Appendices

Reference material for the whole report is collected in the appendices. They define the terms of the study, list every indicator and every table with its base, set out the method in full, and register the 216 companion indicator statistical analyses. Through them, each figure in the report can be traced to the companion analysis that computed it.

Appendix A defines each term the report uses, once, in the form the Study Taxonomy fixes. Every indicator is listed in Appendix B by its code, with the question as the survey asked it, the section of the survey that held it, and the companion analysis that reports it. Each table and figure of the report appears in Appendix C with the indicators and the base behind it.

Appendix D sets out the method. It covers the design of the survey, the composition of the study set, the construction of the four core study measures, and the tests, corrections, and checks applied to every indicator. Its final section states the limits of the evidence in full.

Each companion analysis is described section by section in Appendix E. Its register lists all 216 by the Part of the report that summarizes them. How artificial intelligence was used in the analysis and preparation of the report is disclosed in Appendix G. The appendices close with the consolidated list of sources, an index, and a description of Akholi.

Appendix A defines each term the report uses, in the sense in which the report uses it. The terms follow the Study Taxonomy. It fixes one term for each concept and uses it without variation. Where the survey used a different word, the survey’s word appears only inside a quotation of a question or an answer.

Every indicator the report analyzes is listed in Appendix B, in the order of its code. Each row gives the question as the survey asked it, in the survey’s own words, including its use of “customer” for the financial institution. Its section column shows where the question appeared in the survey. Its class column records whether the indicator measures the respondent, records a fact about the engagement, or records a fact as the respondent perceived it. A paper block column names the section of the report that summarizes the indicator, and a companion analysis column names the file that holds its full statistics.

Each table and figure of the report is listed in Appendix C with the indicators it draws on and its base. Where one base covers the whole exhibit, Appendix C gives it. Where the base differs from row to row, the exhibit prints it beside each figure. From any table or figure, a reader can move to the companion analyses that hold the full statistics behind it. Where a table combines several indicators, each indicator is listed.

A single survey of vendor delivery teams supplied every figure in the report. Respondents numbered 850, on 150 agentic AI outsourcing engagements serving Tier 1 financial institutions, and they answered between January and June 2026. Every indicator was analyzed under one protocol, applied in full in the companion analyses and summarized in the report. Appendix D sets out that protocol, from the design of the survey to the limits of its evidence.

4.1 Survey Design

Each respondent answered about the one engagement on which he or she worked. Five sections made up the survey. Its first section covered the respondent’s role, experience, and intent to stay in the job. A second covered the engagement, including its timeline, scale, commercial structure, performance, and governing regulations. Questions on the financial institution followed, on its profile, its AI strategy, and its relationship with the vendor. A fourth section asked about the vendor, the respondent’s manager, and the respondent’s relationship with the employer.

In the fifth section, the survey examined the challenges of the engagement in depth. Each of the five failure categories, financial institution readiness, vendor capability, co-management, AI technology, and the third-party ecosystem, received 20 rated indicators. An open question closed each category, inviting comment on what its ratings did not capture.

A final group of 18 questions covered manual alteration. Three methods were asked about directly, manual alteration of financial-institution-provided data, of financial institution processes, and of AI outputs. Each method received six questions. A yes-or-no question asked whether the respondent had made the alteration, and a written question asked why. Two further yes-or-no questions asked whether the respondent had discussed it with the vendor’s management and whether the respondent or the respondent’s management had discussed it with the financial institution. A rating of the wider team’s practice and an open question for comment completed the set.

Three types of question appeared in the survey. Rated questions used a scale from 1 to 10. Categorical questions recorded attributes such as role or outsourcing model, together with the yes-or-no answers on manual alteration and its disclosure. Open questions asked for a written answer. In place of a rating, respondents could answer “I do not know” or “I prefer not to answer,” and both answers were counted apart from the ratings.

Rated questions ran in both directions. Of the 146 rated questions with a favorable and an unfavorable end, 138 were worded so that a low rating is bad and eight so that a high rating is bad. Eight further rated questions had neither end. Another 18 numeric questions recorded typed numbers, such as months of experience, team size, or contract term.

Survey answers fill 216 columns of the study data. Five were set aside before analysis, namely the respondent and engagement identifiers, two codes that repeated another answer, and the location of the financial institution’s headquarters. What remains is 211 survey questions, of which 172 are numeric, 23 categorical, and 16 open. Analysis also covers 12 engagement fields, recorded once for each engagement.

4.2 Sampling and Respondents

Every respondent was employed by the vendor delivering the engagement and answered once. No vendor or financial institution was approached for data, validation, or commentary, and no model provider or third-party provider was surveyed. Nothing that could identify a financial institution, the people it serves, or the vendor was recorded.

A specialist provider built the candidate pool from professional networks, published profiles, and vetted panels, and added referrals from early respondents. Participation was compensated, and recruitment reached more than 900 completed responses. Screening removed anyone whose answers showed no working knowledge of banking, the outsourcing industry, or agentic AI. The screen ran before any analysis and could not be tuned toward a preferred result. It retained 850 respondents. In a validation round, 122 respondents completed short telephone interviews to confirm their roles and responses, and no material discrepancy appeared. A third-party research firm administered the survey on its own platform, retained the raw responses, and certified that nothing identifying a financial institution, a vendor, or the people a financial institution serves had been collected.

Selection of the engagements was made with partial intent to cover a range of financial institution types, outsourcing models, and use cases. How many engagements of a type the study set holds does not reflect how common that type is in the industry. Each figure in the report describes the 150 engagements studied, and none is an estimate for the industry as a whole.

Three lifecycle stages are represented, with 60 engagements in Build-Out, 60 in Transition/Cut-Over, and 30 in Production. By the study date, 56 engagements had been canceled and 94 remained active. Customer experience management accounted for 42 engagements, information technology outsourcing for 40, business process outsourcing for 37, and knowledge process outsourcing for 31. Line-of-business operations held 76 engagements and corporate functions 74. Engagement age had a mean of 5.69 months and a median of 5 months (n = 150 engagements).

Individual contributors made up 667 of the 850 respondents (78.47%). Alongside them were 150 engagement managers, one on each engagement, 31 portfolio directors, and two customer management executives.

By vendor type, an IT services major employed 316 respondents and a BPO specialist 198. A consulting firm employed 117, a boutique or specialist provider 87, and a pure-play agentic AI firm 75.

Vendor type was left blank by 57 respondents. For each of them the value was recovered from the engagement, on which vendor type was constant in all 150 cases, and is marked as derived.

Bases vary from question to question, since each is the number of respondents who answered. Intent to leave, for example, was answered by 798 of the 850 respondents. Every mean and percentage in the report and the companion analyses carries its base.

4.3 The Study Data

Two sheets and the written answers make up the study data. One row for each respondent and one column for each survey question form the Data sheet. On the Engagements sheet, each of the 150 engagements has one row of 17 columns. Those columns record the delivery location, the type of financial institution, the outsourcing model, the class of work, the use case, the age of the engagement, and its lifecycle stage. Cancellation, the return on the engagement, customer satisfaction, headcount, and the number of respondents in each role fill the remaining columns.

On any conflict, the Engagements sheet governs. Both sheets were checked against each other before analysis. Where a respondent reported a lifecycle stage and the engagement record carried one, the two agreed in 501 of 501 cases. Cancellation, customer satisfaction, the return on the engagement, and the lifecycle stage are taken from the Engagements sheet.

Fixed rules governed the preparation of the data. One of the 851 rows of the Data sheet held only a column mean and was dropped. Its removal left 850 respondents, equal to the respondent total of the Engagements sheet and to the sum across roles. A value of 0 in any career or tenure question means less than one year. Return on the engagement is blank for the 120 engagements that had not reached Production.

Use cases were recorded as free text, in 121 distinct strings across the 150 engagements. A fixed rule applied to the recorded string assigned each engagement to one of eight process families, taking the first match. The free-text use case never divides the study set.

Each indicator carries a class that governs how its answers are counted. A respondent-level indicator measures the respondent and is computed across those who answered. An engagement attribute is one fact about the engagement, repeated across its respondents and taken from the engagement record. A perceived fact is a fact about the engagement as the respondent understood it. On such a question, an answer of “I do not know” is analyzed as a finding about who held the information.

No change is ever made to the study data file. One program computes every statistic in the report and the companion analyses from it. A second run of the program produces an identical file. Random draws start from a fixed value. In the companion analyses, figures were placed into the text and tables by script rather than typed.

4.4 Construction of the Four Core Study Measures

Cancellation rate, customer satisfaction, engagement indicator mean, and intent to leave are applied to every indicator, in that order. For comparisons across the four, each is turned so that a higher value is the better outcome. Under that orientation, an engagement not canceled, a higher satisfaction, a higher engagement indicator mean, and a lower intent to leave each count as better.

Cancellation rate is computed from the cancellation flag of the Engagements sheet, at engagement level only. By the study date, 56 of the 150 engagements had been canceled (37.33%). A rate across respondents would weight each engagement by the size of its delivery team, and none is computed. Canceled engagements were rated after cancellation. Every association with cancellation is therefore a descriptive contrast.

Customer satisfaction is the figure recorded against each engagement on a 100-point scale. Its mean across the 150 engagements was 43.85%. Among the 94 active engagements, its mean was 64.10%. Figures for canceled engagements were recorded after cancellation. For that reason the clean estimate of every association with customer satisfaction uses the 94 active engagements, and a descriptive contrast across all 150 is reported beside it.

Three steps produce the engagement indicator mean. Each respondent’s ratings on the 100 rated indicators are averaged, the respondent means are averaged within each engagement, and the figure for the study set is the mean of the 150 engagement means. Its study-set value was 4.02 on the 10-point scale (n = 150 engagements), with a standard deviation of 1.48 and a median of 3.81. Engagement means ranged from 1.16 to 7.28.

Each rated indicator is one of the 100 components of that mean. Tested against a mean that contains it, an indicator would be correlated in part with itself. When a rated indicator is tested, the mean is recomputed without it. Its companion analysis prints the correlation with the indicator included and the change. An indicator outside the 100 components is tested against the full mean.

Intent to leave is the answer to the question of how likely the respondent is to leave the current job in the next 12 months, rated from 1 to 10. Measured at respondent level, it is reported on its original scale. Its mean was 6.77 (n = 798 respondents). A rating of 7 or above was given by 430 of the 798 respondents (53.88%). An engagement-level form, the mean of each engagement’s respondents, is reported alongside.

Close interdependence marks the four measures. Every pair was correlated at 0.72 or more in absolute value across the 150 engagements. One component accounted for 88.23% of their variance (n = 150 engagements). Each rated indicator was also tested for an association with cancellation, customer satisfaction, or intent to leave beyond what the engagement indicator mean accounts for. In 33 of those 300 partial tests, p was under 0.05, against 15 expected by chance. None held after the false discovery rate was controlled across the 300 tests.

Two supplementary measures appear where their bases allow. Measured return exists for the 30 engagements in Production. A return was reported by 17 of them, with a mean of 24.71% (n = 17 engagements). Indicators distressed is the share of an engagement’s ratings on the rated indicators that are distressed. Computed for each respondent, averaged within the engagement and then across the 150 engagements, it was 42.49% (n = 150 engagements). Indicator quintiles divide the 150 engagements into five bands of 30 by engagement indicator mean, with ties broken by engagement identifier. A rise in that mean across the bands is true by construction and is never reported as a finding.

4.5 Tests of Association and Difference

Every test is two-sided at the 0.05 level. At 150 engagements, the smallest correlation the tests could detect with 80% power was 0.23. Correlations are reported as Pearson’s r with a 95% confidence interval, a p value, and a base. Spearman’s rho is computed beside each as a check on ranks.

Numeric indicators were tested against each measure in the form that fits the measure. Against cancellation, the indicator’s engagement mean was correlated with the cancellation flag across the 150 engagements as a point-biserial r. An odds ratio per rating point and the area under the curve accompany it. Fisher’s exact test compared the cancellation rate of engagements distressed on the indicator with that of the other engagements.

For customer satisfaction, the clean estimate correlates the indicator’s engagement mean with satisfaction across the 94 active engagements. A slope, in points of satisfaction per rating point, and the share of variance explained accompany it. Against the engagement indicator mean, the correlation uses all 150 engagements and the leave-one-out mean where the indicator is a component.

Intent to leave is correlated with the indicator across the respondents who answered both questions. Its p value uses standard errors that allow for clustering by engagement, and its interval comes from a cluster bootstrap. Respondents distressed on the indicator are compared with the other respondents on the share at 7 or above. Beside the respondent-level result sits the engagement-level correlation, computed on engagement means.

Categorical indicators were tested by comparing their levels. For cancellation, the chi-square test and Cramér’s V compared the levels. Fisher’s exact test replaced it for a two-by-two table, and a Monte Carlo test for a larger one, wherever an expected count fell below five. Welch’s analysis of variance compared the levels on the other measures, reported with omega squared and its 90% confidence interval. Levels with fewer than 10 engagements were left out of these tests and reported as counts.

Ratings were also compared across the segments of the study set. Seven segment families define them, namely the outsourcing model, the type of financial institution, the process family, the class of work, the engagement’s status and lifecycle stage, the respondent’s role, and the vendor type. Welch’s analysis of variance and the Kruskal-Wallis test compared the ratings within each family. For the share of distressed ratings, the chi-square test applied, with Fisher’s exact test or a Monte Carlo test where an expected count fell below five. Families recorded once for each engagement were tested on engagement means, and respondent role and vendor type on respondents.

Within every level of every segment family, each association with each measure was repeated. Status and lifecycle stage form four levels, the active engagements in each of the three stages and the canceled engagements. A level under 30 is reported without an interval, and a level under 10 as counts only.

Across the four measures, an indicator’s associations were compared in pairs, six pairs in all. Each comparison used the engagement-level forms of the measures, oriented as in section 4.4, and the Steiger z test for correlations that share a variable. From that comparison, a reader can see which measure the indicator tracked most closely.

Each indicator was also set against the study norm. For each measure, a Fisher z test compared its oriented correlation with the median correlation across the study’s indicators. An indicator whose deviation held on no measure is reported as following the study’s typical pattern.

Intervals and simulated p values were computed to a stated precision.

Bootstrap intervals were drawn in batches of 10,000 resamples, up to 500,000, until each end of the interval had a standard error of 0.0025 or less. Monte Carlo tests ran in batches of 10,000, up to 2,000,000, until the simulation error on p was 0.0005 or less.

4.6 Correction for Multiple Testing

Each companion analysis runs several tests of the same kind, and some would pass the 0.05 threshold by chance alone. To correct for the number of tests, the study applies the Holm method. Holm orders the p values of a family from smallest to largest. It multiplies the smallest by the number of tests in the family, the next by one fewer, and so on, keeping each adjusted value at least as large as the one before it.

Four families receive the correction in each companion analysis.

1. The tests of difference across the segment families.

2. The association with one measure across the study set and every segment level, run separately for each of the four measures.

3. The six pairwise comparisons of the four associations.

4. The four comparisons with the study norm.

An association is reported as held only when its Holm-adjusted p value was under 0.05. Companion analyses print the unadjusted p value beside the adjusted one. A result below 0.05 before correction and not after is reported as an association that did not hold.

For the 300 partial tests of the rated indicators net of the engagement indicator mean, the study controlled the false discovery rate, the expected share of false positives among the results that pass the test. Its smallest adjusted value was 0.323. The largest partial correlation in absolute value was 0.25.

4.7 Robustness Checks

Associations of a numeric indicator with the four measures were repeated under 22 robustness checks. Each check changes one choice in the primary test and asks whether the conclusion survives.

Seven types of check make up the 22, each applied to the measures where it applies.

1. Rank correlation in place of Pearson’s r, applied to all four measures.

2. Non-answers excluded and then imputed at the engagement mean, applied to intent to leave.

3. Active engagements only, applied to the engagement indicator mean and intent to leave.

4. Removal of the most influential engagement, the one with the largest Cook’s distance, applied to all four measures.

5. Engagements weighted by the number of their respondents, applied to cancellation, customer satisfaction, and the engagement indicator mean.

6. A bootstrap of engagements in place of the analytic interval, applied to all four measures.

7. Removal of the thin-base segments, applied to all four measures.

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. For a check built on an interval, the interval must exclude zero exactly when the primary test does. A check is reported as not computable when fewer than 90% of its resamples were valid.

Thin-base segments are the levels of the outsourcing model or the process family with fewer than 30 engagements. Type of financial institution is not used for that check, since every one of its levels holds fewer than 30 engagements.

Categorical indicators received checks suited to a comparison of levels. In one, the Kruskal-Wallis test replaced Welch’s analysis of variance. Another repeated the test on active engagements only. A third repeated it at respondent level, as a Wald F test with standard errors clustered by engagement.

4.8 Equivalence Tests

A result above the 0.05 threshold can mean that no association exists or that the test could not detect one. Equivalence tests separate the two. Its bound is 0.30 in absolute value of r, a correlation whose square, the share of variance explained, is under 0.09. An association inside the bound is treated as negligible.

Two one-sided tests were run on each null result. One asks whether the correlation lies above −0.30, and the other whether it lies below 0.30. When both have p below 0.05, the null is equivalent to zero, and an association large enough to matter can be ruled out. For categorical indicators, the bound applies to omega squared at 0.09.

Each equivalence table prints the achieved power of the test and the smallest effect it could detect at 80% power, beside the detection floor of 0.23 at 150 engagements. Intent to leave was tested on its effective sample size rather than its count of respondents, since clustered answers carry less information than independent ones.

A null that could not be shown equivalent to zero is reported as inconclusive. An inconclusive null is not evidence of no association. A small base can leave a null inconclusive. Among the structural attributes of the engagements, equivalence to zero on all four core study measures held for 10 of the 19 attributes tested (Part 1).

4.9 Clustering Within Engagements

Respondents on one engagement share its circumstances, and their answers are not independent. Each companion analysis measures clustering with the intraclass correlation, the share of the variance that lies between engagements rather than within them. A design effect follows from it, equal to one plus the intraclass correlation multiplied by the mean number of answers per engagement less one. Dividing the count of respondents who answered by the design effect gives the effective sample size.

Clustering more than halved the information in respondent answers on the two indicators that follow. On whether the engagement was within budget, the intraclass correlation was 0.61 and the design effect 2.63. Its 554 respondent answers carried the information of about 211 independent answers. For the quality of the financial institution’s data across the systems the engagement depends on, the intraclass correlation was 0.57. Those 795 answers carried the information of about 230.

Three provisions allow for clustering. Engagement-level figures use one value per engagement and are free of it by construction. Respondent-level intervals use the cluster bootstrap, resampling whole engagements, wherever the intraclass correlation is 0.05 or above. Respondent-level tests against intent to leave use standard errors clustered by engagement.

Tests of difference by respondent role and by vendor type treat respondents as independent. Each companion analysis reports its effective sample size, from which a reader can judge how far clustering reduced the information behind those tests.

4.10 Treatment of the Written Answers

Written answers number 8,918, given to 16 open questions.

Under written coding rules, 6,908 were coded. Another 615 were stock non-answers, such as “none,” “n/a,” or “same as above,” classified by exact text. Answers on the laws and regulations governing the engagement, 548 in all, were not coded, and the obligations each named were extracted into a list. Use-case answers, 847 in all, were classified into the eight process families.

Before coding began, the coding rules were fixed. They link each answer to 211 measures drawn from the survey questions and the engagement record, of which 149 have a favorable and an unfavorable end and were scored. They add 52 themes in 11 groups for points the survey questions do not capture, seven emotional states, and four types of non-answer. A flag marks an answer that describes another engagement, and such answers are kept out of any link to the respondent’s own engagement.

A mark needs the answer’s own words to state it, and a meaning that could be inferred was not enough. Where the words fit two measures without settling which one the respondent meant, neither was marked.

Coding followed five steps.

1. The question and the whole answer were read.

2. The answer was split into statements, each the shortest run of the respondent’s own words that makes one point. Together the statements rebuild the answer word for word.

3. Each statement was marked with every measure, theme, and emotional state its words state.

4. Each marked measure with a favorable and an unfavorable end was scored.

5. A second pass checked every mark against its rule.

Scores take five values. A score of −10 describes a subject at or near its worst, and a score of −5 a clear but partial problem. Zero marks no position, or good and bad in balance. A score of +5 describes a clear but partial strength, and a score of +10 a subject at or near its best. Either extreme needs words of totality or constancy, serious harm, or the respondent’s strongest language.

Coding produced 28,513 statements and 22,024 marks across the 6,908 answers. Before acceptance, a validator checked each output file. It confirmed that every code existed in the coding rules, that scores used only the five values, and that the statements rebuilt the answer exactly. Each of the 153 output files passed.

A random tenth of the production answers and a calibration set of 100 were coded a second time, by a separate coder who could not see the first coding. Agreement was measured with Krippendorff’s alpha against a standard of 0.80, set in advance. Across the 781 answers coded twice, alpha was 0.919 on topical marks, 0.926 on themes, 0.904 on emotional states, and 0.969 on scores. In the production tenth, topical marks reached 0.916, just below the lowest figure of earlier coding rounds of the study, 0.918.

Two coders differed on at least one mark or score in 288 of the 681 production answers coded twice.

An answer key of 48 hard answers, each carrying a ruling by the author, served as a check on the final coding. Every one of the 48 rulings was reproduced.

On 9 October 2026, the author ruled that every quotation carries a sentiment score. Scores on the same five values were then added for 34 measures with no favorable or unfavorable end that had carried a mark, 3,599 scores in all. An overall score was also given to each of 6,874 substantive answers. A second, independent coding of 777 answers gave an alpha of 0.945 on the overall scores and 0.930 on the 429 added measure scores.

In the companion analyses, a written response is a written answer about the respondent’s own engagement that addresses the indicator. A primary response answered the open question of the indicator’s survey section, and a secondary response answered another question. Written answers were voluntary. Counts of written responses describe the respondents who chose to write, and the frequency of a practice in the study set is measured by the structured answers. Fewer than 10 responses are reported as counts only.

Quotations reproduce the respondents’ words with spelling and grammar unchanged. Words are omitted only where an ellipsis marks the omission, and nothing is inserted. Each quotation is attributed by role, outsourcing model, and process family, and each is used once in the report.

4.11 Limits of the Evidence

Reported experience is the evidence of the study, the account that members of vendor delivery teams gave of their own engagements. Nobody observed the engagements, inspected their records, or interviewed the financial institutions they served. Every rating, count, and written response is the respondent’s account. Ratings of the financial institution describe the circumstances the delivery team worked within.

One source supplied the evidence. No financial institution, model provider, or third-party provider contributed. Customer satisfaction and the return on the engagement are the figures recorded for the engagement, and the study holds no record from the financial institution against which to confirm them.

Each engagement was observed once. A single cross-section can show that two measures were associated but cannot show which came first. No finding in the report is stated as a cause. No change over time can be read from the data.

Selection covered a range of types with partial intent. Within the study set, the share of any type of financial institution, outsourcing model, or use case does not reflect its frequency in the industry. Findings describe the 150 engagements studied.

Canceled engagements, 56 in all, were rated after cancellation, and their respondents’ ratings may reflect the cancellation itself. Associations with cancellation are descriptive contrasts and are never presented as forecasts. Computed on the 94 active engagements, the clean estimate against customer satisfaction is the report’s primary form.

Indicators and measures move together. One component accounted for 88.23% of the variance of the four core study measures (n = 150 engagements). No rated indicator was associated with an outcome beyond what the engagement indicator mean accounts for. For that reason the report does not rank one indicator above another on its association with outcomes.

Role and experience cannot be separated. Individual contributors reported a mean career of 1.9 years (n = 614 respondents), engagement managers 10.0 years (n = 150 respondents), and portfolio directors 15.1 years (n = 31 respondents). A difference between roles is also a difference in experience.

Non-response shapes every base. On whether the engagement was within budget, 554 respondents gave a rating and 242 answered “I do not know.” On a question of perceived fact, such an answer records that the respondent did not hold the information. Every figure describes the respondents who answered the question.

Several bases are thin. Customer management executives, two in all, are reported as counts only, and the 31 portfolio directors form a thin base. Every type of financial institution is served by fewer than 30 engagements. Three process families are served by fewer than 10 engagements each, and their results are reported as counts.

Some null results follow from the answers themselves. Vendor experience of delivering agentic AI outsourcing was placed at two years or less in every answer given (n = 571 respondents). Where nearly every respondent gave the same answer, no association could appear, and such a null is not evidence that the attribute does not matter.

Respondent-level figures carry less information than their counts suggest. Clustering reduced 554 answers to the equivalent of about 211 independent answers, and 795 answers to about 230, on the two indicators of section 4.9.

Written answers were voluntary and describe the respondents who wrote them. Two themes that record how often manual work occurred are marked only where the answer uses wording on a fixed list, and their counts are floors. Topical marks in the production tenth reached an alpha of 0.916, below the lowest figure of earlier coding rounds.

Disclosure is measured from the respondent’s side. A respondent who reported that the financial institution was not told of a manual alteration reported what he or she did and whom he or she told. Nothing in the study shows what any financial institution knew, and any other route by which an alteration came to its attention lies outside the study.

No regulatory or legal breach can be determined from the study. Obligations the respondents named are reported as context for an audit and never as a legal conclusion.

Three lifecycle stages were measured. Preparation and procurement, the stages that precede Build-Out, were not recorded, and no engagement in the study set was observed in them. Measured return rests on the 17 engagements that reported a return, of the 30 in Production.

Companion indicator statistical analyses form the most detailed level of the study. Each reports one indicator, engagement field, study measure, or derived analysis in full, with every table, test, and check behind the result the report summarizes. All 216 were built to one template from the same statistics record.

Analyses are grouped by the Part of the report that summarizes them. A first group of 46 analyses (Part 1) covers the three study measures computed from the engagement record, seven engagement fields, 31 survey questions on the engagement and the financial institution, and five derived analyses. Those derived analyses treat the active engagements, the four measures together, the indicator quintiles, measured return, and the associations net of the engagement indicator mean.

The second group of 101 analyses (Part 2) covers the 100 rated indicators of the five failure categories and one question on the vendor’s preparedness to manage outsourcing engagements. The third group holds 48 analyses (Part 4), 47 on questions about the delivery workforce and one a derived analysis of experience by intent to leave. The fourth group holds 21 analyses (Part 5), 19 on manual alteration, its disclosure, and reporting alteration, and two derived analyses of manual alteration by role.

Every analysis opens with the indicator’s specification. Its rows give the code, the question as asked, the failure category where one applies, the place in the survey, the class, the type, and the response scale. They also give the answering base, the count of each non-answer, and the count of written responses that address the indicator. A summary of findings follows, with a table of the association with the four core study measures.

Descriptive sections come next. Ratings are distributed at respondent level and at engagement level, with the share in the distressed range. Written responses are counted by the question they answered and by their sentiment score. For a rated indicator, two further sections compare each respondent’s rating with the sentiment of his or her written response and profile the ratings of the respondents who wrote. Descriptive statistics follow for every level of every segment family.

Tests follow in a fixed order. Tests of difference compare the indicator across the segment families, and a further section compares written sentiment across segments where the written responses allow. Four sections test the association with cancellation rate, customer satisfaction, the engagement indicator mean, and intent to leave, each for the study set and for every segment level. A section on written sentiment tests its association with the same four measures. In the comparison of associations, the four results are set against one another in pairs.

Checks on those results close the statistical sections. Each analysis reports clustering and effective sample size, the equivalence tests of null results, and the robustness checks. A section on measure independence shows that each measure was computed from data the indicator does not contain. Comparison with the study norm sets the indicator’s associations against the median of the study’s indicators. Further sections compare the indicator’s written sentiment with the other indicators of its Part and list the indicators and themes that appeared in the same written responses.

Interpretation of the data, limitations, definitions, and sources end each analysis. Its sources section names the study data file, the program that computed the statistics, the statistics record, and the coded written answers, each with its file fingerprint. It also gives the date of computation and the version of the template. A derived analysis adds a section of analysis results, as the analysis of associations net of the engagement indicator mean does.

Each indicator block of the report summarizes one companion analysis. In it, a reader finds the mean rating and its base and the result against each core study measure with its state. The block also reports the tests of difference that held and the written evidence, with selected quotations. Everything the block leaves out remains in the companion analysis, including every segment table, every robustness check, the equivalence tests, and the full statistics of the written responses. Appendix B links each indicator to its block and to the file that holds its companion analysis.

All 216 companion analyses are published at akholi.com.

The index lists each term defined in Appendix A with the Part and section in which it appears. A term used in more than 40 sections is listed by Part.

The author prepared this report with the assistance of an artificial intelligence system, Claude from Anthropic. The system worked as a directed tool on defined tasks, and the author reviewed its output. Claude assisted in the statistical analysis of the study data and the coding of the written answers, and helped structure and draft the text. Claude also checked the data and the report to confirm that no financial institution and none of the people a financial institution serves could be identified.

One program computed the statistics from the study data, and the study data file is never modified. The program applies definitions fixed in advance, and a second run produces an identical result. In the companion analyses, figures were placed into the text and tables by script rather than typed. The author ruled on each test, threshold, and convention the program applies.

Claude coded the written answers under written coding rules fixed before coding began. Each coder assigned every mark by reading and was not permitted to use keyword or pattern matching. A tenth of the answers were coded a second time by a separate coder to measure agreement. An answer key of 48 hard answers, each settled by a ruling of the author, served as a check. The final coding reproduced 48 of the 48 rulings.

Drafts of the text were prepared with Claude under written standards set by the author for vocabulary, sentence construction, the treatment of numbers and bases, and quotation. The author reviewed the drafts and ruled on their vocabulary, structure, and content. Every quotation reproduces the respondent’s own words as submitted.

The observations, findings, and conclusions in this report are the author’s own. The artificial intelligence did not generate them, and responsibility for the content of the report rests with the author.

Phil Hatch is a global outsourcing and workforce-development executive with more than 30 years of experience helping organizations and governments build, manage, and transform large-scale talent and service-delivery operations. He served as the CEO of ManpowerGroup/Experis global outsourcing business, where he led the development and growth of a global outsourcing organization serving clients across industries and geographies. His work spans multiple industries and geographies, with a particular focus on connecting business strategy, talent, and economic opportunity.

Beyond his corporate leadership, Hatch has served as a senior labor and economic development advisor to governments, helping public-sector leaders develop workforce strategies to attract investment, expand employment opportunities, and build globally competitive talent ecosystems. His experience sits at the intersection of outsourcing, workforce policy, economic development, and international business, giving him a practical perspective on how governments and companies can work together to create sustainable economic growth.

Following Russia’s invasion of Ukraine in 2022, Hatch was asked by the U.S. Department of State and USAID to serve as a formal advisor to Ukraine on workforce and economic development. In that role, he has focused on how technology, outsourcing, skills development, and international investment support Ukraine’s economic resilience and long-term recovery.

Today, through Akholi, Hatch continues to advise organizations and governments on the changing global outsourcing landscape, including the implications of AI and agentic AI for work, talent, and service delivery.

Akholi is a research-led advisory firm for agentic-era outsourcing, service engagements, and global capability centers, established 13 years ago. Its public authority rests on primary evidence rather than opinion, and on one dataset in particular: the 2026 study of 150 agentic AI outsourcing engagements at Tier 1 financial institutions, told by the 850 people who deliver them.

From that record the firm draws a single governing claim, stated at the strength the evidence supports and no further. In the engagements studied, outcomes track the conditions the buying institution builds rather than the deals it signs. Failure there is institutional in origin rather than technical, and the determinants sit within the institution’s own authority to change. That last clause is the whole posture: the levers are in the institution’s hands. The finding is agency, not fault.

Akholi exists to help boards and C-suites act on it: to diagnose the conditions that decide an engagement, to learn the operating discipline the agentic era demands, and to run the hardest engagements with a firm that has counted the failures. The Report is the evidence. The Academy and the Practice are how institutions put it to work.