My greater team regularly alters AI outputs

The statement is rated from 1 to 10 with no good or bad end. The mean rating was 4.4 and the median 4.0 (n = 665 respondents). The rating was associated with all four study measures. Higher ratings came with more cancellation (r = 0.59, 95% CI 0.47 to 0.68, p < 0.001, n = 150 engagements), lower customer satisfaction among active engagements (r = −0.40, 95% CI −0.56 to −0.22, p < 0.001, n = 94 engagements), a lower engagement indicator mean (r = −0.66, 95% CI −0.74 to −0.55, p < 0.001, n = 150 engagements) and higher intent to leave (r = 0.58, cluster bootstrap 95% CI 0.51 to 0.65, p < 0.001 with standard errors clustered by engagement, n = 620 respondents). After Holm correction, ratings differed by engagement status and lifecycle stage (Welch F 26.25, Holm p < 0.001, omega squared 0.356) and respondent role (Welch F 46.44, Holm p < 0.001, omega squared 0.043).

All 23 written responses were negative. The 19 responses to C216 described the correction of AI outputs as constant work for the team, done before anyone outside it saw the output, seldom logged or raised and seldom reported to the financial institution.

The scale has no good or bad end, so no distressed range applies.

The primary responses numbered 20, written in answer to C216 and C202, the written-answer questions of the indicator's survey section. The remaining 3 responses, 13.04% of the total, were secondary and came from 2 other questions. Respondents who wrote both a primary and a secondary response numbered 0. Negative responses made up 100.00% of primary responses (20 of 20) and 3 of the 3 secondary responses.

Every one of the 19 responses to C216 placed the corrections at a frequency: every day, every week, most days, almost every task or more than once a week, and three gave a share of six, seven or eight cases out of ten. Ten responses across all questions said the corrections were not logged, reported, raised or written up, and two said the team was told to keep the figures steady or to handle the problem without raising it. The team lead or manager knew or was told in six responses, and three said the financial institution was not told or saw only the corrected output.

Four responses traced the corrections to output that did not reproduce from one run to the next. In three responses, the respondent said others on the team altered the outputs while the respondent did not. A response to C156 described a private sheet of every generated test case the team had rewritten by hand, about sixty percent of the generated cases, kept outside any process document.

Of the 23 responses, 23 (100.00%, 95% CI 100.00% to 100.00%) were negative and 21 (91.30%) carried the strongest negative score of −10. Positive responses numbered 0 (0.00%), and no response was neutral. The mean written score was −9.6 (95% CI −10.0 to −8.9, n = 23 responses), and the median was −10.0.

Of the 23 responses, 22 came from respondents who also rated the indicator. Among them, the written score did not track the rating (Spearman rho −0.21, 95% CI −0.46 to −0.11, p = 0.345, n = 22 responses).

Respondents who wrote about the indicator and gave a rating numbered 22, with a mean rating of 7.1. The 643 respondents who rated the indicator without writing about it had a mean of 4.3. The difference, 2.8 (95% CI 1.9 to 3.6), excluded zero.

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. After Holm correction, ratings differed by engagement status and lifecycle stage (Welch F 26.25, Holm p < 0.001, omega squared 0.356). Canceled engagements rated highest (mean 5.8, n = 228 respondents) and active engagements in Build-Out lowest (mean 3.4, n = 150 respondents). Ratings also differed by respondent role (Welch F 46.44, Holm p < 0.001, omega squared 0.043). Individual contributors rated highest (mean 4.6, n = 556 respondents) and portfolio directors lowest (mean 2.5, n = 10 respondents). Outsourcing model, institution type, process family, class of work and vendor type showed no difference.

Written sentiment differed by engagement status and lifecycle stage across the 18 engagements with a written response (Kruskal-Wallis H 5.00, df 1, p = 0.025). Fewer than two groups held 10 or more responses, so the groups are reported as counts only: active engagements in Build-Out (1), active engagements in Transition/Cut-Over (3), active engagements in Production (1) and canceled engagements (18).

No test across respondent roles was run, because fewer than two roles had 30 or more written responses. Individual contributors wrote 22 responses, engagement managers 1, 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 more cancellation (r = 0.59, 95% CI 0.47 to 0.68, Holm p < 0.001, n = 150 engagements). The association held after Holm correction in eleven 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 lower customer satisfaction (r = −0.40, 95% CI −0.56 to −0.22, Holm p < 0.001, n = 94 engagements). The association held after Holm correction in four 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 does not enter the mean.

Higher ratings came with a lower engagement indicator mean (r = −0.66, 95% CI −0.74 to −0.55, Holm p < 0.001, n = 150 engagements). The association held after Holm correction in thirteen 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 higher intent to leave (r = 0.58, 95% CI 0.53 to 0.63, Holm p < 0.001, n = 620 respondents). The association held after Holm correction in thirty-four of the 37 segment levels tested.

Two comparisons set the engagements with at least one negative response against all other engagements. Of the 130 other engagements, 130 had no written response addressing the indicator. These two comparisons therefore do not isolate written sentiment. Of the 20 engagements with a negative response, 15 (75.00%) were canceled, against 41 of the 130 other engagements (31.54%). The difference reached significance (Fisher exact p < 0.001). The comparison of customer satisfaction was not computed, because one of the two groups held too few active engagements.

Two correlations test written sentiment directly. Across the 20 engagements with a written response, a more negative engagement mean written score came with cancellation (point-biserial r −0.58, p = 0.008). Across 20 respondents, the respondent mean written score was not associated with intent to leave (Spearman rho −0.37, p = 0.105). Of the three 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, zero of the six pairs differed.

Answers clustered within engagements (ICC 0.56). Clustering reduced the 665 respondent answers to the equivalent of about 226 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.

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 departed from the study's typical pattern on all four measures. Its association was weaker than the norm on cancellation, customer satisfaction, the engagement indicator mean and intent to leave.

Among the 13 Part 5 indicators with 30 or more written responses, the indicator's share of negative responses, 100.00%, ranked 1st from the highest, against a median of 98.03%. Its mean written score, −9.6, ranked 2nd from the most negative, against a median of −6.3.

The indicator most often present in the same responses was C211, “I have altered AI outputs”, in 22 of the 23 responses (95.65%). C157, “Rate the factual reliability of the AI tools' outputs on this engagement before any human review.” (20 responses), C212, “I discussed the AI output alteration with internal management” (8 responses), C159, “Rate the consistency of the tools' outputs when given the same input and instructions more than once.” (4 responses) and C018, “How happy are you with your current job?” (3 responses) followed.

The most frequent theme was “Human rework absorbs the failure”, in 18 responses (78.26%). The second was “Occurs every cycle or most days”, in 14 responses.

Stronger agreement that the wider team regularly alters AI outputs came with more cancellation, lower customer satisfaction among active engagements, a lower engagement indicator mean and higher intent to leave, and every association held under each of the 22 robustness checks. Respondents on canceled engagements rated the statement well above those on the three active stages (mean 5.8, n = 228 respondents).

The written evidence describes the correction of AI outputs as constant, unrecorded work that kept the delivered output clean while leaving no trace of the work behind it. The associations do not establish that the corrections produced the outcomes. The written responses describe them as one response to the same instability in the AI outputs that the Part 5 ratings record.

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.

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 (23 responses on 20 engagements). Intervals for written sentiment therefore used a bootstrap by engagement.