Rate the stability of the tools' behavior over the engagement when nothing changed on the engagement side

The mean rating was 4.6 and the median 5.0 (n = 797 respondents). Ratings of 3 or below, the distressed range, made up 33.00% 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 = 0.66, 95% CI 0.53 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 = −0.64, cluster bootstrap 95% CI −0.69 to −0.58, p < 0.001 with standard errors clustered by engagement, n = 754 respondents). After Holm correction, ratings differed by engagement status and lifecycle stage (Welch F 52.03, Holm p < 0.001, omega squared 0.434), respondent role (Welch F 9.75, Holm p < 0.001, omega squared 0.017) and vendor type (Welch F 5.42, Holm p < 0.001, omega squared 0.028).

Of the 71 written responses, 57 were negative and 14 positive. Nearly every negative response described output that changed from one run, week or month to the next with nothing changed on the engagement side. More than half of all responses came from the questions asking why respondents altered AI outputs or data provided by the financial institution. Those respondents said they corrected the drift by hand before it went out, in some cases without reporting it. Across engagements, a more negative engagement mean written score came with cancellation, and respondents whose written responses were more negative reported higher intent to leave.

Ratings of 3 or below, the distressed range, accounted for 263 of the 797 answers (33.00%). The engagement mean fell in the distressed range in 36 of 150 engagements.

The primary responses numbered 30, 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 41 responses, 57.75% of the total, were secondary and came from 5 other questions. That secondary share ranked 16th 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 0. Negative responses made up 53.33% of primary responses (16 of 30) and 100.00% of secondary responses (41 of 41).

Thirty-seven responses came from four questions: C214, on why the respondent altered AI outputs, C202, on why the respondent altered data provided by the financial institution, and the comment fields that followed each. Every one was negative, and they gave drift as the reason for altering the output: same input, same release and same configuration, a different result the next run or the next day. Respondents described correcting the output before it reached a report, an advisor, a relationship manager or a submission, every run in several cases. Three described choosing between runs: taking whichever version looked closest to last time, reverting to the last version that passed, or rerunning and keeping whichever output looked stable. Five said the correction went unreported or unreviewed, one because the team was measured on throughput, one because the team was told to make it work and not to raise it.

Responses to C177 described the same drift in operational terms: a ticket type closed one week and held the next, a portfolio dropped from a composite one week and kept the next, a role mapping request approved one day and flagged the next, a fee question answered correctly on Monday and wrongly on Wednesday. Four described standing routines kept because of it, such as an analyst re-running the previous week's sample every Monday and comparing it line by line. Two respondents asked for a model version they could hold fixed, with warning of updates, and two contrasted the tools with older bots that made the same mistake every day and could therefore be corrected once. Six responses described governance that learned of drift from the queue, had no forum for it, or met monthly while the behavior changed between meetings.

The fourteen positive responses said behavior had held steady, in several cases for four to ten months, and one said the behavior settled by month six. Two qualified the judgment: one respondent used the tools for a single narrow task, and another attributed the instability the team saw to the platform beneath the agent rather than the agent itself.

Of the 71 responses, 57 (80.28%, 95% CI 70.77% to 88.89%) were negative and 3 (4.23%) carried the strongest negative score of −10. Positive responses numbered 14 (19.72%), and no response was neutral. The mean written score was −3.1 (95% CI −4.1 to −2.0, n = 71 responses), and the median was −5.0.

Of the 71 responses, 66 came from respondents who also rated the indicator. Among them, the written score rose with the rating (Spearman rho 0.58, 95% CI 0.39 to 0.73, p < 0.001, n = 66 responses). The written responses corroborated the rating. Of the 52 negative responses with a rating, 31 (59.62%) came from respondents whose rating was in the distressed range.

Respondents who wrote about the indicator and gave a rating numbered 62, with a mean rating of 4.0. The 735 respondents who rated the indicator without writing about it had a mean of 4.6. The difference, −0.6 (95% CI −1.2 to 0.0), included zero. Ratings in the distressed range made up 46.77% of the ratings from respondents who wrote (29 of 62) and 31.84% of the ratings from respondents who did not (234 of 735).

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 52.03, Holm p < 0.001, omega squared 0.434). Active engagements in Transition/Cut-Over rated highest (mean 5.5, n = 236 respondents) and canceled engagements lowest (mean 3.1, n = 300 respondents). Ratings also differed by respondent role (Welch F 9.75, Holm p < 0.001, omega squared 0.017). Portfolio directors rated highest (mean 5.5, n = 31 respondents) and individual contributors lowest (mean 4.4, n = 614 respondents). Ratings also differed by vendor type (Welch F 5.42, Holm p < 0.001, omega squared 0.028). Consulting firms rated highest (mean 5.4, n = 112 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 did not differ by engagement status and lifecycle stage across the 56 engagements with a written response (Kruskal-Wallis H 6.69, df 3, p = 0.083). Of the groups with 10 or more responses, responses from canceled engagements were the most negative (mean −4.6, 34 responses) and responses from active engagements in Transition/Cut-Over the least negative (mean −1.4, 18 responses). Responses from active engagements in Production (9) are reported as counts only.

No test across respondent roles was run, because fewer than two roles had 30 or more written responses. Individual contributors wrote 57, engagement managers wrote 10 and portfolio directors wrote 4 responses.

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 = 0.66, 95% CI 0.53 to 0.76, Holm p < 0.001, n = 94 engagements). The association held after Holm correction in eleven 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 = −0.64, 95% CI −0.68 to −0.59, Holm p < 0.001, n = 754 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 104 other engagements, 94 had no written response addressing the indicator. These two comparisons therefore do not isolate written sentiment. Of the 46 engagements with a negative response, 22 (47.83%) were canceled, against 34 of the 104 other engagements (32.69%). The difference did not reach significance (Fisher exact p = 0.099). Customer satisfaction among active engagements did not differ between the two groups (mean 66.46% on 24 engagements against 63.29% on 70, Welch t 1.17, p = 0.249).

Two correlations test written sentiment directly. Across the 56 engagements with a written response, a more negative engagement mean written score came with cancellation (point-biserial r −0.30, p = 0.023). Across 59 respondents, a more negative respondent mean written score came with higher intent to leave (Spearman rho −0.31, p = 0.019). Of the four comparisons, two 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 = 0.400).

Answers clustered within engagements (ICC 0.57). Clustering reduced the 797 respondent answers to the equivalent of about 231 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, 80.28%, ranked 37th from the highest, against a median of 80.82%. Its mean written score, −3.1, ranked 60th 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 26 of the 71 responses (36.62%). C157, “Rate the factual reliability of the AI tools' outputs on this engagement before any human review.” (25 responses), C159, “Rate the consistency of the tools' outputs when given the same input and instructions more than once.” (13 responses), C176, “Rate the quality of the tools' logs, traces, and usage telemetry for detecting and diagnosing failures.” (8 responses) and C161, “Rate the predictability of the tools' token consumption for a given volume of work.” (5 responses) followed.

The most frequent theme was “Human rework absorbs the failure”, in 22 responses (30.99%). The second was “Occurs every cycle or most days”, in 13 responses.

About a third of all ratings of the stability of the tools' behavior fell in the distressed range (263 of 797 answers, 33.00%). 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 = 300 respondents).

The written evidence links unstable tool behavior to a control problem for the financial institution. When the same input gave a different result, staff corrected or selected the output by hand, and much of that correction went unrecorded. The numbers that reached reports, advisors and submissions were therefore partly the product of undocumented human adjustment, and the governance forums met on a cadence slower than the drift they were meant to govern. Written sentiment was associated with cancellation and with 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 (71 responses on 56 engagements). Intervals for written sentiment therefore used a bootstrap by engagement.