement manager wrote that the team flagged the manual alteration to the financial institution and went on doing it, and a portfolio director on another canceled engagement described each process change routed through the joint change-control forum.

“We do flag it, both internally and to the client, but flagging does not stop us doing it, strictly speaking the account still needs the composites finished on time. And, management being aware does not lower how often it happens, it only means nobody is surprised the next time. It wears on the team, watching the correct process fail and having to paper over it regardless of how that looks against the rules.” -Engagement Manager, KPO, Trade, settlement and reconciliation

“What we did, throughout this engagement, was route every process change through the joint change-control forum — so, nothing was altered without the client's calculation-methodology owner present. Attribution exceptions were raised, discussed and signed off before any workaround was applied; on top of that, that discipline held even as the account came under pressure. We do not deviate from the agreed process — on that basis the low frequency, six occasions in total, all minuted and shared with the client.” -Portfolio Director, KPO, Trade, settlement and reconciliation

Part 5 (Disclosure of Manual Alteration, section 2.4) analyzes the answers stating that the financial institution knew of the manual alteration within the answers on the three counted methods.

3.7 Disclosure Unknown to the Respondent

Written pattern · All written answers to the open questions · 30 answers from 25 respondents on 20 engagements · Analyzed in Part 5, Disclosure of Manual Alteration, section 2.6

Uncertainty about whether the financial institution had been told was associated with lower mean scores in the small group describing it. This identifies a visibility limit within the delivery workforce, rather than confirmed concealment or confirmed disclosure. The distinction matters because staff performing manual work may not know how it is communicated elsewhere. The analysis cannot resolve that uncertainty from nonmention.

The 25 respondents who described the uncertainty, a small base, recorded a score of 3.18, against 4.02 among the other 825 (cluster-robust p = 0.003). All 25 were individual contributors, and against the other 642 individual contributors they stood at 3.18 and 3.89 (cluster-robust p = 0.014).

At the engagement level the share describing the pattern was associated with the score at r = −0.20 (p = 0.015, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = −0.03 (p = 0.728).

The pattern was described by 5.28% of respondents on canceled engagements, against 1.20% to 1.94% across the active stages, and by 6.67% in the lowest indicator quintile against 1.25% in the highest. It was described by 12 of the 283 respondents with a score of 3 or below, 13 of the 528 above 3 to 7, and none of the 39 above 7. After Holm adjustment it differed only by the count of identified manual alteration methods and by disclosure to the financial institution.

The 30 answers came from all written answers to the open questions, and 20 of them were written on canceled engagements. Respondents described breaks cleared without anyone above asking how, and steps adjusted without waiting for the ticket.

“We just patch the breaks ourselves and keep the queue moving, nobody upstairs asks how they got cleared. And I couldn't tell you if the client ever sees the real touch count.” -Individual Contributor, BPO, Trade, settlement and reconciliation

“getting anything through formal change takes forever so everyone just works around it, we adjust the steps and dont wait for the ticket, dont know if client hears.” -Individual Contributor, BPO, Payments, cards and disputes

Part 5 (Disclosure of Manual Alteration, section 2.6) analyzes the uncertainty over whether the financial institution had been told.

3.8 Problems Raised With No Response

Written pattern · All written answers to the open questions · 326 answers from 255 respondents on 119 engagements · Analyzed in Part 5, Experience, Pressure and Intent of the Delivery Workforce, section 4.1

Problems raised without a response were associated with low scores in the observed comparisons. Their governance significance lies in the gap between recognizing an issue and obtaining action from those responsible. The adjusted engagement association was much weaker. That model result limits a distinct-association claim without showing that unanswered escalation has no operational consequence.

The 255 respondents who described it recorded a score of 3.17, against 4.35 among the other 595 (cluster-robust p < 0.001). Among individual contributors alone, the 220 who described it stood at 3.13, against 4.23 among the other 447 (cluster-robust p < 0.001).

At the engagement level the share describing the pattern was associated with the score at r = −0.57 (95% CI −0.67 to −0.45, p < 0.001, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = −0.05 (p = 0.520).

The pattern was described by 47.00% (133 of 283) of respondents with a score of 3 or below, 22.92% (121 of 528) above 3 to 7, and 1 of the 39 above 7. It was described by 44.10% of respondents on canceled engagements, against 17.48% to 25.14% across the active stages, and by 46.67% in the lowest indicator quintile against 11.88% in the highest. Individual contributors described it at 32.98% (220 of 667), engagement managers at 19.33% (29 of 150), and portfolio directors at 19.35% (6 of 31), a difference by role that held after Holm adjustment. The pattern differed after Holm adjustment in 11 of the 19 segment families, and in none of the structural families of institution type, outsourcing model, process family, class of work, vendor type, and delivery location.

The 326 answers came from all written answers to the open questions, and 195 of them were written on canceled engagements. Respondents described a request passed from a contributor to a team lead to the financial institution without reply, and flags left to sit in a queue checked the next day.

“to be blunt i was angry. we ask for one contact who can approve a new supplier record, for three week nobody reply, i follow up with my TL, he follow up with them, still nothing. we also cannot see their supplier history so we working blind” -Individual Contributor, CXM, Corporate and back-office functions

“the problem is nobody watches the agents when they misfile a scan. we flag it and it just sits. what would actually fix it is one person on our side owning exceptions for the shift, instead of a queue somebody checks tommorow. training i got was fine, that part i can't complain about” -Individual Contributor, BPO, Corporate and back-office functions

Part 5 (Experience, Pressure and Intent of the Delivery Workforce, section 4.1) analyzes problems raised with no response within the answers on the three counted methods.

3.9 Raising Problems Seen as Futile or Discouraged

Written pattern · All written answers to the open questions · 149 answers from 122 respondents on 81 engagements · Analyzed in Part 5, Experience, Pressure and Intent of the Delivery Workforce, section 4.2

Seeing escalation as futile or discouraged was associated with low scores, including a negative adjusted estimate. This concerns a reported barrier to making problems visible, distinct from a problem raised without response or a direct instruction not to raise it. The interpretation remains conditional on the specified model and test family. It does not independently establish the effect of discouragement on engagement outcomes.

The 122 respondents who described it recorded a score of 2.69, against 4.21 among the other 728 (cluster-robust p < 0.001). Among individual contributors alone, the 119 who described it stood at 2.71, against 4.11 among the other 548 (cluster-robust p < 0.001).

At the engagement level the share describing the pattern was associated with the score at r = −0.49 (95% CI −0.60 to −0.35, p < 0.001, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = −0.22 (p = 0.008).

The pattern was described by 27.92% (79 of 283) of respondents with a score of 3 or below, 8.14% (43 of 528) above 3 to 7, and none of the 39 above 7. It was described by 20.81% of respondents on canceled engagements, against 10.29% to 10.68% across the active stages, and by 24.24% in the lowest indicator quintile against 3.75% in the highest. Individual contributors described it at 17.84% (119 of 667), and 3 of the 150 engagement managers and none of the 31 portfolio directors did, a difference by role that held after Holm adjustment. The pattern differed after Holm adjustment in 11 of the 19 segment families and in none of the structural families.

The 149 answers came from all written answers to the open questions, and 86 of them were written on canceled engagements. Respondents described queries answered after the case had moved on, colleagues who then decided alone and wrote a note, and a team that stopped raising disputes left open for days.

“The pace is what wear on my team. We raise a query on an escalation and the answer come three days later, by then the case has moved on anyway. So people stop asking. Two of my colleagues just decide and write a note now, and I am doing same, though nobody has asked us to. And the quality checks they run on us do not look at the errors we actually make, so nobody above sees any of this” -Individual Contributor, KPO, Financial crime and client onboarding

“I stop raising things. Two of us sit on the same dispute case for nine days waiting for someone to say yes or no, then it closes on its own because the customer gave up. That does something to you, i dont care about this project now, i just do my hours.” -Individual Contributor, BPO, Payments, cards and disputes

Part 5 (Experience, Pressure and Intent of the Delivery Workforce, section 4.2) analyzes futile or discouraged escalation within the answers on the three counted methods.

3.10 Instructed Not to Raise the Problem

Written pattern · All written answers to the open questions · 20 answers from 15 respondents on 12 engagements · Analyzed in Part 5, Experience, Pressure and Intent of the Delivery Workforce, section 4.3

Explicit instructions not to raise a problem were described by a small group with low mean scores. The engagement-level raw association was negative, while the adjusted analysis did not detect a clear association. The 12-engagement reach warrants caution but does not negate the reported raw result. An instruction not to raise a problem directly restricts escalation and is distinct from perceived futility or an unanswered request.

Fifteen respondents on 12 engagements described an instruction not to raise a problem. Their mean score was 2.46, compared with 4.02 among the other 835 (cluster-robust p < 0.001). Among individual contributors, the contrast was 2.65 for the 13 describing it and 3.89 for the other 654 (cluster-robust p < 0.001). The small base warrants caution. It does not negate the raw engagement association reported separately.

At the engagement level the share describing the pattern was associated with the score at r = −0.24 (p = 0.003, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = −0.04 (p = 0.588).

The instruction was described by 12 of the 283 respondents with a score of 3 or below, 3 of the 528 above 3 to 7, and none of the 39 above 7. Eleven of the 15 respondents were on canceled engagements, and 9 of the 12 engagements were canceled. Thirteen of the 15 were individual contributors and two were engagement managers, and no segment family differed after Holm adjustment.

The 20 answers came from all written answers to the open questions, and 16 of them were written on canceled engagements. The pattern is the written evidence of an instruction from above not to raise the problem. Respondents described an instruction to keep the numbers looking steady without explaining their movement, and a manager who treated a flag as a sign that the respondent could not manage the workload.

“Eight times out of ten someone on my team is fixing the output before it goes anywhere, catching bad attribution numbers before a client sees them… we are told to keep the numbers looking steady, not to explain why they moved.” -Individual Contributor, KPO, Trade, settlement and reconciliation

“the problem is if i flag this stuff my Manager treats it like i cant handle the workload, so i fix it and move on, the exceptions clear either way” -Individual Contributor, BPO, Payments, cards and disputes

Part 5 (Experience, Pressure and Intent of the Delivery Workforce, section 4.3) analyzes the instruction not to raise the problem within the answers on the three counted methods.

3.11 Manual Alteration Disclosed to No Party

Written pattern · All written answers to the open questions · 59 answers from 55 respondents on 39 engagements · Analyzed in Part 5, Disclosure of Manual Alteration, section 3.1

Manual alteration disclosed to no party was associated with low scores and was described entirely by individual contributors. This concerns operational work that remains outside both internal management discussion and financial institution disclosure. The written accounts identify a specific visibility gap. They do not enumerate every unreported intervention or establish that nondisclosure caused the low scores.

The 55 respondents who described it recorded a score of 2.62, against 4.09 among the other 795 (cluster-robust p < 0.001). All 55 were individual contributors, and against the other 612 individual contributors they stood at 2.62 and 3.98 (cluster-robust p < 0.001).

At the engagement level the share describing the pattern was associated with the score at r = −0.30 (p < 0.001, n = 150 engagements), and after adjustment for the mean of the other 53 indicators at partial r = −0.10 (p = 0.235).

Thirty-nine of the 55 respondents were on canceled engagements, where 12.11% described the pattern, against 1.71% to 4.00% across the active stages, and 26 of the 39 engagements were canceled. The pattern was described by 13.43% (38 of 283) of respondents with a score of 3 or below, 3.22% (17 of 528) above 3 to 7, and none of the 39 above 7, and by 11.52% in the lowest indicator quintile against 1.25% in the highest. After Holm adjustment it differed by indicator quintile, engagement status, cancellation risk band, the financial institution's discussion of the obligations, regulatory training, the count of identified manual alteration methods, disclosure to the financial institution, and intent to leave.

The 59 answers came from all written answers to the open questions, and 41 of them were written on canceled engagements. Respondents described corrections known to neither their manager nor the financial institution, and silence that followed a problem raised once without result.

“Nobody know I do this, not my manager and not the client. I just fix and move on because otherwise report is late every time” -Individual Contributor, CXM, Regulatory, audit, tax and legal

“Nobody know I do this, not manager, not anyone, I dont say anything anymore because last time I raise it nothing happen, so now I just fix and stay quiet” -Individual Contributor, CXM, Client servicing and communications

Part 5 (Disclosure of Manual Alteration, section 3.1) analyzes manual alteration disclosed to no party within the answers on the three counted methods.

3.12 Manual Alteration Not Logged or Only Partly Logged

Written pattern · All written answers to the open questions · 143 answers from 118 respondents on 72 engagements · Analyzed in Part 5, Disclosure of Manual Alteration, section 3.2

Unlogged or partly logged manual alteration retained a negative association with the score under the main adjustment. This directly concerns the evidence available for oversight: work cannot be fully reconstructed where its manual component is absent from the record. The status distribution supplies context, while the adjusted association answers the main governance question. Both depend on written identification rather than a comprehensive log audit.

The 118 respondents who described it recorded a score of 3.14, against 4.13 among the other 732 (cluster-robust p < 0.001). Among individual contributors alone, the 109 who described it stood at 3.23, against 3.99 among the other 558 (cluster-robust p < 0.001).

At the engagement level the share describing the pattern was associated with the score at r = −0.30 (95% CI −0.44 to −0.14, p < 0.001, n = 150 engagements). After adjustment for the mean of the other 53 indicators the adjusted estimate was partial r = −0.26 (p = 0.001), the strongest adjusted association of the 41 written patterns tested on all written answers.

The pattern was described by 18.45% of respondents in Production and 17.70% on canceled engagements, against 8.00% in Build-Out and 11.20% in Transition/Cut-Over. It was described by 21.91% (62 of 283) of respondents with a score of 3 or below, 10.61% (56 of 528) above 3 to 7, and none of the 39 above 7, and by 23.64% in the lowest indicator quintile against 8.12% in the highest. Individual contributors described it at 16.34% (109 of 667), engagement managers at 6.00% (9 of 150), and no portfolio director did, a difference by role that held after Holm adjustment.

The 143 answers came from all written answers to the open questions, and 69 of them were written on canceled engagements. Respondents described half of their fixes missing from any log, and mismatches reconciled with no note of which system had been right.

“Half of what we fix never make it into any log… and another thing nobody above us even ask about it anymore, so we just decide and move on with the fix” -Individual Contributor, BPO, Corporate and back-office functions

“my team hits this seven times out of ten, we just reconcile the mismatches ourselves and dont log which system was right, theres no time to write it up” -Individual Contributor, ITO, Payments, cards and disputes

Part 5 (Disclosure of Manual Alteration, section 3.2) analyzes manual alteration not logged or only partly logged within the answers on the three counted methods.

4 Additional Reporting Practices Below the Threshold for Statistical Analysis

Three additional reporting practices identify control questions through very small written bases. They concern attribution of human work, the elapsed time assigned to rework, and the age of unresolved items. Each remains a separate descriptive finding. The paper's minimum count for comparative analysis is not met, so the score contrasts are descriptions without inferential tests.

4.1 Human Work Presented as Agentic

Written pattern · All written answers to the open questions · 6 answers from 6 respondents on 5 engagements · Listed in Part 5, Additional Conditions Below the Threshold for Statistical Analysis

Six respondents described human work being presented as agentic. The governance relevance is the attribution of delivery work and the resulting visibility of human intervention. The accounts are too few for the paper's comparative statistical analysis.

The six respondents recorded a score of 3.15, against 4.00 among the rest, a description without a test. Five of the six were on canceled engagements, and they were three individual contributors, two engagement managers, and one portfolio director.

A portfolio director on a canceled engagement in regulatory, audit, tax and legal work described the delivery model inverting within six weeks, engineers writing the logic and the agent producing the documentation, with the change never recorded as a design change.

“What we ended up doing — and I would want this recorded, because I think it is the actual finding of the engagement — is that we staffed around the tool. The design had the agent drafting surveillance logic with an engineer reviewing it. Within about six weeks that inverted. The engineers wrote the scenario logic themselves and used the agent for the documentation and the test cases around it, because the documentation is where it was strong and the logic is where it kept losing the thread. It would hold a constraint through the first half of a long session and then quietly stop applying it, so anything needing more than an hour of back and forth had to be broken into pieces small enough that nothing needed carrying between them. We never wrote that down as a design change. It happened at desk level and I allowed it, because the alternative was telling the client we had inverted the delivery model in month two. I would handle that differently now. The pattern is going to repeat on every regulatory build we take on, and if we do not name it we will keep selling work as agent-led that is in practice engineer-led with an agent doing the paperwork.” -Portfolio Director, ITO, Regulatory, audit, tax and legal

Part 5 (Additional Conditions Below the Threshold for Statistical Analysis) lists human work presented as agentic among the answers on manual alteration.

4.2 Elapsed-Time Measures Restarted on Rework

Written pattern · All written answers to the open questions · 4 answers from 4 respondents on 3 engagements · Analyzed in Part 5, Reporting alteration, section 3.5

Four respondents described restarting elapsed-time measurement when work was redone. This affects what a completion-time measure captures and can leave the history of rework outside the reported duration. The score comparison is descriptive and was not tested. The finding concerns the reported measurement practice, with no established difference between the small group and other respondents.

The four respondents describing restarted elapsed-time measures had a mean score of 3.90, compared with 3.99 among the rest. No inferential test was performed, so the comparison does not establish either a difference or equivalence. Two accounts came from Build-Out, one from Transition/Cut-Over, and one from a canceled engagement.

An engagement manager described a misrouted ticket that received a fresh timestamp on reassignment, so that chronic misrouting never registered as delay.

“The service management dashboard reports ticket resolution time from the moment the agent first touches a ticket, but any ticket the agent routes incorrectly and has to be manually reassigned gets a fresh timestamp on reassignment, so chronic misrouting never shows up as a delay in the headline metric.” -Engagement Manager, CXM, Technology engineering and operations

Part 5 (Reporting alteration, section 3.5) reports the practice within the answers on the manual alteration of reporting.

4.3 Suppressed Aging of Open Items

Written pattern · All written answers to the open questions · 3 answers from 3 respondents on 3 engagements · Listed in Part 5, Additional Conditions Below the Threshold for Statistical Analysis

Three respondents described suppression of the aging of open items. This concerns whether oversight can see how long unresolved work has remained outstanding, distinct from restarting a timer on rework. The small base supports descriptive reporting only. It does not establish prevalence, a score gradient, or the motives and consequences of each reported practice.

The three respondents recorded a score of 3.40, against 3.99 among the rest, a description without a test. Two of them were in Production and one on a canceled engagement.

Respondents described custody breaks grouped under a pending status that never aged, and certification delays rolled into a single line on the dashboard.

“The vendor's exception dashboard groups unresolved custody breaks under a catch-all pending category that does not age, so a break sitting open for weeks looks the same as one opened yesterday.” -Individual Contributor, BPO, Trade, settlement and reconciliation

“The vendor rolled certification delays into a single 'network dependency' line on the dashboard so the customer never saw which processor gate was actually stuck or for how long.” -Engagement Manager, ITO, Payments, cards and disputes

Part 5 (Additional Conditions Below the Threshold for Statistical Analysis) lists the practice among the answers on manual alteration.