Where did you gain your AI knowledge?

Of the 850 respondents, 800 answered the question. The answers spread almost evenly across the six sources, from 125 to 140 respondents each: vendor and partner training, on the job, online courses, university coursework, self-taught and the employer's training program. The source did not differ by role, vendor type or any segment family recorded at engagement level.

The source assigned to each engagement was not associated with cancellation (chi-square 2.39, p = 0.792, n = 94 engagements), customer satisfaction among active engagements (Welch F 0.43, p = 0.521, n = 22 engagements), the engagement indicator mean (Welch F 0.24, p = 0.940, n = 94 engagements) or intent to leave (Welch F 0.36, p = 0.871, n = 94 engagements). The engagement indicator mean and intent to leave were equivalent to zero, and every robustness check held.

The written responses addressing the indicator numbered 37, all from individual contributors, on 33 engagements. Every response was secondary. Of the 37 responses, 19 (51.35%) were negative and 14 (37.84%) positive. Whatever source they named in the answer, the respondents who wrote described learning the agents from colleagues, from outside material and on live work, after training of slides and videos.

Each engagement was assigned the source given by the largest number of its respondents. Fifty-six engagements were tied and carry no assigned source, and 94 carry one. Every source was the assigned answer on 10 or more engagements.

All 37 responses were secondary. They came from 2 questions outside the indicator's own survey section (C135 with 28 responses and C030 with 9 responses). Each of the nine responses to C030, on what the respondent loves most about the job, listed agentic skill picked up on the job or firsthand.

The 28 responses to C135, the comment field on the employer, described how the respondents learned the agents. Eleven described the training as slides, recorded videos or a recorded session and a PDF, in several cases followed straight by live work, and one said the course came three months after the respondent had started the work. Eleven said the respondent learned from a lead, a colleague or a senior, and two described a note file the team kept and passed to new joiners in place of the official material. Seven described learning alone, from outside videos at night, search, vendor documents or old test cases and broken runs. Eleven said no one senior, or no one above the respondent, knew the agents or had done the work before, so that questions went unanswered for days or came back to the respondent.

Of the 37 responses, 19 (51.35%, 95% CI 34.29% to 68.29%) were negative and 1 carried the strongest negative score of −10. Positive responses numbered 14 (37.84%), and 4 were neutral. The responses had a mean written score of −0.8 (95% CI −2.4 to 0.8, n = 37 responses on 33 engagements) and a median of −5.0.

Every response to C030 was positive. The responses to C135 were mostly negative (19 of 28), Of their five positive responses, four paired thin training with a lead, a colleague or the team's own notes that made up for it, and the other found the agentic courses practical and placed the gap in banking knowledge.

Each share is computed on the respondents who answered in that row. Rows with fewer than 10 respondents report counts only.

Segment families recorded at engagement level were tested on engagements, and respondent role and vendor type on respondents. After Holm correction, the source differed in no family.

Written sentiment did not differ by engagement status and lifecycle stage (Kruskal-Wallis H 0.66, df 3, p = 0.883, n = 33 engagements). All 37 responses came from individual contributors, so no test across roles was run.

Cancellation rate is the share of engagements canceled by the study date, measured at engagement level (150 engagements, 56 canceled).

Across the 94 engagements with an assigned source, cancellation did not differ by source (chi-square 2.39, df = 5, p = 0.792, Cramer's V 0.16).

Customer satisfaction is the percentage recorded against each engagement. The engagement-level estimate uses the active engagements in the sources with 10 or more active engagements.

Customer satisfaction did not differ by source (Welch F 0.43, p = 0.521, n = 22 engagements), on a base of 22 active engagements in the sources with 10 or more active engagements.

The engagement indicator mean is the mean of an engagement's ratings across the rated indicators, measured at engagement level (150 engagements). The source is not a rated indicator and does not enter the mean.

The engagement indicator mean did not differ by source (Welch F 0.24, p = 0.940, n = 94 engagements).

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 by up to 798 respondents. For the indicator, the engagement-level comparison uses the mean intent to leave of each engagement.

Intent to leave did not differ by source (Welch F 0.36, p = 0.871, n = 94 engagements).

The comparisons set engagements with a negative written response against all other engagements, most of which had no written response addressing the indicator. Cancellation affected 5 of the 17 engagements with a negative response (29.41%), against 51 of the 133 other engagements (38.35%), a difference that did not reach significance at 0.05 (Fisher exact p = 0.598). Customer satisfaction was 63.33% on the 12 active engagements with a negative response, against 64.21% on the 82 other active engagements (Welch t −0.26, p = 0.797), a difference that did not reach significance at 0.05. The engagement mean written score against cancellation (r = −0.03, p = 0.854, n = 33 engagements) did not reach significance at 0.05, and the respondent mean written score against intent to leave (rho = −0.13, p = 0.464, n = 32 respondents) did not reach significance at 0.05. No written outcome test reached significance.

Steiger's z compares two correlations, and a categorical indicator yields an effect size across its levels and not a correlation. The pairs of measures were therefore not compared. No association reached significance.

Answers did not cluster within engagements (ICC 0.00): respondents on the same engagement named different sources. The design effect was 1.00, and the effective n equaled the 800 answers. The respondent-level tests used standard errors clustered by engagement.

The engagement indicator mean (equivalence p = 0.002) and intent to leave (equivalence p = 0.005) were equivalent to zero, each on 94 engagements. Customer satisfaction was not (equivalence p = 0.187, n = 22 engagements), and its result is inconclusive rather than evidence of no association. The cancellation contrast carries no equivalence test.

All nine checks held. A check holds when its test falls on the same side of 0.05 as the primary test. The respondent-level test, with standard errors clustered by engagement, found no difference for any of the four measures (p from 0.647 to 0.809).

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.

A categorical indicator has no signed correlation with the study measures and therefore no profile against the study norm.

Among the 11 indicators of Part 4 with 30 or more written responses, the indicator's share of negative responses, 51.35%, ranked 5 of 11 from the highest, against a median of 33.33%. Its mean written score (−0.8, n = 37 responses) ranked 5 of 11 from the most negative, against a median of 0.8 across the 11 indicators.

The indicator most often present in the same responses was C123, on the adequacy of the employer's investment in training its workforce, in 21 of the 37 responses (56.76%), followed by C124, on how well the employer's training programs build practical agentic AI skills, in 10. The most frequent theme was “Skills learned” (9 responses), every one a response to C030.

Respondents named six sources of their AI knowledge in almost equal numbers, and the source did not separate engagements on cancellation, customer satisfaction, the engagement indicator mean or intent to leave. Two of these results were equivalent to zero, and every robustness check held.

The written responses describe a different picture from the answer itself. Individual contributors who wrote said they learned the agents on live work, from colleagues and from outside material, after training of slides and videos, and that often no one above them knew the agents. The answers record where knowledge was said to come from, and the written responses record how thin the formal route was for those who wrote.

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.

The respondent gave one source. Most people learn from several, so the answer records the source the respondent judged most important.

Customer satisfaction was tested on 22 active engagements, a small base.

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 (37 responses on 33 engagements), and intervals for written sentiment therefore used a bootstrap by engagement. In the sample read by a second reader, the two readers agreed on the presence of the indicator in 5 of the 5 responses either reader marked, and on the score in 2 of the 3 responses both readers scored.