Rate the accuracy of citations, references, and quotations the tools produced against the real sources

The mean rating was 4.5 and the median 5.0 (n = 791 respondents). Ratings of 3 or below, the distressed range, made up 33.50% of the answers. The rating was associated with all four study measures. Higher ratings came with less cancellation (r = −0.65, 95% CI −0.73 to −0.54, 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 = 0.86, 95% CI 0.81 to 0.89, 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 = 746 respondents). After Holm correction, ratings differed by engagement status and lifecycle stage (Welch F 43.71, Holm p < 0.001, omega squared 0.403), respondent role (Welch F 8.09, Holm p = 0.003, omega squared 0.014) and vendor type (Welch F 5.32, Holm p < 0.001, omega squared 0.031).

Of the 36 written responses, nearly all from C177, the comment field on the AI tools, 31 were negative. Respondents described citations to clauses, sections and values that did not exist, and citations to real documents that did not say what the output claimed. The response was to check every reference by hand. Written sentiment was not associated with the study measures.

Ratings of 3 or below, the distressed range, accounted for 265 of the 791 answers (33.50%). The engagement mean fell in the distressed range in 32 of 150 engagements.

The primary responses numbered 35, 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 1 responses, 2.78% of the total, were secondary and came from 1 other questions. That secondary share ranked 56th 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 85.71% of primary responses (30 of 35) and 1 of the 1 secondary responses.

Fourteen responses described citations, quotations or values that did not exist in the source: a leave policy clause absent from the whole handbook, a vendor policy number, a section of the financial institution's own API guide for an endpoint renamed two versions earlier, a clause not in the contract, a reference number on a sanctions notice, transactions quoted but not in the file. On a collections account the tool quoted guidance clauses to customers in hardship, and the respondent estimated that about one time in five the clause did not say what the tool said, an error the respondent called a compliance breach with a recording attached.

Nine responses described citations to a real source that did not support the output, or to the wrong version, period or entity: a footnote from the wrong composite period, a leave section replaced the previous year, a hardship clause from a retired policy, a news piece about a different company with a similar name. In one meeting a forecast pack built on the tool's run report cited a real document that did not contain the figure, and the pack was withdrawn in the room. Thirteen responses described checking every citation or quoted figure against the source by hand, in one case finding that roughly a third of the references did not survive the check.

The five positive responses said quoting and links had become accurate or were correct when checked, and placed the weakness elsewhere: retries, forgotten instructions, or runs reported complete that were not.

Of the 36 responses, 31 (86.11%, 95% CI 75.00% to 97.06%) were negative and 1 (2.78%) carried the strongest negative score of −10. Positive responses numbered 5 (13.89%), and no response was neutral. The mean written score was −3.8 (95% CI −4.9 to −2.5, n = 36 responses), and the median was −5.0.

Of the 36 responses, 36 came from respondents who also rated the indicator. Among them, the written score rose with the rating (Spearman rho 0.43, 95% CI 0.16 to 0.66, p = 0.009, n = 36 responses). The written responses corroborated the rating. Of the 31 negative responses with a rating, 22 (70.97%) came from respondents whose rating was in the distressed range.

Respondents who wrote about the indicator and gave a rating numbered 36, with a mean rating of 3.4. The 755 respondents who rated the indicator without writing about it had a mean of 4.6. The difference, −1.2 (95% CI −2.0 to −0.4), excluded zero. Ratings in the distressed range made up 61.11% of the ratings from respondents who wrote (22 of 36) and 32.19% of the ratings from respondents who did not (243 of 755).

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 43.71, Holm p < 0.001, omega squared 0.403). Active engagements in Build-Out rated highest (mean 5.4, n = 163 respondents) and canceled engagements lowest (mean 3.2, n = 301 respondents). Ratings also differed by respondent role (Welch F 8.09, Holm p = 0.003, omega squared 0.014). Portfolio directors rated highest (mean 5.4, n = 29 respondents) and individual contributors lowest (mean 4.4, n = 611 respondents). Ratings also differed by vendor type (Welch F 5.32, Holm p < 0.001, omega squared 0.031). Consulting firms rated highest (mean 5.3, n = 105 respondents) and BPO specialists lowest (mean 4.0, n = 189 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 30 engagements with a written response (Kruskal-Wallis H 2.63, df 3, p = 0.452). Of the groups with 10 or more responses, responses from canceled engagements were the most negative (mean −5.0, 12 responses) and responses from active engagements in Transition/Cut-Over the least negative (mean −2.9, 14 responses). Responses from active engagements in Build-Out (6) and active engagements in Production (4) 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 26, engagement managers wrote 5 and portfolio directors wrote 5 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 = −0.65, 95% CI −0.73 to −0.54, Holm p < 0.001, n = 150 engagements). The association held after Holm correction in fifteen 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 thirteen 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 = 0.86, 95% CI 0.81 to 0.89, 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 = 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 123 other engagements, 120 had no written response addressing the indicator. These two comparisons therefore do not isolate written sentiment. Of the 27 engagements with a negative response, 10 (37.04%) were canceled, against 46 of the 123 other engagements (37.40%). The difference did not reach significance (Fisher exact p = 1.000). Customer satisfaction among active engagements did not differ between the two groups (mean 60.82% on 17 engagements against 64.82% on 77, Welch t −1.13, p = 0.273).

Two correlations test written sentiment directly. Across the 30 engagements with a written response, the engagement mean written score was not associated with cancellation (point-biserial r −0.28, p = 0.141). Across 34 respondents, the respondent mean written score was not associated with intent to leave (Spearman rho −0.29, p = 0.095). Of the four comparisons, none 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 = 0.450).

Answers clustered within engagements (ICC 0.58). Clustering reduced the 791 respondent answers to the equivalent of about 227 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 0.86 to 0.86.

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, 86.11%, ranked 23rd from the highest, against a median of 80.82%. Its mean written score, −3.8, ranked 52nd from the most negative, against a median of −4.7.

The indicator most often present in the same responses was C157, “Rate the factual reliability of the AI tools' outputs on this engagement before any human review.”, in 13 of the 36 responses (36.11%). C172, “Rate how accurately the tools' own reports of success or failure matched the actual outcomes of their work.” (7 responses), C173, “Rate the tools' ability to explain and evidence how each output was produced, at the depth an audit requires.” (6 responses), C164, “Rate how reliably the tools retained instructions and constraints given earlier in long tasks and sessions.” (4 responses) and C171, “Rate the overall production readiness of the AI tools for bank-grade work as experienced on this engagement.” (4 responses) followed.

The most frequent theme was “Occurs every cycle or most days”, in 9 responses (25.00%). The second was “Human rework absorbs the failure”, in 5 responses.

About a third of all ratings of the accuracy of the tools' citations, references and quotations fell in the distressed range (265 of 791 answers, 33.50%). 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 3.2, n = 301 respondents).

The written evidence describes citations that read as authoritative and could not be relied on, so that each one became a task for a person. Respondents reported the checking as the larger cost: a draft the tool produced in minutes took an hour to trust. Written sentiment was not associated with cancellation, customer satisfaction or 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 1 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 (36 responses on 30 engagements). Intervals for written sentiment therefore used a bootstrap by engagement.