Manual alteration was reported across the full range of engagement indicator means, although four methods were more common on canceled engagements and engagements with lower indicator means. Respondents on all 56 canceled engagements described manual alteration of financial institution processes and AI outputs. They described manual alteration of financial-institution-provided data and reporting alteration on 55 of those engagements, and third-party work done by hand on 29.
The observed cancellation rate was higher among engagements on which respondents reported the four most common methods than among engagements with no such report. For AI-output manual alteration, cancellation was recorded on 45.90% of the 122 engagements with a report and on none of the 28 without one. The corresponding contrast for third-party work done by hand was smaller: 39.19% of 74 engagements with a report, compared with 35.53% of 76 without one. Absence of a report does not independently establish absence of the practice.
Method | Engagements with the method | Canceled | Engagements without the method | Canceled |
|---|---|---|---|---|
Manual alteration of financial-institution-provided data | 133 | 55 (41.35%) | 17 | 1 (5.88%) |
Manual alteration of financial institution processes | 138 | 56 (40.58%) | 12 | 0 (0.00%) |
Manual alteration of AI outputs | 122 | 56 (45.90%) | 28 | 0 (0.00%) |
Reporting alteration | 126 | 55 (43.65%) | 24 | 1 (4.17%) |
Third-party work done by hand | 74 | 29 (39.19%) | 76 | 27 (35.53%) |
Note: Cancellation was recorded on 56 of the 150 engagements. Method presence means at least one qualifying report on the engagement. The study observed each engagement once. These comparisons do not establish whether manual alteration preceded or followed cancellation. They are descriptive contrasts, not forecasts of cancellation.
Every indicator quintile included reports of manual alteration on all 30 engagements. An indicator quintile is one of five groups of 30 engagements ordered by engagement indicator mean. Four methods were less common in higher quintiles. AI-output manual alteration was reported on all 30 engagements in the lowest quintile and 13 in the highest. Third-party work done by hand showed no comparable ordered pattern, with 16 engagements in both the lowest and highest quintiles.
Method | Quintile 1 | Quintile 2 | Quintile 3 | Quintile 4 | Quintile 5 |
|---|---|---|---|---|---|
Manual alteration of financial-institution-provided data | 30 of 30 | 29 of 30 | 29 of 30 | 26 of 30 | 19 of 30 |
Manual alteration of financial institution processes | 30 of 30 | 30 of 30 | 30 of 30 | 29 of 30 | 19 of 30 |
Manual alteration of AI outputs | 30 of 30 | 29 of 30 | 27 of 30 | 23 of 30 | 13 of 30 |
Reporting alteration | 30 of 30 | 29 of 30 | 26 of 30 | 25 of 30 | 16 of 30 |
Third-party work done by hand | 16 of 30 | 17 of 30 | 13 of 30 | 12 of 30 | 16 of 30 |
Any of the five methods | 30 of 30 | 30 of 30 | 30 of 30 | 30 of 30 | 30 of 30 |
Note: Quintile 1 contains the 30 engagements with the lowest engagement indicator mean and quintile 5 the 30 with the highest. Cells count engagements with at least one qualifying report of the method. Methods overlap. Rows are not additive. Indicator quintiles describe relative positions within this study set rather than validated categories of engagement performance.