Rate the customer’s capacity to govern and oversee the risks of AI agents doing this work

Ratings of the financial institution’s capacity to govern and oversee the risks of AI agents doing the work (mean 3.5, median 3.0, n = 797 respondents) were associated with all four study measures. Engagements with higher ratings were canceled less often (r = −0.57, 95% CI −0.67 to −0.45, p < 0.001, n = 150 engagements). Higher ratings came with higher customer satisfaction on active engagements (r = 0.75, 95% CI 0.64 to 0.83, p < 0.001, n = 94 active engagements). Ratings rose with the engagement indicator mean, computed with this item left out (r = 0.83, 95% CI 0.77 to 0.88, p < 0.001, n = 150 engagements). Respondents who gave higher ratings reported lower intent to leave (r = −0.61, cluster bootstrap 95% CI −0.67 to −0.54, p < 0.001, n = 750 respondents). Ratings differed after Holm control by engagement status and lifecycle stage (Welch F 28.99, Holm p < 0.001, omega squared 0.318) and by respondent role (Welch F 7.81, Holm p 0.005, omega squared 0.015).

The indicator had 45 written responses coded to it, from 44 of 850 respondents on 38 of 150 engagements. Negative responses numbered 37 of the 45 scored responses (82.22%). The strongest negative score, −10, was carried by 18 of the 45 scored responses (40.00%). Secondary responses, written in answer to questions outside the indicator’s own survey section, numbered 37 of the 45 responses (82.22%). Respondents who wrote and respondents who did not rated the indicator equally to one decimal (mean 3.5, n = 44 respondents, against mean 3.5, n = 753 respondents).

The study set mean was 3.5 on 797 answering respondents, with a median of 3.0. 417 answers, 52.32% of the 797, were at the distressed end. A distressed engagement mean was recorded in 71 of 150 engagements.

Of the 45 responses, 8 (17.78%) were primary and 37 (82.22%) were secondary. The secondary responses came from 37 respondents on 32 engagements, in answer to 5 questions outside the indicator’s own survey section (most often C029 with 30 responses, C030 with 3 responses and C208 with 2 responses). One respondent wrote both a primary and a secondary response. Among the Part 2 indicators with 30 or more scored responses, the secondary share of the indicator ranked 4 of 71 from the highest. The median secondary share was 18.92% across the 71 indicators.

Negative responses made up 34 of the 37 secondary responses (91.89%), against 3 of the 8 primary responses. The difference reached significance at 0.05 (Fisher exact p = 0.002, n = 45 responses).

Of the 45 scored responses, 37 (82.22%) were negative, 18 (40.00%) carried the strongest negative score and 7 (15.56%) were positive. The responses had a mean written score of −5.3 (95% CI −6.9 to −3.5, n = 45 responses on 38 engagements).

Respondents who rated the indicator 3 or below wrote 23 of the 37 negative responses with a rating. Written score and rating showed a moderate positive correlation that reached significance at 0.05 (Spearman rho 0.32, 95% CI 0.02 to 0.57, p = 0.033, n = 45 responses on 38 engagements). Its interval excluded zero. Where a written score diverges from the same respondent’s rating, the rating is the measure of the indicator.

Respondents who wrote about the indicator and respondents who did not rated it equally to one decimal (mean 3.5, n = 44 respondents, against mean 3.5, n = 753 respondents). Ratings of 3 or below came from 24 of the 44 respondents who wrote and gave a rating (54.55%), against 393 of the 753 respondents who did not write and gave a rating (52.19%). Respondents who wrote formed a self-selected group and not a random sample of the answering base, because written responses were voluntary.

Figures are respondent level. Means are on the goodness view. A level under 30 prints no interval and a level under 10 prints counts only.

Engagement families were tested on engagement means, and respondent role and vendor type were tested on respondents. A distressed p marked exact is a Fisher exact test for a two by two table, and one marked simulated is a Monte Carlo test run until its p is stable to the third decimal, both used where an expected count is under 5. Engagement status and lifecycle stage differed after Holm control (Welch F 28.99, Holm p < 0.001, tested on engagement means). In Descriptive Statistics by Segment, Active, Production had the highest mean, 4.4 (n = 98 respondents), and Canceled the lowest, 2.3 (n = 306 respondents). Respondent role differed after Holm control (Welch F 7.81, Holm p 0.005, tested on respondents). In Descriptive Statistics by Segment, Portfolio Director had the highest mean, 4.5 (n = 30 respondents), and Individual Contributor the lowest, 3.4 (n = 615 respondents). No difference held after Holm control for outsourcing model, institution type, process family, class of work and vendor type.

Scored responses came from 5 engagements in the Active, Build-Out group, 10 engagements in the Active, Transition/Cut-Over group, 6 engagements in the Active, Production group and 17 engagements in the Canceled group. A test of mean written scores across the groups, with each engagement counted once, showed no difference that reached significance at 0.05 (Kruskal-Wallis H 5.21, df 3, p = 0.157, n = 38 engagements). Groups under 10 responses are reported as counts only. Negative responses numbered 3 of 5 in the Active, Build-Out group and 6 of 7 in the Active, Production group.

Across respondent roles, no test was run, because fewer than two roles reached 30 scored responses. Roles under 10 responses are reported as counts only. Among engagement managers, negative responses numbered 4 of 8. Among portfolio directors, negative responses numbered 3 of 4.

The share of engagements canceled on the study date, measured at engagement level (base 150 engagements, 56 canceled). The association is a descriptive contrast, inflated because canceled engagements were rated after cancellation, and is not predictive.

The study set r was −0.57 (95% CI −0.67 to −0.45, Holm p < 0.001, n = 150 engagements). Of the 20 tested segment levels, 12 held after Holm control.

The satisfaction score recorded against each engagement on a 100-point scale. The primary estimate uses the 94 active engagements. A descriptive contrast uses all 150 engagements.

The study set r was 0.75 (95% CI 0.64 to 0.83, Holm p < 0.001, n = 94 engagements). Of the 16 tested segment levels, 15 held after Holm control.

The mean of an engagement’s 100 indicator ratings, measured at engagement level (base 150 engagements). Where this indicator is one of the 100, the mean is recomputed without it.

The study set r was 0.83 (95% CI 0.77 to 0.88, Holm p < 0.001, n = 150 engagements). Of the 24 tested segment levels, 24 held after Holm control.

The respondent’s stated likelihood of leaving the current job within 12 months, rated 1 to 10 (C019). Measured at respondent level on up to 798 respondents, with the engagement-level form reported alongside.

The study set r was −0.61 (95% CI −0.65 to −0.56, Holm p < 0.001, n = 750 respondents). Of the 37 tested segment levels, 35 held after Holm control.

The cancellation and customer satisfaction contrasts compared engagements with a negative written response against all other engagements. The comparison group comprised 118 other engagements, of which 112 had no scored written response coded to the indicator and 6 had only scored responses that were not negative. The two contrasts therefore did not test written sentiment by themselves. Cancellation was recorded for 15 of the 32 engagements with a negative response (46.88%), against 41 of the 118 other engagements (34.75%). The difference did not reach significance at 0.05 (Fisher exact p = 0.222). Mean customer satisfaction was 60.88% on the 17 active engagements with a negative response, against 64.81% on the 77 other active engagements. The difference did not reach significance at 0.05 (Welch t −1.13, p = 0.270). The association of written sentiment itself with the study measures was tested by two correlations. The association of the engagement mean written score with cancellation did not reach significance at 0.05 (r = −0.12, p = 0.489, n = 38 engagements). The association of the respondent mean written score with intent to leave did not reach significance at 0.05 (rho = −0.29, p = 0.062, n = 43 respondents). None of the four outcome comparisons reached significance at 0.05. Both comparisons with cancellation were descriptive, because responses on canceled engagements were written after cancellation.

Pairs were compared on the engagement level form of every measure, each turned so that a higher value was the better outcome, on the 150 engagements where the indicator and all four measures existed. Of the 6 pairs, 5 differed after Holm control.

Clustering reduced the 797 respondent answers to about 222 independent answers. Respondent level intervals therefore used the cluster bootstrap, and respondent level tests on intent to leave used cluster-robust errors.

No measure showed a tested null. No result was inconclusive. The detection floor column is the study floor at 150 engagements. Intent to leave was tested on its effective n.

Each change was computed before rounding, so it can differ by 0.01 from the difference of the two printed values.

Every check held. A check holds when its test lands on the same side of 0.05 as the primary test and, where the primary test is below 0.05, its r keeps the primary sign. Thin-base rule: engagements in any level under 30 engagements of outsourcing model or process family removed. Institution type is not used, because every one of its levels is under 30.

Each of the four measures was computed from data this indicator does not contain, so each result is a finding rather than an arithmetic consequence. This indicator is one of the 100 components, so the engagement indicator mean was recomputed without it, and the correction changed r by 0.00, from 0.83 to 0.83.

Each deviation was computed before rounding, so it can differ by 0.01 from the difference of the two printed values.

The comparison ran on 150 engagements. The indicator followed the study’s typical pattern on every measure.

The peer set comprised 71 indicators of Part 2 with 30 or more scored responses. The indicator was within the peer set. Its share of negative responses ranked 32 of 71 from the highest. Its mean written score (−5.3, n = 45 responses) ranked 27 of 71 from the most negative. The peer median of the mean written score was −4.7 across the 71 indicators.

The responses coded to the indicator were most often also coded to “Rate the customer’s ability to provision the system access and identities that AI agents require” (C108, 9 responses). The second most frequent co-coded indicator was “Rate the availability of the customer’s subject matter experts when this engagement needs them” (C106, 6 responses). The themes of the coding rulebook tied as most frequent in the same responses were “Comparison to a prior engagement,” “Documentation lags the build” and “Liability or regulatory accountability unclear,” each in 2 responses.

Distressed answers made up more than half of the study set (417 of 797 answers, 52.32%). Active, Production had the highest mean by lifecycle stage (mean 4.4, n = 98 respondents), and Canceled had the lowest (mean 2.3, n = 306 respondents). Portfolio Director respondents gave the highest ratings by role (mean 4.5, n = 30 respondents), and Individual Contributor respondents gave the lowest (mean 3.4, n = 615 respondents). The associations held under every robustness check run on this item (22 of 22 checks).

The written responses corroborated the rating at respondent level, with a moderate positive correlation between written score and rating. Written responses were voluntary. They recorded what the respondents who wrote reported, and how common it was is measured by the ratings.

The evidence is reported experience: what the delivery team said about its own engagement, gathered in one cross-section. The figures showed associations, not causes. Indicators about the financial institution were reported at second hand by vendor staff. Cancellation is a descriptive contrast, because canceled engagements were rated after they ended.

Tests across respondent role and vendor type were run at respondent level and treated respondents as independent. The section Clustering and Effective Sample Size states whether, and by how much, clustering within engagements reduced the information in respondent level figures.

Customer Management Executive answers were at anecdote strength (n = 2). Answers clustered within engagements, which reduced the effective number of independent answers (ICC 0.60, 797 answers worth about 222 independent answers).

Shares of written sentiment were shares of responses, not of the study set of respondents. The written score took five fixed values, so its mean and median were coarse. Responses clustered within engagements. Intervals for written sentiment used a bootstrap by engagement. Coding was performed by AI coders under a fixed rulebook. Agreement on whether a response was coded to the indicator rested on 5 responses second-coded under rulebook 0.5, with the coders agreeing on 4 of 5. Agreement on the score rested on 4 responses second-scored under rulebook 0.5, with the scores agreeing in 4 of 4. These bases of second-coded responses are too small for an indicator-level alpha.