Introduction

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.

The Full Study Report sets out the complete findings of a study of agentic AI outsourcing at Tier 1 financial institutions. The report covers every indicator the study measured, every study measure and every null result, together with the written evidence the delivery teams provided. The report is the complete account of the study and the source from which every other publication of the study draws its findings.

The study is published at three levels. The most detailed level is a set of 216 companion statistical analyses, one for each indicator, field and derived analysis, each with its full tables, tests and robustness checks. This report is the middle level. It reports the findings of every companion analysis at the depth a reader needs to act on them and directs the reader to the companion analysis for the full statistics. The top level consists of shorter papers, each written for a particular audience, such as the boards of directors of financial institutions.

The evidence is the reported experience of 850 members of vendor delivery teams on 150 agentic AI outsourcing engagements serving Tier 1 financial institutions, collected once, between January and June 2026.

The report is organized in six Parts.

1. The outcomes of the engagements on the four core study measures, and the finding that the structural attributes of an engagement do not separate good outcomes from poor ones (Part 1).

2. The five failure categories of financial institution readiness, co-management, vendor capability, AI technology and the third-party ecosystem (Part 2).

3. The controls meant to govern work performed by agents, and the ways those controls were bypassed or absent (Part 3).

4. The delivery workforce, its experience and its intent to leave (Part 4).

5. Manual alteration by the delivery team, the reasons given for it and its disclosure to the financial institution (Part 5).

6. The executive action plan (Part 6).

The appendices hold the definitions, the indicator inventory, the method, the list of companion analyses and an index.

2 Executive Summary

Agentic AI outsourcing at the Tier 1 financial institutions in this study was performing poorly. By the study date, 56 of the 150 engagements in the study had been canceled (37.33%). The delivery teams rated cancellation at least moderately likely on a further 58 of the 94 active engagements. Combined, 114 of the 150 engagements (76.00%) had been canceled or were at risk of cancellation. The mean of customer satisfaction across the 150 engagements was 43.85%.

Measured returns were scarce. Production had been reached by 30 of the 150 engagements by the time of data collection. A measured return was reported by 17 of the 30, with a mean of 24.71% (n = 17 engagements). Asked to rate the statement that the engagement was not achieving its stated key performance indicators, 84 of 782 respondents (10.74%) disagreed firmly, with a raw rating of 3 or below.

The study’s four key findings concern manual alteration, the controls over agent work, the immaturity of the industry and the design of the financial institution’s processes.

1. The delivery team sustained the engagements by manual alteration, almost never disclosed to the financial institution. Across the five methods, manual alteration was reported on all 150 engagements. The delivery teams reported all five methods, manual alteration of financial-institution-provided data, manual alteration of financial institution processes, manual alteration of AI outputs, reporting alteration and third-party work done by hand. Respondents who had altered data, processes or agent output reported that they or their management had discussed the alteration with the financial institution in 13 of 1,004 answers (1.29%). Among the delivery workforce that holds the knowledge of these corrections, likelihood of leaving within 12 months was rated 7 or above by 430 of 798 respondents (53.88%).

2. The controls meant to govern agent work did not operate as designed. Ratings of joint human review, change control and quality assurance placed close to half of the answers in the distressed range. The capacity of joint human review to keep up with the volume of agent output was distressed for 381 of 805 respondents (47.33%). The pace of joint change control against changes in agent behavior was distressed for 399 of 798 respondents (50.00%). The reliability of the joint quality assurance loop in catching agent errors before they reached live operations was distressed for 377 of 787 respondents (47.90%).

3. Agentic AI outsourcing was in its infancy on the engagements studied. The vendor’s experience of delivering agentic AI outsourcing was two years or less in every answer given (n = 571 respondents). Personal experience with agentic AI did not exceed one year for any of the 793 respondents who answered. The maturity of the vendor’s playbook for taking agentic use cases from concept to governed production was distressed for 392 of 807 respondents (48.57%). Ratings of the agentic AI knowledge and skills held by financial institution staff on the engagement were distressed in 418 of 793 answers (52.71%).

4. The financial institutions had not redesigned their processes for agentic AI, on the delivery teams’ account. Willingness to redesign processes for agentic AI, rather than automate them as they stood, was distressed in 429 of 804 answers (53.36%). Written responses on the subject numbered 43, and 34 were negative. The negative responses described a financial institution that wanted its existing process copied exactly, with the agent taught to reproduce steps designed for people.

None of the five failure categories performed well. On the 10-point scale, the indicators of financial institution readiness had a mean of 3.55 across the 150 engagements, the lowest of the five categories. Co-management followed at 3.65 and vendor capability at 3.67. The two technology categories were rated higher and remained under the midpoint of the scale, with AI technology at 4.60 and the third-party ecosystem at 4.64 (n = 150 engagements). The three lowest-rated categories lay within the financial institution’s control or within its reach through the vendor and the contract.

The study cannot establish that any regulatory or legal obligation was breached. It records what the delivery teams did and whom they told, and it holds no account from any financial institution. Its findings stand as a call for every financial institution to audit its agentic AI outsourcing engagements, beginning with the people who deliver them (Part 6).

3 Study Design, Evidentiary Basis, and Definitions

The study is a single cross-section of the reported experience of vendor delivery teams on agentic AI outsourcing engagements serving Tier 1 financial institutions. Each respondent answered once, between January and June 2026, about one engagement. The design establishes what the delivery teams reported and how their reports were associated with the outcomes of their engagements. It does not establish the order in which events occurred, what any financial institution knew, or how far the findings extend beyond the 150 engagements studied.

Four conventions govern the report. Ratings were given on a 10-point scale, and a rating of 3 or below on the goodness view is distressed. Four core study measures, cancellation rate, customer satisfaction, engagement indicator mean and intent to leave, are applied to every indicator in the same order. Manual alteration by the delivery team is counted by five methods and is measured separately from its disclosure. Every result is reported in one of four states, from an association that held to a null result equivalent to zero.

The full method appears in Appendix D. Every test statistic summarized in the report appears in full in the companion analysis for the indicator concerned.

3.1 Composition of the Study Set

The study has two units of analysis. The engagement is the outsourcing engagement between one vendor and one financial institution, and the study holds 150. The respondent is a member of the vendor’s delivery team on one of the 150 engagements, and the study holds 850. Engagement-level figures are computed with one value per engagement. Respondent-level figures are computed across the respondents and are clustered within their engagements. No figure in the report mixes the two units.

The study set spans the stages of the engagement lifecycle, the four outsourcing models, both classes of work, eight process families and 10 types of financial institutions. Each composition is reported at its own base in sections 3.1.1 to 3.1.3. Individual contributors were the largest role, and 120 of the 150 engagements (80.00%) had not reached Production by the study date.

3.1.1 Respondents, Roles, and Engagements

The study covers 150 engagements and 850 respondents. Individual contributors, the role that performs the engagement’s tasks, numbered 667 of the 850 respondents (78.47%). Engagement managers numbered 150 of the 850 respondents (17.65%), exactly one on each engagement. Portfolio directors, who oversee several engagements, numbered 31 of the 850 respondents (3.65%). Two respondents were customer management executives.

The roles differ in size from 667 individual contributors to two customer management executives, and the report states every small base. A figure computed on fewer than 30 units is a thin base and prints its count with n. A figure computed on fewer than 10 units is of anecdote strength and is reported as a count, never as a rate. The two customer management executives fall under that threshold in every cut and are reported as counts only.

3.1.2 Structured Fields and Written Answers

The survey collected two forms of evidence. The structured answers number 149,619, given to 172 numeric questions and 23 categorical questions, 195 in total. The numeric questions comprise the ratings on the 10-point scale and typed numbers such as months, team size or contract term. The 100 rated engagement indicators, 20 in each of the five failure categories, form the core of the numeric questions. The categorical questions record attributes such as the type of financial institution, the outsourcing model and the respondent’s role, and the yes-or-no questions on manual alteration and its disclosure.

The written answers number 8,918, given to the study’s 16 open questions. The open questions asked respondents to name the challenges that most reduced their job satisfaction and what they valued most in the job. Others asked for the engagement’s use case, the laws and regulations that governed it, and how the vendor altered the performance reports the financial institution received. The remainder asked why the respondent had altered the financial institution’s data, processes or AI outputs, and invited comment on each part of the engagement. Each substantive written response was assigned a sentiment score on a scale from −10 to +10.

Written answers were voluntary. A written response records what the respondent who wrote it reported and describes the respondents who chose to write. The frequency of a practice or a failure in the study set is measured by the structured answers, never by a count of written responses.

3.1.3 Distribution of Engagements by Lifecycle State, Model, and Institution Type

The engagements were spread across three lifecycle stages. Build-Out held 60 of the 150 engagements, Transition/Cut-Over 60 and Production 30. By the study date, 56 engagements had been canceled and 94 remained active. Cancellation occurred in every stage, and its rate did not differ across the stages (chi-square 5.39, p = 0.068, n = 150 engagements).

The four outsourcing models were represented in similar numbers. Customer experience management (CXM) accounted for 42 engagements, information technology outsourcing (ITO) for 40, business process outsourcing (BPO) for 37 and knowledge process outsourcing (KPO) for 31. The two classes of work were nearly even, with 76 engagements in a line-of-business operation and 74 in a corporate function. Each engagement’s use case was assigned to one of eight process families, shown in Table 1.

The engagements served 10 types of financial institutions, shown in Table 2. Insurers were served by 14 engagements, and private equity and private capital firms by nine. The 23 engagements serving those two types fall outside the banking types. The report therefore uses the term financial institution throughout.

Delivery from India accounted for 92 of the 150 engagements (61.33%). Five types of vendor delivered the engagements, from IT services majors to pure-play agentic AI firms. Both compositions are reported with their bases in the analysis of structural attributes (Part 1).

3.2 Evidentiary Basis and Its Limits

The evidence of the entire report is reported experience, the account that members of the vendor delivery teams gave of their own engagements. The study did not observe the engagements, inspect their records or interview the financial institutions they served. Every rating, count and written response is the respondent’s account.

Three limits follow and bind every Part. First, the report makes no claim about what a financial institution knew, received or would have detected. Where the report states that the delivery team did not inform the financial institution of a manual alteration, it records what the respondent did and to whom the respondent reported it. Second, the study observed each engagement once. It can show that two measures were associated but not which came first. Third, the engagements were selected to cover a range of engagement types, and the findings describe the 150 engagements studied rather than the industry as a whole.

3.2.1 Vendor Delivery Staff as the Sole Source of Evidence

Every respondent was employed by the vendor delivering the engagement. No financial institution was approached or provided commentary for this study, and no model provider or third-party provider was surveyed. The study therefore sees each engagement from one side, through the people who performed its tasks and managed its delivery.

The engagements were selected with partial intent to cover a range of financial institution types, outsourcing models and use cases. The number of engagements of any type in the study set does not reflect how common that type is in the industry. The study recorded no information that could identify a financial institution, the people it serves or the vendor.

3.2.2 Financial Institution Conditions Reported at Second Hand

Financial institution readiness is described by 20 of the 100 rated indicators, covering its data, its processes, its staff, its governance and its speed of response. Every rating of the financial institution is the delivery team’s account of the circumstances it worked within. A low rating of the financial institution’s data records that the delivery team found the data deficient for the tasks it was asked to perform.

The same rule governs disclosure. A respondent who reported that the financial institution was not told of a manual alteration reported what the respondent did and whom the respondent told. The statement does not establish what the financial institution knew. Any other route by which the alteration came to the financial institution’s attention lies outside the study. The report states disclosure in those terms throughout.

3.2.3 Observational Reach of Each Respondent Role

Each role sees a different part of the engagement. Individual contributors see the data, the agents and the daily tasks at first hand. Engagement managers see the engagement as a whole and the relationship with the financial institution. Portfolio directors see several engagements and the commercial terms behind them.

The answers of “I do not know” show where each role’s view ends. The commercial and institutional facts of an engagement include the asset size of the financial institution and the annual value of the contract. On such facts, the share of respondents who answered “I do not know” was higher in each less senior role (Part 1). An answer of “I do not know” on such a question is treated as a finding about who held the information.

Role is also closely associated with experience. 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 therefore cannot be separated from a difference in experience.

3.2.4 Standing of a Finding: Association Without Causal Direction

The report distinguishes two forms of estimate. The clean estimate is computed among the 94 active engagements, against the customer satisfaction recorded for each. No cancellation had occurred on those engagements for a respondent to react to, and the clean estimate is the report’s primary form. The descriptive contrast is computed across all 150 engagements, against realized cancellation. The 56 canceled engagements were rated after they had been canceled, and their respondents’ ratings may reflect the cancellation itself. Descriptive contrasts can therefore exceed clean estimates and are never presented as forecasts of cancellation.

Findings are stated as associations. The report says that an indicator was associated with a measure, that the association held or did not hold, or that it was equivalent to zero. It does not attribute an outcome to an indicator. A single cross-section cannot establish the order of events.

3.3 The Ten-Point Rating Scale

Respondents rated each engagement indicator, and the other rated questions, on a scale from 1 to 10. The questions were not all worded the same way. Some asked how well a part of the engagement worked, where a higher rating was better, and others asked how badly something stood, where a higher rating was worse. A small number had neither a favorable nor an unfavorable end. Three conventions, the goodness view, the distressed threshold and the base convention, put every rating on a common footing. Ratings of differently worded questions can then be compared, and their associations interpreted on the same basis.

3.3.1 Scale Direction and the Goodness View

On the 10-point scale, 1 is the worst outcome a respondent can report and 10 the best. Among the 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. The eight reverse-worded questions include the respondent’s intent to leave, the likelihood that the financial institution would cancel the engagement and the statement that the engagement is not achieving its stated key performance indicators.

The report presents reverse-worded questions on the goodness view, in which each rating is subtracted from 11. On the goodness view, a higher figure is always the better outcome. The statement that the engagement is not achieving its stated key performance indicators, for example, received a mean rating of 6.2 (n = 782 respondents), equal to 4.8 on the goodness view. Intent to leave is the exception and is reported on its original scale, where a higher rating means a respondent more likely to leave.

Eight rated questions have neither a favorable nor an unfavorable end. They ask, for example, whether the financial institution is scaling its agentic AI outsourcing or shifting work to its own global capability centers. For these questions the report states whether each association was positive or negative, without treating either end of the scale as the bad end.

3.3.2 The Distressed Threshold

A distressed rating is a rating at the bad end of the question’s own scale. It is a rating of 3 or below where a low rating is bad, and of 8 or above where a high rating is bad. On the goodness view, every distressed rating is 3 or below. The threshold marks the ratings a respondent gave when that part of the engagement was, in the respondent’s judgment, failing.

The threshold also applies to engagements. An engagement is distressed on a question when the mean of its respondents’ ratings falls in the distressed range. On the statement that the engagement is not achieving its stated key performance indicators, 241 of 782 answers were distressed (30.82%), on 32 of the 150 engagements.

Indicators distressed is the share of an engagement’s indicator ratings that are distressed. It is computed for each respondent across the 100 rated indicators, averaged within the engagement and then across the 150 engagements. Across the study set, 42.49% of indicator ratings were distressed (n = 150 engagements).

3.3.3 Response Counts and the Base Convention

Every percentage in the report prints its base, the count of respondents, answers or engagements on which it was computed. The base is the number who answered the question, not the number who were asked. Percentages are printed to two decimals. Where a figure rests on fewer than 30 units, n is printed beside it.

The yes-or-no questions on manual alteration also allowed the answers “I do not know” and “I prefer not to answer.” Their rates are computed on the ruled base of yes plus no. To the question of whether the respondent had altered financial-institution-provided data, for example, 421 of 845 respondents answered yes and 418 answered no, and the rate is 421 of 839 (50.18%).

On questions of perceived fact, such as the asset size of the financial institution or the contract term, an answer of “I do not know” is reported as a finding. It records that the respondent did not hold the information.

3.4 Cancellation and Cancellation Risk

The study records two views of cancellation. Realized cancellation records, for every engagement, whether it had been canceled by the study date. It is drawn from the Engagements sheet of the study data and does not depend on any respondent’s judgment. On that sheet, 56 of the 150 engagements had been canceled.

Forward risk is the delivery team’s own estimate, for the engagements still active, of how likely the financial institution was to cancel. It is a rating, given by the respondents, and is reported as their judgment. The two views combine in a single population, the engagements already canceled together with the active engagements their delivery teams rated at risk.

3.4.1 Cancellation Likelihood on Active Engagements

Cancellation likelihood is the answer to the question of how likely the financial institution was to cancel the engagement, worded for engagements not yet canceled. It was rated from 1 to 10, where 10 is most likely, and is reported on the goodness view, on which a higher figure marks an engagement judged less likely to be canceled. The mean rating on the goodness view was 5.0 (n = 328 respondents), and 269 respondents answered “I do not know.”

The rating records the delivery team’s estimate. It is not a prediction of cancellation, and the report does not test it as one. Its value lies in what the delivery team believed about engagements that were still running when the study was conducted.

3.4.2 High Risk, Moderate Risk, and the Canceled-or-at-Risk Population

Each of the 94 active engagements was placed in a risk band by the mean of its respondents’ ratings of cancellation likelihood, on the original scale. An engagement whose mean was 7 or above is high risk, and one whose mean was 5 or above and below 7 is moderate risk. By that rule, 19 active engagements were high risk, 39 moderate risk and 36 below moderate risk.

The canceled-or-at-risk population joins the 56 canceled engagements to the 19 high-risk and 39 moderate-risk engagements. The population numbers 114 of the 150 engagements (76.00%). It is the most consequential single figure in the report. Three engagements in four had been canceled, or were rated by their own delivery teams at least moderately likely to be canceled.

3.5 The Four Core Study Measures and Supplementary Measures

Four core study measures are applied to every indicator in the report, always in the same order. They are cancellation rate, customer satisfaction, engagement indicator mean and intent to leave. Together they describe whether an engagement survived, the satisfaction recorded against it, the circumstances it was delivered under and whether its delivery team intended to stay.

The four measures are closely interdependent. Every pair was correlated at 0.72 or more in absolute value across the 150 engagements, and one component accounted for 88.23% of their variance. A result on one measure is therefore interpreted alongside the other three. Two supplementary measures, measured return and indicators distressed, are reported where their bases allow.

3.5.1 Engagement Indicator Mean

The engagement indicator mean summarizes the circumstances an engagement was delivered under. It is computed in three steps. Each respondent’s ratings on the 100 rated indicators are averaged, the respondent means within each engagement are then averaged, and the figure for the study set is the mean across the 150 engagements. Across the study set it was 4.02 on the 10-point scale (n = 150 engagements).

Each indicator is also one of the 100 components of the mean. When an indicator is tested against the engagement indicator mean, the mean is recomputed without that indicator. The leave-one-out rule prevents an indicator from being correlated with a measure that contains it.

3.5.2 Cancellation Rate

The cancellation rate is the share of a named population of engagements that had been canceled by the study date. Across the study set, 56 of the 150 engagements had been canceled (37.33%). The rate is computed at engagement level only. Cancellation belongs to the engagement, and a rate computed across respondents would weight each engagement by the size of its delivery team.

3.5.3 Customer Satisfaction

Customer satisfaction is the satisfaction figure recorded against each engagement on a 100-point scale. The figure was reported by the vendor’s delivery team for the engagement, and the study holds no record from the financial institution against which to confirm it. Across the 150 engagements its mean was 43.85%. Its mean among the active engagements was 64.10% (n = 94 engagements).

The clean estimate of any association with customer satisfaction uses the 94 active engagements. The figures for canceled engagements were recorded after cancellation. The report uses the measure’s full name in every instance.

3.5.4 Intent to Leave

Intent to leave is each respondent’s rating of how likely he or she is to leave the current job in the next 12 months, from 1 to 10. It is reported on its original scale, where a higher rating marks a respondent more likely to leave. The mean rating was 6.77 (n = 798 respondents), and 430 of the 798 respondents (53.88%) rated their likelihood of leaving at 7 or above.

Intent to leave is measured at respondent level. Respondents on the same engagement share its circumstances, and tests against respondent-level questions therefore use standard errors clustered by engagement. Where an engagement-level form is needed, the measure is the mean of its respondents’ ratings.

3.5.5 Measured Return and Indicators Distressed

Measured return is the return on the engagement, recorded only for the 30 engagements that had reached Production. By the study date, 10 of the 30 had been canceled and had not generated a return. A return was reported by 17 engagements, with a mean of 24.71%, and the remaining three active engagements in Production had not yet generated one. Because the measure exists only for engagements in Production, it is never printed in a cut of the whole study set.

Indicators distressed, defined in section 3.3.2, is a study measure but not one of the four core measures. It is reported where the share of distressed ratings is the point at issue.

3.6 Manual Alteration as Defined and Counted

Manual alteration is the study’s term for the delivery team changing, by hand, the material of the financial institution, agent output or the reporting of the engagement. A respondent who corrected records in a data file before an agent processed them had made a manual alteration. The same held for a respondent who decided a process step that the financial institution had not decided, or who adjusted agent output before it was sent.

Manual alteration is a different object from the five failure categories. The failure categories are groups of rated indicators describing the conditions of the engagement. Manual alteration is what the delivery team did within those conditions. The report counts it by method, measures its disclosure separately and makes no assumption that any alteration was wrong in itself.

3.6.1 The Five Manual Alteration Methods

The study counts manual alteration by five methods, shown in Table 3. Three methods were counted by a direct yes-or-no question, each followed by a written question asking why. Respondents reported altering financial-institution-provided data in 421 of 839 answers (50.18%), the financial institution’s processes in 490 of 841 (58.26%) and AI outputs in 278 of 828 (33.57%).

The fourth method, reporting alteration, concerns the performance reports and dashboards the financial institution received. Every respondent was asked how the vendor altered those reports. The method is counted from the written answers other than “I do not know.” Such an answer was given by 484 of the 775 respondents who answered (62.45%), on 127 engagements.

Third-party work done by hand, the last of the five methods, is the delivery team performing by hand a task that a third-party system did not perform. Respondents described pulling records from a third-party portal that offered no interface to an agent and keying in data that a broken feed did not deliver. Others described reopening records that a change by the third party had stopped. The method was rated on two questions and described in 91 written answers from 87 respondents on 63 engagements.

3.6.2 Counting Conventions and Coverage

Two coverage figures are reported, and they measure different populations. On the three methods counted by a direct question, at least one manual alteration was reported by 619 respondents on 146 of the 150 engagements. Across all five methods, manual alteration was found on all 150 engagements. The second figure holds after excluding every written answer that expressed no knowledge, denied that alteration occurred, contained no content or criticized the survey.

The two figures are never presented as measures of the same population. The first counts respondents who answered yes to a direct question. The second adds the reporting alteration and third-party work described in the written answers.

3.6.3 Disclosure as a Separate Object

Disclosure is what followed a manual alteration, and the report measures it apart from the alteration itself. Each respondent who reported altering data, processes or AI outputs was asked two further questions. The first asked whether the respondent had discussed the alteration with the vendor’s own management. The second asked whether the respondent or the respondent’s management had discussed it with the financial institution.

Internal disclosure, to the vendor’s management, was reported in 783 of 1,113 answers (70.35%). Financial institution disclosure was reported in 13 of 1,004 answers (1.29%). Both counts sum the answers to three questions, one for each method, and are counts of answers rather than of respondents.

3.7 Statistical Methods and Reporting Conventions

Every indicator, field and derived analysis was tested in its own companion analysis against the four core study measures. The tests follow one protocol throughout, and this report states their results in a fixed form, giving the statistic, its interval where one applies, its p value and its base. The protocol fixes the tests, the four states in which a result is reported and the treatment of null results and written evidence.

Two thresholds govern the interpretation of every result. At the study’s 150 engagements, the smallest correlation the tests could reliably detect was 0.23, at a 5% significance level and 80% power. An association smaller than 0.30 in absolute value, the study’s bound for a negligible association, can be shown to be equivalent to zero.

3.7.1 Primary Estimates, Descriptive Contrasts, and Partial Tests

Each indicator was correlated with each of the four core measures, and the result is reported as Pearson’s r with its 95% confidence interval. The test against cancellation uses all 150 engagements and is a descriptive contrast. The test against customer satisfaction uses the 94 active engagements and is the clean estimate. The test against the engagement indicator mean uses all 150 engagements, with the indicator removed from the mean. The test against intent to leave uses the respondents, with standard errors and a bootstrap confidence interval clustered by engagement.

A partial test asks whether an indicator was associated with a measure beyond what the engagement indicator mean already accounts for. Each of the 100 rated indicators was tested in this way against cancellation, customer satisfaction and intent to leave, 300 tests in total, with the false discovery rate controlled across the full set. Comparisons across segments of the study set were corrected for the number of comparisons by the Holm method.

3.7.2 The Four States of a Reported Result

Every result in the report is stated in one of four states. An association held when its test reached significance at 0.05, survived correction for multiple comparisons and kept its sign and significance under the robustness checks. Each of the 100 rated indicators was put through 22 robustness checks. Each check repeats the test on a subset of the study set or under an alternative method.

An association did not hold when it reached significance on the primary test but failed correction or one or more robustness checks. A null equivalent to zero is a result that did not reach significance and whose interval lay within the equivalence bound of 0.30. Any association large enough to matter can then be ruled out. An inconclusive null did not reach significance but could not be shown equivalent to zero, usually because its base was small. An inconclusive null is not evidence of no association.

3.7.3 Tests for Unordered Attributes

Attributes without an order, such as the outsourcing model or the type of financial institution, were tested by comparing the levels of the attribute. Cancellation, a yes-or-no outcome, was compared across levels with the chi-square test, or with Fisher’s exact test or a Monte Carlo test where an expected count was under five. The other measures were compared with Welch’s analysis of variance. That test does not assume equal spread across levels.

Levels with fewer than 10 engagements were excluded from these tests and are reported as counts. Where an attribute was recorded once for each engagement, it was tested on engagements. Respondent role and vendor type were tested on respondents.

3.7.4 Engagement Indicator Quintiles

The indicator quintiles divide the 150 engagements into five bands of 30, ranked by engagement indicator mean, with ties broken by engagement identifier. The lowest band holds the 30 engagements delivered under the poorest conditions and the highest band the 30 delivered under the best. The quintiles allow every study measure to be compared across the full range of conditions in five steps. Because the bands are built on the engagement indicator mean, the rise of that mean from band to band is true by construction and is not reported as a finding.

3.7.5 Reporting of Null Results

A null result is reported with the same prominence as a difference. Many of the explanations a reader might expect for poor outcomes, such as the type of financial institution, the outsourcing model or the age of the engagement, returned null results equivalent to zero. Those results rule the explanations out, and the report states them as findings.

Some null results follow from the answers themselves. Where nearly every respondent gave the same answer, no association could appear. Every respondent who answered, for example, placed the vendor’s experience of agentic AI outsourcing at zero, one or two years (n = 571 respondents). A null forced by so narrow a range is identified as such and is not presented as evidence that the attribute does not matter.

3.7.6 The Written Answers as Evidence

Each substantive written response is assigned a sentiment score from −10 to +10, in five fixed steps. A score of −10 marks a part of the engagement described at or near its worst, and −5 a clear but partial problem. A score of 0 marks no position, or good and bad in balance. A score of +5 marks a clear but partial strength, and +10 a part of the engagement described at or near its best. Each response is also identified by the indicators it addresses and by the recurring themes it contains.

A response is primary when it answers the indicator’s own open question and secondary when it addresses the indicator in answer to another question. A respondent asked for comment on the financial institution may describe its data, its processes and its speed of response in one answer, and that answer is evidence for each of the indicators it describes. The report quotes written responses as the respondents wrote them, spelling and grammar unchanged, with words omitted only where an ellipsis marks the omission. Each quotation is attributed by role, outsourcing model and process family, and each is used once.

3.8 Reading Conventions

The report is a reference work. Few readers will study it from beginning to end, and it is written so that a reader can enter at the Part, section or indicator that concerns them. Each section reports its own evidence in full, with the bases and tests it relies on, and states its findings without depending on the section before it. Three conventions govern where each figure comes from and how a single section relates to the rest of the study.

3.8.1 Provenance of Every Figure

Every figure in the report is computed from the study data, comprising the structured answers, the written answers and the Engagements sheet. Each figure appears in the companion analysis for the indicator concerned, where its computation can be followed in full. Terms are used as the Study Taxonomy defines them, one term for each concept. The sources of the study are listed once, in a consolidated list at the end of the report, and are not cited in the text.

3.8.2 Relationship to the Companion Indicator Analyses

Each section that reports an indicator summarizes one of the 216 companion analyses. The companion analysis contains what the summary omits, including the full distribution of ratings, the tables by segment, every robustness check, the equivalence tests and the complete statistics of the written evidence. The report gives the result each companion analysis established, its base and its standing, and selects the written responses that best show what the ratings measured. A reader who needs the statistics behind a sentence will find them in the companion analysis for the indicator named in the section heading.

3.8.3 Use of a Single Block in Isolation

A reader who turns to a single section will find it placed by its parent sections. A section on one indicator sits within a section on its sub-topic. That section in turn sits within a section on its theme, and each higher section states the finding its subsections support. Terms such as distressed, the goodness view, the clean estimate and the four core study measures are defined once, in this Introduction, and are not defined again. A reader entering at a single section can rely on those definitions and on the parent sections to place what the section reports.