Rate the factual reliability of the AI tools' outputs on this engagement before any human review

The mean rating was 4.6 and the median 5.0 (n = 796 respondents). Ratings of 3 or below, the distressed range, made up 32.41% of the answers. The rating was associated with all four study measures. Higher ratings came with less cancellation (r = −0.67, 95% CI −0.75 to −0.57, p < 0.001, n = 150 engagements), higher customer satisfaction among active engagements (r = 0.65, 95% CI 0.52 to 0.76, p < 0.001, n = 94 engagements), a higher engagement indicator mean computed without the indicator (r = 0.87, 95% CI 0.83 to 0.91, p < 0.001, n = 150 engagements) and lower intent to leave (r = −0.65, cluster bootstrap 95% CI −0.70 to −0.60, p < 0.001 with standard errors clustered by engagement, n = 746 respondents). After Holm correction, ratings differed by engagement status and lifecycle stage (Welch F 49.79, Holm p < 0.001, omega squared 0.440), respondent role (Welch F 9.65, Holm p = 0.001, omega squared 0.019) and vendor type (Welch F 4.73, Holm p = 0.002, omega squared 0.023).

The indicator drew 628 written responses from 14 questions, the largest written base in this part. Of these, 548 were negative. The operational questions described outputs that staff corrected before they went on, usually because the same input gave a different result on another run, with the corrections seldom logged and the financial institution often not told. Comments on the tools described invented content, runs reported complete that were not, and logs that could not show why. Written sentiment was associated with customer satisfaction, cancellation at engagement level and intent to leave.

Ratings of 3 or below, the distressed range, accounted for 258 of the 796 answers (32.41%). The engagement mean fell in the distressed range in 30 of 150 engagements.

The primary responses numbered 144, all written in answer to C177, “Add any comments on the agentic AI tools that the ratings in this section do not capture.” The remaining 484 responses, 77.07% of the total, were secondary and came from 13 other questions. That secondary share ranked 7th from the highest among the 71 Part 2 indicators with 30 or more written responses, against a median of 18.92%. Respondents who wrote both a primary and a secondary response numbered 71. Negative responses made up 54.86% of primary responses (79 of 144) and 96.90% of secondary responses (469 of 484).

In answer to C214, 137 of the 141 responses described correcting, rewriting, choosing between or checking the output before it went on, and 128 described output that differed between runs on the same input. Respondents chose the run that looked right, took a blend of several runs, or rewrote the output against the source. Some added that they were told to make the output work and not to raise the problem, or that the change process was too slow to wait for.

In answer to C216, ten responses put the share of outputs they corrected at six to eight in ten. Thirty-two said the corrections were not logged, reported, raised or escalated, and seventeen said their lead knew while the financial institution was not told, or that they did not know what it heard. Seven reported a quality score between one and just above four out of ten. The fifty-one responses to C055 described reporting that kept this rework out of view: tasks counted as agent-handled after a person redid them, accuracy scored only on outputs a reviewer had already corrected or agreed with, resolution counted when the bot replied, and uptime reported as the main status measure while answer accuracy fell.

Responses to C177, the comment field on the tools, were divided. Twelve described content the tool invented, such as a spouse on a loan application, a delivery date, supplier bank details, a section number in the financial institution's own guide, a counterparty and transactions not in the file. Fifteen described runs or tasks reported complete or validated when they were not, and nineteen said the logs could not show why the tool produced a wrong output. Others described the middle of long documents missed, rules forgotten on long sessions and details carried from one account's file into another. The 72 positive responses, the majority to C177, described tools that were accurate on simple work or had improved, often with one of these limits attached. Responses to C156 described errors that neither firm would own or stop, and teams that kept their own lists of what the agent got wrong.

Of the 628 responses, 548 (87.26%, 95% CI 84.09% to 90.11%) were negative and 121 (19.27%) carried the strongest negative score of −10. Positive responses numbered 72 (11.46%), and 8 were neutral. The mean written score was −4.7 (95% CI −5.1 to −4.3, n = 628 responses), and the median was −5.0.

Of the 628 responses, 588 came from respondents who also rated the indicator. Among them, the written score rose with the rating (Spearman rho 0.39, 95% CI 0.31 to 0.47, p < 0.001, n = 588 responses). The written responses corroborated the rating. Of the 511 negative responses with a rating, 309 (60.47%) came from respondents whose rating was in the distressed range.

Respondents who wrote about the indicator and gave a rating numbered 383, with a mean rating of 3.9. The 413 respondents who rated the indicator without writing about it had a mean of 5.3. The difference, −1.4 (95% CI −1.7 to −1.1), excluded zero. Ratings in the distressed range made up 49.35% of the ratings from respondents who wrote (189 of 383) and 16.71% of the ratings from respondents who did not (69 of 413).

The respondents who wrote came disproportionately from the distressed end of the scale. The written evidence therefore gives more weight to the respondents who rated the indicator worst.

Every figure in the segment tables is at respondent level. Segments with fewer than 30 respondents carry no interval. Segments with fewer than 10 report counts only.

Segment families recorded at engagement level were tested on engagement means. Respondent role and vendor type were tested on respondents. Where an expected count fell below 5, the test of the distressed share used Fisher's exact test for a two-by-two table, marked exact, or a Monte Carlo test, marked simulated. After Holm correction, ratings differed by engagement status and lifecycle stage (Welch F 49.79, Holm p < 0.001, omega squared 0.440). Active engagements in Transition/Cut-Over rated highest (mean 5.6, n = 235 respondents) and canceled engagements lowest (mean 3.1, n = 302 respondents). Ratings also differed by respondent role (Welch F 9.65, Holm p = 0.001, omega squared 0.019). Portfolio directors rated highest (mean 5.7, n = 31 respondents) and individual contributors lowest (mean 4.4, n = 613 respondents). Ratings also differed by vendor type (Welch F 4.73, Holm p = 0.002, omega squared 0.023). Consulting firms rated highest (mean 5.2, n = 114 respondents) and BPO specialists lowest (mean 4.1, n = 185 respondents). Outsourcing model, institution type, process family and class of work showed no difference.

Written sentiment differed by engagement status and lifecycle stage across the 136 engagements with a written response (Kruskal-Wallis H 38.93, df 3, p < 0.001). Of the groups with 10 or more responses, responses from canceled engagements were the most negative (mean −5.9, 332 responses) and responses from active engagements in Transition/Cut-Over the least negative (mean −3.0, 144 responses).

The role test compared individual contributors and engagement managers, the roles with 30 or more responses. Written sentiment differed (Kruskal-Wallis H 11.63, p < 0.001, n = 397 respondents): the mean written score was −5.0 for individual contributors (544 responses) and −2.8 for engagement managers (75 responses). Portfolio directors wrote 9 responses, reported as counts only.

Cancellation rate is the share of engagements canceled by the study date, measured at engagement level (150 engagements, 56 canceled). Canceled engagements were rated after cancellation. The association is therefore a descriptive contrast and not a forecast of cancellation.

Across the study set, higher ratings came with less cancellation (r = −0.67, 95% CI −0.75 to −0.57, Holm p < 0.001, n = 150 engagements). The association held after Holm correction in fourteen of the 20 segment levels tested.

Customer satisfaction is the percentage recorded against each engagement. The primary estimate uses the 94 active engagements. A descriptive contrast uses all 150 engagements.

Across the active engagements, higher ratings came with higher customer satisfaction (r = 0.65, 95% CI 0.52 to 0.76, Holm p < 0.001, n = 94 engagements). The association held after Holm correction in ten of the 16 segment levels tested.

The engagement indicator mean is the mean of an engagement's ratings across the rated indicators, measured at engagement level (150 engagements). The indicator is one of the rated indicators. Each comparison therefore uses the mean recomputed without it.

Higher ratings came with a higher engagement indicator mean (r = 0.87, 95% CI 0.83 to 0.91, Holm p < 0.001, n = 150 engagements). The association held after Holm correction in twenty-four of the 24 segment levels tested.

Intent to leave is the answer to C019, “How likely are you to leave your current job in the next 12 months?”, rated 1 to 10. It is measured at respondent level on up to 798 respondents, with the engagement-level form reported alongside.

Respondents who gave higher ratings reported lower intent to leave (r = −0.65, 95% CI −0.69 to −0.61, Holm p < 0.001, n = 746 respondents). The association held after Holm correction in thirty-six of the 37 segment levels tested.

Two comparisons set the engagements with at least one negative response against all other engagements. Of the 20 other engagements, 14 had no written response addressing the indicator. These two comparisons therefore do not isolate written sentiment. Of the 130 engagements with a negative response, 52 (40.00%) were canceled, against 4 of the 20 other engagements (20.00%). The difference did not reach significance (Fisher exact p = 0.135). Customer satisfaction among active engagements differed between the two groups (mean 62.77% on 78 engagements against 70.56% on 16, Welch t −2.50, p = 0.021).

Two correlations test written sentiment directly. Across the 136 engagements with a written response, a more negative engagement mean written score came with cancellation (point-biserial r −0.43, p < 0.001). Across 381 respondents, a more negative respondent mean written score came with higher intent to leave (Spearman rho −0.41, p < 0.001). Of the four comparisons, three reached significance at 0.05.

The four associations were compared in pairs on their engagement-level forms, each oriented so that a higher value is the better outcome, across the 150 engagements with all four measures. After Holm correction, five of the six pairs differed. The association with the engagement indicator mean was the strongest and the association with intent to leave the weakest. The associations with customer satisfaction and intent to leave did not differ (Holm p = 0.437).

Answers clustered within engagements (ICC 0.55). Clustering reduced the 796 respondent answers to the equivalent of about 235 independent answers. Respondent-level intervals therefore used the cluster bootstrap. Respondent-level tests of intent to leave used standard errors that allow for clustering by engagement.

No association was equivalent to zero. Each of the four differed from zero. The detection floor shown is the study floor at 150 engagements. Intent to leave was tested on its effective sample size.

Each change was computed before rounding and may differ by 0.01 from the difference between the two printed values.

All 22 checks held. A check holds when its test falls on the same side of 0.05 as the primary test. Where the primary test falls below 0.05, the check must also keep the sign of the primary correlation. The thin-base check removes the engagements in any level of outsourcing model or process family with fewer than 30 engagements. Institution type is not used for that check, because every one of its levels has fewer than 30 engagements.

Each of the four measures was computed from data the indicator does not contain. Each result is therefore a finding and not an arithmetic consequence. The indicator is one of the rated indicators, and the engagement indicator mean was recomputed without it. The correction changed r by 0.00, from 0.88 to 0.87.

Each deviation was computed before rounding and may differ by 0.01 from the difference between the two printed values.

The comparison covered 150 engagements. The indicator followed the study's typical pattern on all four measures, and no deviation held after Holm correction.

Among the 71 Part 2 indicators with 30 or more written responses, the indicator's share of negative responses, 87.26%, ranked 20th from the highest, against a median of 80.82%. Its mean written score, −4.7, ranked 35th from the most negative, against a median of −4.7.

The indicator most often present in the same responses was C211, “I have altered AI outputs”, in 313 of the 628 responses (49.84%). C159, “Rate the consistency of the tools' outputs when given the same input and instructions more than once.” (152 responses), C212, “I discussed the AI output alteration with internal management” (51 responses), C052, “The dashboard and performance reports given to the customer are accurate.” (50 responses) and C053, “The vendor manipulates the dashboard and performance reports given to the customer.” (49 responses) followed.

The most frequent theme was “Human rework absorbs the failure”, in 307 responses (48.89%). The second was “Occurs every cycle or most days”, in 106 responses.

About a third of all ratings of the factual reliability of the AI tools' outputs before human review fell in the distressed range (258 of 796 answers, 32.41%). The rating was associated with all four study measures, and every association held under each of the 22 robustness checks. Canceled engagements rated far below the three active stages (mean 3.1, n = 302 respondents).

The written evidence is the largest for any indicator in this part and points one way. Outputs reached the financial institution correct because delivery staff corrected them, and the corrections were seldom recorded or reported, while the vendor's measures counted the corrected output as the agent's. The reliability the financial institution saw was therefore partly the work of people it did not see. Written sentiment was associated with lower customer satisfaction among active engagements, with cancellation at engagement level and with higher intent to leave.

The evidence is the reported experience of the vendor's delivery teams, collected in a single cross-section. The associations reported here are not causal.

Tests by respondent role and by vendor type treated respondents as independent. The section on clustering reports how far clustering within engagements reduced the information in respondent-level figures.

Customer management executives gave 2 answers, a base of anecdote strength.

Shares of written sentiment are shares of responses and not of the study set of respondents. The written score takes five fixed values, so its mean and median are coarse. Responses clustered within engagements (628 responses on 136 engagements). Intervals for written sentiment therefore used a bootstrap by engagement.