Source: https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/appendices/appendix-d-method/
Author: Phil Hatch
Publisher: Akholi

Appendix D: Method

The 2026 Report · Appendices

# Appendix D: Method

Survey design, sampling, measures, statistical methods, coding, and limitations

By Phil Hatch · Akholi · Version 1.8.2

Complete study · PDFAppendices · PDFStudy overview / Appendices / Appendix D: Method

The study analyzed responses from 850 vendor delivery staff associated with 150 agentic AI outsourcing engagements serving institutions classified in the study as Tier 1. The reported fieldwork period was January–June 2026. A common analytical framework was applied, with tests selected according to variable type, analytical unit, and available sample size.

## 4.1 Survey Design

Each respondent answered about the one engagement on which they 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

Respondents were vendor employees answering about a sampled engagement. The study records state that financial institutions, vendors, model providers, and other third parties did not supply independent validation. The analytical files available to the author contained deidentified records. This does not establish that identifying information was never collected during recruitment or administration.

The reported recruitment process used professional networks, published profiles, vetted panels, and referrals through a specialist provider. Participation was compensated. More than 900 completed responses were reportedly screened for working knowledge of banking, outsourcing, and agentic AI, leaving 850 respondents. Study documentation also describes telephone validation with 122 respondents and reports no material discrepancies. These are descriptions of the administration process; the files supporting the present analysis do not provide a complete recruitment disposition, response rate, screening instrument, or validation record.

Engagement selection sought variation in institutional type, outsourcing model, and use case. The sample is not established as probability-based or representative of the industry. Estimates describe the sampled engagements; confidence intervals and significance tests do not remove this selection limitation.

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.

Among the 793 respondents with an originally recorded vendor type, 316 reported an IT services major, 198 a BPO specialist, 117 a consulting firm, 87 a boutique or specialist provider, and 75 a pure-play agentic AI firm.

Vendor type was originally blank for 57 respondents. The analytical preparation derived these values from the engagement-level classification, which was internally consistent within each engagement. Derived values remain distinguishable from originally reported values; the preceding counts describe the original answers.

The valid base varies by question and analysis. Numeric summaries exclude nonnumeric answers unless a stated sensitivity analysis specifies otherwise. Categorical analyses identify which answers are retained as levels. Counts of written answers, respondents, and engagements are reported separately.

## 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.

The analysis uses the Engagements sheet as the governing source when it conflicts with respondent-level records. The two sources agreed on lifecycle stage in all 501 comparable cases. Cancellation, customer satisfaction, return, and lifecycle stage were taken from the engagement record; this source hierarchy does not independently validate the recorded values.

Data preparation excluded one of 851 rows because it contained only a column mean, leaving 850 respondent records. Zero values in career and tenure fields were interpreted under the study rules as less than one year. Return was blank for the 120 engagements outside Production; among the 30 production records, zero and positive values were retained as recorded.

The 150 engagement records contain 121 distinct use-case strings. The computational engine matches each string exactly to the crosswalk in Study Meta.md, section 9, dated September 21, 2026, which assigns one of eight process families. The supplied metadata states that these assignments agree with the process_family values in the coded-answer records. This is an exact lookup against an established crosswalk, not a new first-match keyword classification during statistical computation.

Analytical classes determine the unit and treatment of a variable. Respondent-level indicators describe individual answers. Engagement attributes use one engagement-level value. Perceived facts describe the respondent’s understanding of engagement conditions. “I do not know” identifies unavailable or uncertain information for that respondent; it does not establish that the organization lacked the information.

The statistical record identifies compute_statistics.py as the computational engine and records input fingerprints. Its v16 layer specifies random seed 20261006, sequential bootstrap settings, and Monte Carlo settings. The older base layer records different settings. These records support traceability but do not establish independent reproducibility in a separate environment.

## 4.4 Construction of the Four Core Study Measures

The four core measures are cancellation, customer satisfaction, engagement indicator mean, and intent to leave. Their original directions are retained in ordinary association reporting: cancellation and intent to leave are adverse outcomes. For formal comparisons of association strength, the measures are reoriented so that higher values consistently indicate more favorable outcomes. Ratings converted to the favorable direction are identified where used.

Cancellation is recorded as one binary value per engagement; 56 of 150 engagements were canceled (37.33%). Engagement-level analysis gives each engagement equal weight unless a sensitivity analysis specifies otherwise. Respondent-weighted comparisons answer a different question and are identified separately. Ratings on canceled engagements were collected after cancellation, so associations with cancellation are retrospective descriptive contrasts.

Customer satisfaction is a recorded engagement score on a 100-point scale. Its mean was 43.85 points across all 150 engagements and 64.10 points across the 94 active engagements. The source field is labeled as a percentage, but these values should not be interpreted as the percentage of customers satisfied. Primary satisfaction associations use active engagements; all-engagement comparisons include post-cancellation scores and are descriptive.

The engagement indicator mean is constructed by averaging each respondent’s available favorable-direction ratings across the 100 rated indicators, then averaging respondent means within each engagement. Study-level summaries give each engagement equal weight. The mean was 4.02 on the 10-point scale (n = 150 engagements), with a standard deviation of 1.48, median of 3.81, and range of 1.16–7.28. Differences in missing-item coverage can affect the composition of respondent means.

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.

The four measures were strongly interrelated: all pairwise engagement-level correlations had absolute values of at least 0.72, and one component accounted for 88.23% of their variance (n = 150 engagements). The analysis also examined 300 partial associations between individual rated indicators and cancellation, satisfaction, or intent to leave, controlling for the engagement indicator mean. Thirty-three unadjusted p values were below 0.05, but none remained significant after false-discovery-rate correction. This does not establish that every independent effect is zero.

Supplementary results include recorded return, the share of distressed indicator ratings, and indicator quintiles. Of 30 production engagements, 17 had positive recorded returns, averaging 24.71%; the other 13 had recorded zero values. The reporting period, financial denominator, and treatment of costs were not established in the supplied analytical records. The mean distressed-rating share was 42.49%, calculated within respondents, then engagements, and then across engagements. Quintiles contain 30 engagements each, ordered by indicator mean with identifier-based tie breaking; differences in that mean across quintiles are true by construction.

## 4.5 Tests of Association and Difference

Association and difference tests generally use a two-sided significance threshold of 0.05; equivalence tests use two one-sided tests as described below. Under the specified assumptions, the approximate detectable correlation at 80% power is |r| = 0.23 for 150 independent engagements. This is a design calculation, not a guarantee of sensitivity in every analysis. Companion results report effect sizes, intervals, p values, and applicable bases; the main text summarizes selected results.

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.

Primary customer-satisfaction analysis correlates the indicator’s engagement mean with satisfaction across 94 active engagements. Companion results also report a slope and squared correlation where applicable. Analysis against the engagement indicator mean uses all eligible engagements and excludes the indicator itself from that mean when it 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 are compared across response levels. Engagement assignments use the unique most frequent response, with ties excluded. Cancellation comparisons use contingency-table tests, with Fisher’s exact test for sparse two-by-two tables or Monte Carlo inference for larger sparse tables. Continuous-outcome comparisons use Welch’s analysis of variance. Omega squared and its interval are calculated separately from the conventional analysis-of-variance decomposition. Levels with fewer than 10 eligible engagements are excluded from inferential comparisons and retained descriptively.

Segment comparisons cover outsourcing model, institutional type, process family, work class, engagement status and stage, respondent role, and vendor type. Numeric comparisons use Welch and Kruskal–Wallis tests; categorical comparisons use contingency-table methods. Engagement-defined families use engagement-level values. Some role and vendor-type comparisons use respondents without clustering adjustments, whereas specified respondent-level outcome tests use cluster-robust methods. The analytical unit and adjustment must be read for each result.

Segment analyses use four status-and-stage groups: active engagements in Build-Out, Transition/Cut-Over, and Production, plus canceled engagements. The protocol suppresses segment intervals below 30 observations and reports groups below 10 descriptively. The supplementary return analysis is a separate analysis class and reports intervals for 17 positive-return engagements; those exploratory intervals require particular caution and should not be treated as meeting the segment-reporting threshold.

Formal pairwise comparisons of numeric-indicator associations use engagement-level versions of all four measures on a common available base, oriented so higher values are favorable. Steiger’s test compares dependent correlations sharing the indicator. These tests concern associations, not differences in causal importance. Categorical effect sizes are not compared with Steiger’s test.

The computational protocol also compares oriented correlations with the study’s median correlation using Fisher-z calculations. Because the indicator and reference correlations arise from shared data and correlated variables, these comparisons should be treated as descriptive benchmarks rather than evidence of a reliable ranking of indicators.

The v16 statistical layer specifies sequential simulation targets rather than a fixed number of draws for every result.

Bootstrap sampling uses batches of 10,000, a maximum of 500,000, and a target standard error of 0.0025 for each interval endpoint. Monte Carlo testing uses batches of 10,000, a maximum of 2,000,000, and a target simulation error of 0.0005 for p. Reaching a draw limit does not itself establish that the precision target was met; analysis-specific computation records govern that determination.

## 4.6 Correction for Multiple Testing

Multiple related tests increase the opportunity for false-positive findings. The study uses Holm adjustment within specified families, rather than treating every unadjusted p value below 0.05 as confirmatory evidence. Holm’s procedure orders the p values and applies sequential multiplicity factors while preserving monotonic adjusted values.

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.

Holm adjustment applies to these defined families. The statistical record also retains unadjusted primary p values, and some main-text results report those values. An unadjusted p value should not be interpreted as adjusted unless explicitly identified. These within-analysis corrections do not control multiplicity across every indicator, written-response comparison, and analysis in the report.

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

The numeric-indicator protocol contains up to 22 checks distributed across the four measures. It does not apply 22 separate checks to every individual association. Checks are omitted or reported as noncomputable where an identity, unavailable base, or other constraint prevents estimation.

For a respondent-level numeric indicator with all checks computable, the protocol assigns five checks to cancellation, five to customer satisfaction, six to the engagement indicator mean, and six to intent to leave. The seven check types and their allocation are as follows.

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 is recorded as holding when its result falls on the same side of the specified significance threshold as the primary test and, for a significant primary association, retains its direction. Interval-based checks assess whether zero is included. This criterion concerns consistency under the specified alternatives; it is not proof that an estimate is unbiased or that an association is causal. A bootstrap check is noncomputable if fewer than 90% of resamples are 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 nonsignificant association test does not establish absence of an effect. The study’s equivalence analysis asks whether an association is sufficiently small to fall within a prespecified region. For correlations, the region is −0.30 < r < 0.30. This is a study convention; effects within that interval may still matter in particular operational contexts.

For correlations, both one-sided tests must reject their respective boundary nulls at 0.05 to support equivalence within the stated interval. Equivalence does not mean a literal zero correlation. Categorical comparisons use an omega-squared bound of 0.09; for cancellation, this is calculated from the binary flag and is distinct from the contingency-table significance test. A small association can be statistically different from zero and still meet the equivalence criterion.

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.

When an association is nonsignificant and does not meet the equivalence criterion, it is inconclusive. Limited sample size can produce this result. Ten of the 19 structural attributes examined met the study’s equivalence criteria for all four core measures; the remaining results require measure-specific interpretation.

## 4.9 Clustering Within Engagements

Respondents within an engagement share circumstances, so their observations may be correlated. The protocol estimates an intraclass correlation and approximates the design effect as 1 + (mean cluster size − 1) × max(ICC, 0). Nominal sample size divided by this design effect gives an approximate effective sample size. The calculation summarizes information loss under its assumptions; it does not replace a cluster-adjusted model or fully capture unequal cluster sizes and higher-level dependence.

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.

Aggregation removes duplication of respondent observations within an engagement but does not establish independence between engagements sharing institutions or vendors. Respondent-level intent-to-leave analyses use cluster-robust standard errors and cluster bootstrap intervals. Other intervals and tests follow their stated analytical units and methods.

Some segment comparisons by role and vendor type treat respondents as independent. Effective-sample-size summaries describe the extent of clustering but do not correct those tests. Separately reported cluster-robust outcome comparisons, including the role analysis, should not be confused with these unadjusted segment tests.

## 4.10 Treatment of the Written Answers

The data contain 8,918 written entries across the open-response and use-case fields. This is distinct from the broader total of 12,088 entries, which also includes 3,170 numbers entered in text boxes.

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.

Rulebook version 0.5 linked answers to 211 measures, including 149 scored measures, 52 additional themes, seven expressed emotional states, and four non-answer types. Other-engagement accounts were flagged and excluded from links to the respondent’s own engagement. Initial rules were fixed before production coding, but later documented revisions expanded frequency-theme wording and added sentiment scores under version 0.6.

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.

AI language-model coders using Claude processed batches of answers under the rulebook. The documented workflow required 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 each assigned mark against the rulebook. This was an AI coding procedure, not independent human review of every answer.

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 6,908 answers. Automated validation checked permitted codes and scores, placement of marks, and exact reconstruction of the source answer; all 153 output files passed. These structural checks do not establish semantic accuracy of every code.

A separate AI coder, without access to the first coding, recoded 681 production answers and 100 calibration answers. Across the 781 double-coded answers, Krippendorff’s alpha was 0.919 for topical marks, 0.926 for themes, 0.904 for emotional states, and 0.969 for scores, exceeding the prespecified 0.80 threshold. Production topical agreement was 0.916. These results measure consistency between AI coding runs; they are not independent human validation or proof of construct validity.

The two AI codings differed on at least one mark or score in 288 of the 681 production answers. High aggregate agreement can coexist with substantial record-level disagreement.

An author-ruled set of 48 difficult answers served as an additional check, and the final coding reproduced all 48 rulings. This targeted check does not establish accuracy across the full corpus.

On October 9, 2026, rulebook version 0.6 added 3,599 sentiment scores to previously unscored marks on 34 measures and an overall score to each of 6,874 substantive answers. Separate AI recoding of 777 answers yielded alpha values of 0.945 for overall scores and 0.930 for 429 paired additional measure scores. These additions were documented after the initial coding and should not be described as part of an unchanged original protocol.

A written response is an answer about the respondent’s own engagement coded to an indicator. Primary responses come from the designated section’s open question; secondary responses come from other questions. Counts describe the coded material and may include several answers per respondent. Structured ratings measure responses to their own wording, not the prevalence of every practice described in the comments. Fewer than 10 written responses support illustrative description only; larger qualitative samples remain self-selected.

Respondent quotations retain the supplied spelling and grammar. Ellipses identify omissions, and the supplied material includes the name substitutions described below. Attributions identify role, outsourcing model, and process family.

The study’s documented preparation process states that a third party replaced financial-institution and vendor names before the material reached the author. The quotations retain those supplied substitutions. This describes deidentification of the analytical material, not a finding that no identifying information existed at any earlier stage.

## 4.11 Limits of the Evidence

The study primarily records vendor staff members’ reported experiences, supplemented by the engagement-level fields in the analytical record. It does not include independent observation of delivery, inspection of institutional records, or validation by the financial institutions served.

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.

The analysis is cross-sectional. Associations do not establish temporal order or causal direction. Retrospective five-year items are perceptions collected at one time, not repeated observations; their interpretation is further limited by the short careers reported by many respondents.

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.

Ratings from the 56 canceled engagements were collected after cancellation and may reflect it. Cancellation associations are retrospective contrasts. Primary satisfaction associations use the 94 active engagements, which avoids combining those scores with post-cancellation scores but introduces its own restriction to surviving engagements.

The measures and indicators were strongly interrelated. None of the 300 partial indicator–outcome associations remained significant after false-discovery-rate correction. This limits claims of distinct explanatory contributions; it does not establish that independent effects are absent. Descriptive rankings of means or coefficients should not be interpreted as tested rankings of importance.

Role and experience were closely related. Mean career experience was 1.9 years for individual contributors, 10.0 for engagement managers, and 15.1 for portfolio directors. The reported unadjusted comparisons do not isolate their respective contributions. Other shared respondent and engagement characteristics may also affect associations.

Missing and nonnumeric responses alter analytical coverage. For example, 554 respondents rated whether the engagement was within budget, while 242 answered “I do not know.” Such an answer may reflect lack of information or uncertainty, rather than organizational ignorance. The study does not establish that excluded responses are missing at random.

Several groups have limited coverage. The two customer management executives support counts only. Portfolio-director analyses include at most 31 respondents, with clustering reducing effective information. Every institutional-type group contains fewer than 30 engagements, and three process families contain fewer than 10. Study thresholds determine which results are descriptive or eligible for inference.

Restricted variation limits inference. For example, every recorded answer on vendor experience in agentic AI outsourcing was two years or less (n = 571 respondents). Such a narrow range can reduce the ability to identify an association and cannot establish the value of substantially greater experience.

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 are self-selected and influenced by question framing. Frequency-theme counts depend on documented wording rules, including a post-coding expansion, and may miss descriptions outside those rules. AI coding agreement measures consistency, while structural validation checks format and source reconstruction; neither establishes independent semantic accuracy for the full corpus.

Disclosure is measured through respondents’ accounts of their own or management’s discussion. Structured answers and coded written accounts have different bases and may differ. Neither establishes all information available to a financial institution through other channels.

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.

The sample covers Build-Out, Transition/Cut-Over, and Production, not directly observed preparation or procurement stages. Return data comprise 30 production records, including 17 positive and 13 zero values, with incomplete documentation of the financial measure. Findings cannot establish industry-wide returns, prospective cancellation risk, or the effectiveness of the proposed executive actions.

← Appendix C: Tables and the Indicators Behind ThemAppendix E: Companion Indicator Statistical Analyses →