Source: https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/part-5-manual-alteration/the-five-manual-alteration-methods/
Author: Phil Hatch
Publisher: Akholi

The Five Manual Alteration Methods

The 2026 Report · Part 5

# The Five Manual Alteration Methods

Intervention in data, processes, AI outputs, reporting, and third-party work

By Phil Hatch · Akholi · Version 1.8.2

Complete study · PDFPart 5 · PDFStudy overview / Part 5 / The Five Manual Alteration Methods

Manual alteration of processes and data was reported on more engagements than any other method, although the five methods commonly overlapped. The evidence also differed by method. Personal manual alteration of data, processes, and AI outputs was asked through yes-or-no questions, with separate ratings of regular manual alteration by the wider delivery team. Reporting alteration and third-party work done by hand were identified in written answers. Their related ratings assess respondents’ judgments about reporting and the retention or replacement of manual practices. Table 5.1 distinguishes the occurrence counts from these related measures.

Table 5.1. The five manual alteration methods and their evidence

| Method | Occurrence evidence and related ratings | Respondents | Engagements |
| --- | --- | --- | --- |
| Manual alteration of financial-institution-provided data | Occurrence: C199 True answers and written answers. Related rating: C203. | 452 | 133 |
| Manual alteration of financial institution processes | Occurrence: C205 True answers and written answers. Related rating: C209. | 546 | 138 |
| Manual alteration of AI outputs | Occurrence: C211 True answers and written answers. Related rating: C215. | 328 | 122 |
| Reporting alteration | Occurrence: written answers, including C055. Related ratings: C052–C054. | 477 | 126 |
| Third-party work done by hand | Occurrence: written answers. Related ratings: C190–C192. | 108 | 74 |

Note: Counts identify respondents and engagements with at least one direct True answer or written description of the method on the engagement, from 850 respondents and 150 engagements. Counts overlap across methods. Related ratings do not add to the occurrence counts. C numbers identify survey questions, including rated indicators. The outsourcing models shown in the paper are business process outsourcing (BPO), information technology outsourcing (ITO), customer experience management (CXM), and knowledge process outsourcing (KPO).

> Weekly status decks reported case-processing speed against targets set for the manual process, so faster agent throughput looked like a win even on cases the agent had gotten factually wrong and compliance later had to unwind.” INDIVIDUAL CONTRIBUTOR · BPO, Financial crime and client onboarding

## 1 Manual Alteration of Financial-Institution-Provided Data

Manual alteration of financial-institution-provided data was reported on 133 of the 150 engagements (88.67%), often as recurring correction rather than an isolated intervention.

For personal manual alteration, the direct yes-or-no question recorded 421 True answers among 839 valid answers (50.18%), representing 127 engagements. The broader method-based compilation across written questions included 588 answers from 382 respondents on 128 engagements. Of those answers, 573 had negative sentiment and 223 described manual alteration every cycle or on most days.

Sentiment classifies an answer’s reported effect on the engagement or respondent on five values: −10, −5, 0, +5, and +10. Negative values indicate harm, zero a neutral assessment, and positive values benefit. An answer may have an overall sentiment score and separate scores for the survey subjects it addresses. These classifications describe the account and do not establish its truth, the legality of a practice, or the adequacy of disclosure.

> “I fix wrong Data every day, team lead know but we do not raise it with client, feedbacks would go nowhere anyway” -Individual Contributor, BPO, Trade, settlement and reconciliation

> “I flag it every week, my manager knows but nothing changes upstream. This Vendor takes zero ownership so we are the ones patching data every single day” -Individual Contributor, CXM, Technology engineering and operations

> “i correct the extracts before loading them, half the fields come swapped or blank before we even open the file,” -Individual Contributor, ITO, Technology engineering and operations

### 1.1 Personal Manual Alteration of Financial-Institution-Provided Data

Indicator C199 · I have altered client-provided data · Full statistical analysis ›

Half of the respondents answering the personal data question reported that they had manually altered financial-institution-provided data: 421 of 839 answered True (50.18%), while 418 answered False (49.82%). The True share was 69.62% among respondents on canceled engagements (n = 316 respondents), compared with 36.59% among those on active engagements in Transition/Cut-Over (n = 246). The notation n identifies the number of respondents or engagements in an analysis. These respondent-level shares describe personal answers, not the proportion of engagements assigned to a comparison group.

Figure 5.2. Share answering True to “I have altered client-provided data”, by respondent role and engagement status

Scroll horizontally to view the complete figure.

Personal manual alteration of financial-institution-provided data differed by role

Respondent role Share answering True

Lifecycle stage / status Share answering True

Exact chart data

| Series | Point index | Category | Value as stored |
| --- | --- | --- | --- |
| Share answering True | 0 | Portfolio director (n = 30) | 0 |
| Share answering True | 1 | Engagement manager (n = 150) | 2 |
| Share answering True | 2 | Individual contributor (n = 659) | 63.429400000000001 |

Exact chart data

| Series | Point index | Category | Value as stored |
| --- | --- | --- | --- |
| Share answering True | 0 | Canceled (n = 316) | 69.6203 |
| Share answering True | 1 | Production (n = 102) | 37.254899999999999 |
| Share answering True | 2 | Transition/Cut-Over (n = 246) | 36.5854 |
| Share answering True | 3 | Build-Out (n = 175) | 41.714300000000001 |

Note: True or Yes answers: 421 of 839 answers (50.18%).

n gives the numeric-answer base for each group. These are descriptive comparisons of reported answers.

The four core measures are cancellation rate, customer satisfaction, engagement indicator mean, and intent to leave. Cancellation rate is the share of a stated group of engagements canceled by the study date. Customer satisfaction is the figure recorded for each engagement on a 100-point scale, with the comparisons here restricted to active engagements. Engagement indicator mean summarizes respondents’ ratings across the 100 engagement indicators, averaged within each engagement. The mean is the arithmetic average of the recorded values, and the median is the middle value in an ordered distribution. Intent to leave is the respondent’s rating of the likelihood of leaving the current job within 12 months.

For categorical comparisons, an engagement is assigned the most frequent answer among its respondents, termed the plurality answer. Sentinel categories such as a preference not to answer remain answer categories under the existing rule, and tied engagements are excluded. This assignment differs from counting an engagement whenever any respondent reports a manual alteration method. An engagement-level analysis uses one summarized observation per engagement, while a respondent-level analysis uses paired answers from individuals.

A chi-square test examines whether a categorical distribution differs from independence. Welch F compares group means while allowing their variances, or spread of ratings, to differ. Degrees of freedom, df, describe the information available to estimate variation. A p value is the probability, under the null model, of a result at least as extreme as the observed result. Statistical significance does not establish the size, operational importance, or cause of a relationship.

Engagements assigned True on the personal data question had a higher observed cancellation rate and less favorable results on the other three core study measures. Cancellation was recorded on 61.84% of the 76 engagements assigned True, compared with 10.17% of the 59 assigned False (chi-square 37.19, df = 1, p < 0.001, n = 135 engagements). Among active engagements, customer satisfaction differed between the groups (Welch F 17.74, p < 0.001, n = 82 engagements). The groups also differed in engagement indicator mean (Welch F 93.60, p < 0.001) and intent to leave (Welch F 55.34, p < 0.001), each on 135 engagements. These associations describe reported experience measured after cancellation where an engagement had already been canceled. They do not establish whether manual alteration preceded cancellation or caused it.

Robustness checks repeat an analysis using alternative assumptions or selections. Their interpretation depends on the particular check and analytical unit. The linked indicator analyses provide the existing check results. The qualifications reported below, including failed or uncomputable checks and differences between engagement-level and respondent-level results, limit any general claim of stability.

The separate indicator-specific compilation for C199 contained 578 written answers from 378 respondents on 128 engagements, of which 571 received negative sentiment scores for this subject. Respondents described data arriving unusable despite earlier cleaning and access restrictions that prevented correction at the source. This indicator-specific count differs from the broader method-based compilation above and is not an additional set of incidents.

> “Last cycle their file was missing tax ids for half the batch, I fixed it myself before run started... nothing gets checked before it lands on our desk, this happens basically every cycle now” -Individual Contributor, BPO, Corporate and back-office functions

> “Last cycle it was three whole cohort missing from one file, I only catch it because I count row against last cycle total and numbers didnt match up, if I hadnt counted it would have gone straight through with people missing pay.” -Individual Contributor, BPO, Corporate and back-office functions

> “last month the cash statement came after cut off again... so i take previous day file and roll forward, thus the break clears on time.” -Individual Contributor, BPO, Trade, settlement and reconciliation

### 1.2 Regular Manual Alteration of Financial-Institution-Provided Data by the Wider Delivery Team

Indicator C203 · My greater team regularly alters client provided data · Full statistical analysis ›

Respondents described data manual alteration as a wider delivery-team practice as well as a personal activity. Agreement with the statement about regular manual alteration had a mean rating of 5.4 and a median of 5.0 (n = 773 respondents). The mean was 6.9 among respondents on canceled engagements (n = 287). Higher ratings indicate stronger agreement that regular manual alteration occurred. The question has no study-defined favorable or unfavorable direction, so these ratings are not classified as distressed.

Figure 5.3. Regular Manual Alteration of Financial-Institution-Provided Data by the Wider Delivery Team: associations with the study measures

Scroll horizontally to view the complete figure.

Regular manual alteration of financial-institution-provided data by the wider delivery team: associations with the study measures

Scroll horizontally to view the complete comparison.

| Comparison | Correlation r | Estimate and analysis base |
| --- | --- | --- |
| Cancellation | | r = 0.70 150 engagements p < 0.001 |
| Customer satisfaction | | r = −0.65 94 active engagements p < 0.001 |
| Indicator mean | | r = −0.84 150 engagements p < 0.001 |
| Intent to leave | | r = 0.66 722 respondents p < 0.001 |
| | −1 0 +1 | |

Exact comparison data

Cancellation

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | 0.69758210399999998 | 1 |

Customer satisfaction

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | -0.6529784917 | 1 |

Indicator mean

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | -0.84431032939999995 | 1 |

Intent to leave

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | 0.66159079249999997 | 1 |

Note: Mean rating: 5.4 (n = 773 respondents).

Survey question: “My greater team regularly alters client provided data”.

Group means and numeric-answer bases: Individual contributor (n = 622): 5.70. Engagement manager (n = 131): 4.28. Portfolio director (n = 20): 2.45. Build-Out (n = 161): 4.68. Transition/Cut-Over (n = 232): 4.48. Production (n = 93): 4.24. Canceled (n = 287): 6.87.

r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational. A confidence interval expresses uncertainty under the stated method. A p value is the probability, under the null model, of a result at least as extreme as the observed result.

A mean is the arithmetic average and a median is the middle ordered rating. The coefficient r measures the direction and strength of a linear association, ranging from −1 to +1. A coefficient near either end describes a stronger linear relationship, while its sign identifies the direction. It does not establish causation.

The correlation coefficient r describes the direction and strength of a linear association. Stronger agreement was associated with higher cancellation rates (r = 0.70, n = 150 engagements) and a lower engagement indicator mean (r = −0.84, n = 150). Among active engagements, it was associated with lower customer satisfaction (r = −0.65, n = 94 engagements). At respondent level, stronger agreement was associated with higher intent to leave (r = 0.66, n = 722 respondents). The distinction between personal correction and a regular delivery-team practice is substantive, although these associations do not establish the sequence of events.

Of the 84 written answers coded to regular data manual alteration, 83 had negative sentiment. Respondents described data that required cleaning again after each delivery, with manual alteration known to the team lead but seldom logged or disclosed to the financial institution. The single positive answer described a data-exception procedure agreed with the financial institution’s platform owner before correction.

> “I keep note of fixes i make each sprint, my lead know about it, client does not. Take maybe extra hour every day and i do not think should be my job to clean their data before agent see it.” -Individual Contributor, ITO, Technology engineering and operations

> The vendor master data is a mess every time it lands with us. We clean it up each cycle and it just arrives dirty again. Their security team wont approve us a proper connection either, so I dont wait anymore, I just fix it myself.” INDIVIDUAL CONTRIBUTOR · BPO, Corporate and back-office functions

## 2 Manual Alteration of Financial Institution Processes

Manual alteration of financial institution processes was the most widely reported method, occurring in respondents’ accounts on 138 of the 150 engagements (92.00%).

For personal manual alteration, the direct question recorded 490 True answers among 841 valid answers (58.26%), representing 132 engagements. The method-based compilation contained 857 written answers from 515 respondents on 137 engagements, including 826 classified as negative. In 143 answers, respondents described manual alteration every cycle or on most days.

> “nobody upstream will say yes or no so i just go with my own call, decision take too long to come back and Payroll deadline is same every cycle so i cannot wait” -Individual Contributor, BPO, Corporate and back-office functions

> “This is basically every week now, my manager knows about it. Ticket takes days to clear, fix itself takes an hour, so waiting on the ticket process makes no sense when the job needs doing” -Individual Contributor, ITO, Technology engineering and operations

### 2.1 Personal Manual Alteration of Financial Institution Processes

Indicator C205 · I have altered client processes · Full statistical analysis ›

A majority of respondents answering the personal process question reported manual alteration: 490 of 841 answered True (58.26%), while 351 answered False (41.74%). The True share was 73.73% among respondents on canceled engagements (n = 316), compared with 45.14% among those on active engagements in Build-Out (n = 175). By outsourcing model, it ranged from 69.42% among respondents on BPO engagements (n = 206) to 50.87% on CXM engagements (n = 230).

Figure 5.4. Share answering True to “I have altered client processes”, by respondent role and engagement status

Scroll horizontally to view the complete figure.

A majority of respondents answering the personal process question reported manual alteration: 490 of 841 answered True (58.26%), while 351 answered False (41.74%).

Respondent role Share answering True

Lifecycle stage / status Share answering True

Exact chart data

| Series | Point index | Category | Value as stored |
| --- | --- | --- | --- |
| Share answering True | 0 | Portfolio director (n = 31) | 0 |
| Share answering True | 1 | Engagement manager (n = 148) | 6.0811000000000002 |
| Share answering True | 2 | Individual contributor (n = 662) | 72.658600000000007 |

Exact chart data

| Series | Point index | Category | Value as stored |
| --- | --- | --- | --- |
| Share answering True | 0 | Canceled (n = 316) | 73.734200000000001 |
| Share answering True | 1 | Production (n = 103) | 45.631100000000004 |
| Share answering True | 2 | Transition/Cut-Over (n = 247) | 53.0364 |
| Share answering True | 3 | Build-Out (n = 175) | 45.142899999999997 |

Note: True or Yes answers: 490 of 841 answers (58.26%).

n gives the numeric-answer base for each group. These are descriptive comparisons of reported answers.

Engagements assigned True on the process question had a cancellation rate of 50.00%, compared with 15.22% for engagements assigned False (94 and 46 engagements, respectively). The difference was supported by the existing categorical comparison (chi-square 15.77, df = 1, p < 0.001, n = 140 engagements). Active-engagement customer satisfaction also differed (Welch F 17.72, p < 0.001, n = 86 engagements), as did engagement indicator mean (Welch F 64.36, p < 0.001) and intent to leave (Welch F 36.02, p < 0.001), each on 140 engagements. These are associations with the assigned answer, not evidence that process manual alteration caused the outcomes.

The indicator-specific compilation for C205 contained 849 written answers from 513 respondents on 137 engagements, including 841 with negative sentiment scores for this subject. Respondents described delayed financial institution decisions, change approval designed for infrequent changes, and procedures that were undocumented or inconsistent with practice. They reported changing processes before recording the changes, or without recording them. The method-based compilation above and this indicator-specific compilation are separate summaries. Their different sentiment counts have not been reconciled and must not be interpreted as a whole and its subset.

> “Escalating a break means waiting on someone who never actually rules on it, we just pick a call and run with it. And whatever that forces into the workflow gets written up later, if it gets written up at all.” -Individual Contributor, BPO, Trade, settlement and reconciliation

> “i changed a step on the reconciliation process last week, nothing about the change process here really matches how the agent behaves day to day, so we adjust as we go and i discuss about it with my lead, its not a big drama” -Individual Contributor, BPO, Trade, settlement and reconciliation

### 2.2 Regular Manual Alteration of Financial Institution Processes by the Wider Delivery Team

Indicator C209 · My greater team regularly alters client processes · Full statistical analysis ›

Regular process manual alteration received stronger agreement than the corresponding questions on data or AI outputs. Its mean rating was 5.7 and its median was 6.0 (n = 779 respondents). Respondents on canceled engagements reported a mean of 7.1 (n = 291). Higher ratings indicate stronger agreement that the wider delivery team regularly performed manual alteration. The question has no study-defined favorable or unfavorable direction.

Figure 5.5. Regular Manual Alteration of Financial Institution Processes by the Wider Delivery Team: associations with the study measures

Scroll horizontally to view the complete figure.

Regular manual alteration of financial institution processes by the wider delivery team: associations with the study measures

Scroll horizontally to view the complete comparison.

| Comparison | Correlation r | Estimate and analysis base |
| --- | --- | --- |
| Cancellation | | r = 0.63 150 engagements p < 0.001 |
| Customer satisfaction | | r = −0.64 94 active engagements p < 0.001 |
| Indicator mean | | r = −0.83 150 engagements p < 0.001 |
| Intent to leave | | r = 0.65 730 respondents p < 0.001 |
| | −1 0 +1 | |

Exact comparison data

Cancellation

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | 0.62819249509999997 | 1 |

Customer satisfaction

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | -0.63864880280000003 | 1 |

Indicator mean

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | -0.82606459359999995 | 1 |

Intent to leave

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | 0.65294885329999997 | 1 |

Note: Mean rating: 5.7 (n = 779 respondents).

Survey question: “My greater team regularly alters client processes”.

Group means and numeric-answer bases: Individual contributor (n = 631): 6.00. Engagement manager (n = 130): 4.56. Portfolio director (n = 18): 3.00. Build-Out (n = 161): 4.63. Transition/Cut-Over (n = 230): 4.84. Production (n = 97): 5.23. Canceled (n = 291): 7.10.

r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational.

Stronger agreement was associated with a lower engagement indicator mean (r = −0.83, n = 150 engagements) and a higher cancellation rate (r = 0.63, n = 150). Among active engagements, it was associated with lower customer satisfaction (r = −0.64, n = 94 engagements). It was also associated with higher intent to leave at respondent level (r = 0.65, n = 730 respondents).

Of the 111 written answers coded to regular process manual alteration, 110 had negative sentiment. Respondents described routine changes made by the delivery team when financial institution decisions were delayed. The team lead was described as knowing about these changes, while disclosure to the financial institution was uncommon in the accounts.

> “This is normal now, my team decide small things ourselves and tell manager after not before. I don't ask if client know, not my place. Happens most weeks, only way to hit the cycle.” -Individual Contributor, BPO, Corporate and back-office functions

> I have always felt confident about how change is handled on this account. When the client's watchlist data arrives incomplete, the team raises it with the client's financial-crime unit, we discuss about the gap together, and any adjustment to triage process is agreed and logged before it is applied. We do not revert back to old workflows without that sign-off. Precedent across our sanctions engagements shows this is consistently how it works, five or six times a quarter at most.” PORTFOLIO DIRECTOR · KPO, Financial crime and client onboarding

## 3 Manual Alteration of AI Outputs

Manual alteration of AI outputs was reported on 122 of the 150 engagements (81.33%), although fewer respondents personally reported this method than personal manual alteration of data or processes.

For personal manual alteration, the direct question recorded 278 True answers among 828 valid answers (33.57%), representing 108 engagements. The method-based compilation contained 429 written answers from 289 respondents on 121 engagements. Of those answers, 413 had negative sentiment and 95 described manual alteration every cycle or on most days.

Within the method-based compilation, 14 written answers on nine engagements described instructions to continue processing outputs without raising the problem. The reasons compilation later in the paper reports a different count for this theme. The two summaries are identified separately because their coverage has not been reconciled.

> “Nobody trusts the raw output so I rewrite it every time, that's the real Job now not what they think we do” -Individual Contributor, CXM, Corporate and back-office functions

> “the problem is nobody talk about how often team do this because everyone just fix quietly on their own, do understand I only track for myself, I do it almost every case I recieve” -Individual Contributor, BPO, Corporate and back-office functions

### 3.1 Personal Manual Alteration of AI Outputs

Indicator C211 · I have altered AI outputs · Full statistical analysis ›

One-third of respondents answering the personal AI-output question reported manual alteration: 278 of 828 answered True (33.57%), while 550 answered False (66.43%). The True share was 54.19% among respondents on canceled engagements (n = 310), compared with 18.52% among those on active engagements in Transition/Cut-Over (n = 243).

Figure 5.6. Share answering True to “I have altered AI outputs”, by respondent role and engagement status

Scroll horizontally to view the complete figure.

One-third of respondents answering the personal AI-output question reported manual alteration: 278 of 828 answered True (33.57%), while 550 answered False (66.43%).

Respondent role Share answering True

Lifecycle stage / status Share answering True

Exact chart data

| Series | Point index | Category | Value as stored |
| --- | --- | --- | --- |
| Share answering True | 0 | Portfolio director (n = 31) | 0 |
| Share answering True | 1 | Engagement manager (n = 147) | 1.3605 |
| Share answering True | 2 | Individual contributor (n = 650) | 42.461500000000001 |

Exact chart data

| Series | Point index | Category | Value as stored |
| --- | --- | --- | --- |
| Share answering True | 0 | Canceled (n = 310) | 54.1935 |
| Share answering True | 1 | Production (n = 103) | 28.1553 |
| Share answering True | 2 | Transition/Cut-Over (n = 243) | 18.5185 |
| Share answering True | 3 | Build-Out (n = 172) | 20.930199999999999 |

Note: True or Yes answers: 278 of 828 answers (33.57%).

n gives the numeric-answer base for each group. These are descriptive comparisons of reported answers.

The cancellation rate was 76.19% among the 42 engagements assigned True on the AI-output question, compared with 22.00% among the 100 assigned False (chi-square 36.86, df = 1, p < 0.001, n = 142 engagements). Active-engagement customer satisfaction differed between the groups (Welch F 11.44, p = 0.005, n = 88 engagements). Engagement indicator mean (Welch F 87.54, p < 0.001) and intent to leave (Welch F 79.99, p < 0.001) also differed, each on 142 engagements. The direction of these associations is consistent with the accounts of difficult delivery, but it does not establish the timing or causal role of manual alteration.

The indicator-specific compilation for C211 contained 407 written answers from 276 respondents on 116 engagements, including 399 with negative sentiment. Respondents described inconsistent outputs from repeated runs on the same input. They reported selecting and correcting outputs by hand before passing them onward because the approval process was slower than the delivery deadline. These counts refer to C211 coding rather than the broader AI-output method compilation.

> “the problem is the ai gives a different match each time i run it, so i correct it myself before it moves on, on that basis i do not wait for approval” -Individual Contributor, BPO, Trade, settlement and reconciliation

> “So what I do is run it, then run it again, and if the two agree I go with it, and if they don't I work it out by hand. That's my process. Nobody designed that, I invented it because I needed something.” -Individual Contributor, BPO, Trade, settlement and reconciliation

> “the problem is change control takes weeks for anything, so whats the point raising it every time, i patch the output myself and keep the queue moving, thats just how this job works and truthfully im quicker at catching it than the process is anyway” -Individual Contributor, CXM, Corporate and back-office functions

### 3.2 Regular Manual Alteration of AI Outputs by the Wider Delivery Team

Indicator C215 · My greater team regularly alters AI outputs · Full statistical analysis ›

Regular manual alteration of AI outputs received the weakest agreement among the three wider delivery-team questions. The mean rating was 4.4 and the median was 4.0 (n = 665 respondents), compared with means of 5.4 for data and 5.7 for processes. Respondents on canceled engagements reported a mean of 5.8 (n = 228). Higher ratings indicate stronger agreement that manual alteration was regular. The question has no study-defined favorable or unfavorable direction.

Figure 5.7. Regular Manual Alteration of AI Outputs by the Wider Delivery Team: associations with the study measures

Scroll horizontally to view the complete figure.

Regular manual alteration of AI outputs by the wider delivery team: associations with the study measures

Scroll horizontally to view the complete comparison.

| Comparison | Correlation r | Estimate and analysis base |
| --- | --- | --- |
| Cancellation | | r = 0.59 150 engagements p < 0.001 |
| Customer satisfaction | | r = −0.40 94 active engagements p < 0.001 |
| Indicator mean | | r = −0.66 150 engagements p < 0.001 |
| Intent to leave | | r = 0.58 620 respondents p < 0.001 |
| | −1 0 +1 | |

Exact comparison data

Cancellation

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | 0.58788869899999996 | 1 |

Customer satisfaction

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | -0.4043261836 | 1 |

Indicator mean

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | -0.6564147239 | 1 |

Intent to leave

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | 0.58313727100000001 | 1 |

Note: Mean rating: 4.4 (n = 665 respondents).

Survey question: “My greater team regularly alters AI outputs”.

Group means and numeric-answer bases: Individual contributor (n = 556): 4.58. Engagement manager (n = 99): 3.59. Portfolio director (n = 10): 2.50. Build-Out (n = 150): 3.41. Transition/Cut-Over (n = 201): 3.61. Production (n = 86): 4.35. Canceled (n = 228): 5.77.

r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational.

Stronger agreement was associated with a higher cancellation rate (r = 0.59, n = 150 engagements) and higher intent to leave (r = 0.58, n = 620 respondents). Its association with customer satisfaction among active engagements was negative (r = −0.40, n = 94 engagements) and weaker than for the corresponding data and process questions. Agreement was also associated with a lower engagement indicator mean (r = −0.66, n = 150 engagements).

All 23 written answers coded to this indicator had negative sentiment. The 19 that addressed AI-output manual alteration described continuing delivery-team correction before outputs were seen outside the delivery team, with little recording of that work.

> “it's f***ing unpredictable, same input different answer. i fix it quiet every time, my lead knows, client doesn't. that's just how it is here.” -Individual Contributor, CXM, Corporate and back-office functions

> What would help is one person owning this end to end. Right now delivery lead, the AI team and quality all give different instruction and we follow whichever came last. Second, when the assist gives a wrong answer to the floor it must be somebody's job to fix the source, not mine to patch it in the ticket, we do that all week and nothing changes upstream” INDIVIDUAL CONTRIBUTOR · CXM, Corporate and back-office functions

## 4 Reporting Alteration

Respondents described reporting alteration on 126 of the 150 engagements (84.00%), including methods that obscured manual correction or overstated agent-completed work.

In the reporting-method question, 484 of 775 respondents named a method (62.45%), while 291 answered “I do not know.” A separate compilation of descriptions concerning the respondent’s own engagement contained 486 written answers from 477 respondents on 126 engagements, including 485 with negative sentiment. Naming a method and describing it on one’s own engagement are different count populations.

> “The vendor rolled unresolved reconciliation exceptions into the next period's dashboard so the error count never showed as owned, and it took a manual pull of the raw exception log to see the real backlog.” -Engagement Manager, BPO, Corporate and back-office functions

### 4.1 Reporting Methods Described by Respondents

Substituting activity counts for outcomes was the most frequently described reporting method, appearing in 112 answers on 78 engagements. This method is termed proxy substitution. Category conflation, which counted human-corrected work as the agent’s, appeared in 52 answers on 45 engagements. Denominator exclusion, which removed failures from the base of a reported rate, appeared in 50 answers on 39 engagements. Table 5.2 defines the reporting methods identified in the coded written answers.

Table 5.2. Reporting methods described in the written answers

| Reporting method | Description | Answers | Engagements |
| --- | --- | --- | --- |
| Proxy substitution | An activity count stands in for an outcome. | 112 | 78 |
| Category conflation | Human-corrected, sandbox or overridden work is counted as agent-completed production work. | 52 | 45 |
| Denominator exclusion | Failures or exceptions are removed from the base of a reported rate. | 50 | 39 |
| Scope restriction | The measure covers only the easy or parseable subset. | 38 | 32 |
| Manual effort omitted from reported measures | The labor does not appear in any reported measure. | 27 | 25 |
| Exceptions omitted from financial institution reports | The exception is recorded internally but omitted from what the financial institution receives. | 23 | 20 |
| Status misrepresentation | A readiness or completion status is reported falsely, letting a gate pass. | 20 | 14 |
| Validation-set selection | Accuracy is reported on a set chosen to exclude the hardest cases. | 18 | 16 |
| Clock reset | Elapsed-time measurement restarts on rework so delay never accrues. | 4 | 3 |
| Aging suppressed | Open items are pooled into a category that does not age, so duration never surfaces. | 3 | 3 |

Note: The base is 486 written answers describing reporting alteration on the respondent’s own engagement, covering 126 engagements. An answer may describe several methods. The method labels summarize coded descriptions, rather than independently verified reporting practices.

> “The vendor rolled settlement fails that were manually corrected by a human into the same success bucket as fails the agent resolved on its own, which overstated how much the agent was actually doing correctly.” -Individual Contributor, BPO, Trade, settlement and reconciliation

> “The vendor's settlement dashboard reported the fails rate as a percentage of trades processed that day rather than trades still open, so a growing backlog of unresolved fails never showed up as a rising number.” -Engagement Manager, BPO, Trade, settlement and reconciliation

> “The help-desk dashboard reports agent deflection rate as tickets the agent handled without escalation, but tickets the agent closes and the same employee reopens within 24 hours are counted as a new, separate ticket rather than a reopened one, which hides the actual repeat-contact rate the deflection number is supposed to reflect.” -Engagement Manager, CXM, Corporate and back-office functions

### 4.2 Accuracy of Dashboards and Performance Reports

Indicator C052 · The dashboard and performance reports given to the customer are accurate · Full statistical analysis ›

Respondents’ ratings of dashboard and performance-report accuracy indicated substantial reservations about the reports the financial institution received. The mean rating was 4.8 and the median was 5.0 (n = 794 respondents). Ratings were distressed in 220 answers (27.71%). Respondents on canceled engagements reported a mean of 3.5 (n = 301). These ratings describe perceived accuracy rather than an independent audit of the reports.

Figure 5.8. Accuracy of Dashboards and Performance Reports: associations with the study measures

Scroll horizontally to view the complete figure.

Accuracy of dashboards and performance reports: associations with the study measures

Scroll horizontally to view the complete comparison.

| Comparison | Correlation r | Estimate and analysis base |
| --- | --- | --- |
| Cancellation | | r = −0.61 150 engagements p < 0.001 |
| Customer satisfaction | | r = 0.65 94 active engagements p < 0.001 |
| Indicator mean | | r = 0.81 150 engagements p < 0.001 |
| Intent to leave | | r = −0.62 745 respondents p < 0.001 |
| | −1 0 +1 | |

Exact comparison data

Cancellation

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | -0.61354076930000001 | 1 |

Customer satisfaction

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | 0.65147018990000005 | 1 |

Indicator mean

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | 0.81057508629999997 | 1 |

Intent to leave

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | -0.62398229329999999 | 1 |

Note: Survey question: “The dashboard and performance reports given to the customer are accurate”.

Distressed ratings: 220 of 794 answers (27.71%).

Group means and numeric-answer bases: Individual contributor (n = 615): 4.66. Engagement manager (n = 150): 5.16. Portfolio director (n = 27): 6.33. Build-Out (n = 162): 5.85. Transition/Cut-Over (n = 233): 5.46. Production (n = 98): 5.67. Canceled (n = 301): 3.47.

r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational.

Distressed ratings are ratings of 3 or below on indicators for which a higher rating is favorable. The percentage uses valid respondent ratings as its base. An engagement with a distressed mean is a separate engagement-level classification. For indicators converted to the favorable direction, the same threshold applies after conversion.

Higher perceived report accuracy was associated with a lower cancellation rate (r = −0.61, n = 150 engagements) and a higher engagement indicator mean (r = 0.81, n = 150). It was also associated with higher customer satisfaction among active engagements (r = 0.65, n = 94 engagements) and lower intent to leave (r = −0.62, n = 745 respondents). The engagement mean for this indicator was distressed on 29 of the 150 engagements.

The indicator-specific compilation contained 468 written answers from 453 respondents on 125 engagements, including 467 with negative sentiment. Respondents described measures that counted a task as complete when an agent acted, excluded cases requiring human correction, or blended and reclassified failures. The single positive answer concerned figures the respondent could substantiate.

> “The vendor's status dashboard counted a task as complete the moment the agent submitted an action, before any human check confirmed the pay run was correct, so failed items looked closed until someone dug into the underlying case log.” -Individual Contributor, BPO, Corporate and back-office functions

> “The vendor's weekly report calculates average handle time using only tickets the agent fully resolves on its own, so the growing share of tickets the agent partially processes and then silently hands back to a human queue never enters the handle-time calculation at all, keeping the headline average artificially low.” -Individual Contributor, CXM, Corporate and back-office functions

> “The vendor's monthly billing accuracy dashboard reports fee exceptions only after they clear a manual override queue, so exceptions the agent catches and auto-corrects same-day never post as exceptions at all, making the vendor's own error rate look better than the fee leakage the agent is actually finding.” -Engagement Manager, BPO, Payments, cards and disputes

### 4.3 Reporting Alteration in Dashboards and Performance Reports

Indicator C053 · The vendor manipulates the dashboard and performance reports given to the customer · Full statistical analysis ›

Respondents’ assessments of reporting alteration were less favorable on canceled engagements. After conversion to the favorable direction, the overall mean rating was 4.7 and the median was 5.0 (n = 786 respondents). The mean among respondents on canceled engagements was 3.4 (n = 302). The conversion subtracts the original rating from 11, so a higher converted rating indicates less perceived reporting alteration. It does not turn a reported judgment into a verified occurrence count.

Figure 5.9. Reporting Alteration in Dashboards and Performance Reports: associations with the study measures

Scroll horizontally to view the complete figure.

Reporting alteration in dashboards and performance reports: associations with the study measures

Scroll horizontally to view the complete comparison.

| Comparison | Correlation r | Estimate and analysis base |
| --- | --- | --- |
| Cancellation | | r = −0.61 150 engagements p < 0.001 |
| Customer satisfaction | | r = 0.62 94 active engagements p < 0.001 |
| Indicator mean | | r = 0.79 150 engagements p < 0.001 |
| Intent to leave | | r = −0.62 739 respondents p < 0.001 |
| | −1 0 +1 | |

Exact comparison data

Cancellation

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | -0.60785053450000004 | 1 |

Customer satisfaction

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | 0.6193936605 | 1 |

Indicator mean

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | 0.78805765910000003 | 1 |

Intent to leave

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | -0.61721375950000001 | 1 |

Note: Survey question: “The vendor manipulates the dashboard and performance reports given to the customer”.

Distressed ratings: 235 of 786 answers (29.90%).

Group means and numeric-answer bases: Individual contributor (n = 613): 4.61. Engagement manager (n = 150): 5.19. Portfolio director (n = 22): 5.41. Build-Out (n = 168): 5.52. Transition/Cut-Over (n = 220): 5.56. Production (n = 96): 5.78. Canceled (n = 302): 3.38.

r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational.

The indicator is converted to the favorable direction as 11 minus the original rating before these comparisons.

Higher converted ratings were associated with a lower cancellation rate (r = −0.61, n = 150 engagements) and a higher engagement indicator mean (r = 0.79, n = 150). They were also associated with higher active-engagement customer satisfaction (r = 0.62, n = 94 engagements) and lower intent to leave (r = −0.62, n = 739 respondents). Distressed ratings were recorded in 235 of the 786 answers (29.90%), and the engagement mean was distressed on 33 of the 150 engagements. For this converted indicator, distressed ratings of 3 or below correspond to original ratings above 7.

The indicator-specific compilation contained 465 written answers from 456 respondents on 125 engagements, including 464 with negative sentiment. Respondents described the use of activity counts, exclusion of failures, and attribution of human-corrected work to the agent in ways that maintained favorable reported figures.

> “Purchase requisitions the agent could not match to a valid supplier record were logged as processed with a manual override flag that never surfaced in the weekly summary the customer received.” -Individual Contributor, CXM, Corporate and back-office functions

> “The vendor's exception dashboard groups unresolved custody breaks under a catch-all pending category that does not age, so a break sitting open for weeks looks the same as one opened yesterday.” -Individual Contributor, BPO, Trade, settlement and reconciliation

> “The vendor's monthly performance report calculates first-contact resolution using only tickets the agent both received and closed within the same reporting window, so any ticket the agent picks up late in the month and carries into the next cycle simply disappears from both months' denominators instead of counting as a miss.” -Individual Contributor, CXM, Technology engineering and operations

> “The vendor's weekly ops dashboard tracks fee break resolution time starting from ticket creation rather than from when the agent first flagged the anomaly internally, so breaks the agent catches and holds for days before a ticket gets opened never count against the resolution clock at all.” -Individual Contributor, BPO, Payments, cards and disputes

### 4.4 Retrospective Comparison of Reporting Alteration

Indicator C054 · Compared to 5 years ago, is manipulation of the dashboards and performance reports given to the customer more or less common now? · Full statistical analysis ›

Respondents' retrospective judgments of reporting alteration were less favorable on engagements with less favorable reported outcomes. The converted mean rating was 4.8 and the median was 5.0 (n = 795 respondents). A higher converted rating indicates less perceived reporting alteration relative to five years earlier. It was associated with a higher engagement indicator mean (r = 0.79, n = 150 engagements) and a lower cancellation rate (r = −0.60, n = 150 engagements). Among active engagements, it was associated with higher customer satisfaction (r = 0.66, n = 94 engagements). At respondent level, it was associated with lower intent to leave (r = −0.60, n = 747 respondents). Ratings were distressed in 247 answers (31.07%), and the engagement mean was distressed on 35 of the 150 engagements.

Figure 5.10. Retrospective Comparison of Reporting Alteration: associations with the study measures

Scroll horizontally to view the complete figure.

Retrospective comparison of reporting alteration: associations with the study measures

Scroll horizontally to view the complete comparison.

| Comparison | Correlation r | Estimate and analysis base |
| --- | --- | --- |
| Cancellation | | r = −0.60 150 engagements p < 0.001 |
| Customer satisfaction | | r = 0.66 94 active engagements p < 0.001 |
| Indicator mean | | r = 0.79 150 engagements p < 0.001 |
| Intent to leave | | r = −0.60 747 respondents p < 0.001 |
| | −1 0 +1 | |

Exact comparison data

Cancellation

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | -0.59980415259999997 | 1 |

Customer satisfaction

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | 0.65879625730000002 | 1 |

Indicator mean

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | 0.79416703109999998 | 1 |

Intent to leave

Exact chart data

| Series | Point index | X | Y |
| --- | --- | --- | --- |
| Reported association | 0 | -0.60325554189999997 | 1 |

Note: Survey question: “Compared to 5 years ago, is manipulation of the dashboards and performance reports given to the customer more or less common now?”.

Distressed ratings: 247 of 795 answers (31.07%).

Group means and numeric-answer bases: Individual contributor (n = 619): 4.67. Engagement manager (n = 150): 5.25. Portfolio director (n = 24): 5.96. Build-Out (n = 164): 5.65. Transition/Cut-Over (n = 232): 5.57. Production (n = 98): 5.95. Canceled (n = 301): 3.43.

r is Pearson’s correlation coefficient, measuring signed linear association from −1 to +1. Whiskers show 95% confidence intervals where reported. Intent-to-leave intervals, where shown, use a cluster bootstrap. The bootstrap resamples engagements to preserve the grouping of respondent answers. Associations are observational.

The indicator is converted to the favorable direction as 11 minus the original rating before these comparisons.

No written answer addressed this comparison. The figures retain the full answering base and do not establish that every respondent had sufficient work experience to make a five-year comparison. They describe retrospective judgments, not a measured change over five years. The present results must not be interpreted as an analysis restricted to respondents with an established five-year work history. Part 3, “Perceived Reporting Change Compared With Five Years Earlier,” section 2.3 under “Oversight of Reporting Provided to the Financial Institution,” uses an experience-eligible base of 176 respondents and examines the governance, oversight, and control score. That analysis has a different base and outcome from the full-answering-base comparisons reported here. Its raw direction also differs from the favorable-direction conversion used in this section.

## 5 Third-Party Work Done by Hand

Third-party work done by hand was described on 74 of the 150 engagements (49.33%), making it the least widely reported of the five methods.

The method-based compilation included 113 written answers from 108 respondents on those engagements, of which 95 answers had negative sentiment. Most descriptions appeared in the comment question on the third-party ecosystem (84 of 113 answers).

### 5.1 Third-Party Work Described by Respondents

Respondents described performing work by hand when third-party systems did not deliver it in a form an agent could use. They reported retrieving records individually from provider portals without an interface, entering data when feeds failed or changed format, and repeating work after unannounced provider changes. In 31 of the 113 answers, the work occurred every cycle or on most days. In 18 answers, respondents said the change became apparent only after a downstream failure.

> “administrator portal is same as day one... still no proper interface, we pull screens by hand at cut off.” -Individual Contributor, BPO, Trade, settlement and reconciliation

> “Everything we use sits on one account with one provider. When that access got suspended in June nobody has a second way in, my team did the tickets by hand for two days” -Individual Contributor, CXM, Corporate and back-office functions

> “I score that section down for one reason only, nobody wrote down what old ops team use to patch by hand before the monitoring agent came” -Individual Contributor, ITO, Technology engineering and operations

### 5.2 Ratings Related to Third-Party Work Done by Hand

Ratings indicated gaps in retaining manual capability, knowing where informal corrections had occurred, and replacing those corrections with explicit controls. Each of the three indicators had a mean below the midpoint of the 10-point scale. They assess different aspects of the delivery team’s preparation for third-party problems and are considered separately after the comparison in Table 5.3.

Table 5.3. Ratings related to third-party work done by hand

| Indicator | Mean | Distressed ratings | r, cancellation | r, customer satisfaction | r, intent to leave |
| --- | --- | --- | --- | --- | --- |
| C190 Retention of the Ability to Perform Manual Workarounds | 4.66 | 249 of 808 (30.82%) | −0.65 | 0.72 | −0.67 |
| C191 Knowledge of Informal Third-Party Corrections | 4.61 | 262 of 792 (33.08%) | −0.66 | 0.61 | −0.66 |
| C192 Replacement of Informal Corrections With Explicit Controls | 4.58 | 267 of 796 (33.54%) | −0.64 | 0.71 | −0.65 |

Note: Means are respondent-level arithmetic means, with bases of 808, 792, and 796 for C190, C191, and C192, respectively. Distressed ratings are ratings of 3 or below. The coefficient r measures the direction and strength of a linear association, from −1 to +1. Cancellation correlations use 150 engagements and customer-satisfaction correlations use the 94 active engagements. Intent-to-leave correlations use paired respondent answers: 757 for C190, 746 for C191, and 748 for C192. Related ratings do not establish the occurrence of third-party work done by hand. Each indicator’s separate discussion below also reports its relationship with engagement indicator mean.

A 95% confidence interval (CI) expresses the uncertainty around an estimated association under the stated method. Cluster-bootstrap intervals resample whole engagements to retain the grouping of respondents within engagements. The intervals below are existing reported results. Correlations with engagement indicator mean exclude the indicator being tested to avoid including it in both quantities.

### 5.3 Retention of the Ability to Perform Manual Workarounds

Retained manual capability received 249 distressed ratings among 808 answers (30.82%), indicating gaps in the delivery team’s reported ability to perform work by hand when third-party systems failed. The mean rating was 4.66 and the median was 5.0. Respondents on canceled engagements reported a mean of 3.2 (n = 311). These ratings concern retained capability, not the occurrence count for third-party work done by hand.

Higher capability ratings were associated with a lower cancellation rate (r = −0.65, 95% CI −0.73 to −0.54, n = 150 engagements), higher active-engagement customer satisfaction (r = 0.72, 95% CI 0.60 to 0.80, n = 94 engagements), and a higher engagement indicator mean calculated without this indicator (r = 0.88, 95% CI 0.84 to 0.91, n = 150 engagements). Higher ratings were also associated with lower intent to leave (r = −0.67, cluster-bootstrap 95% CI −0.73 to −0.62, n = 757 respondents). The engagement mean was distressed on 38 of the 150 engagements.

The indicator-specific compilation contained 43 written answers about manual capability, with 22 negative and 21 positive sentiment classifications. Respondents described both retained skills used during outages and capability lost when experienced delivery team members left without documenting or passing on the work. The existing written-sentiment analysis detected no statistically significant relationship with the core measures. Retaining a capability and relying on undocumented individual knowledge remain distinct aspects of the reported experience.

### 5.4 Knowledge of Informal Third-Party Corrections

Knowledge of where informal third-party corrections had occurred received 262 distressed ratings among 792 answers (33.08%). The mean rating was 4.61 and the median was 5.0. Respondents on canceled engagements reported a mean of 3.1 (n = 305). This indicator concerns knowledge of earlier manual practices rather than the delivery team’s present ability to perform them.

Higher knowledge ratings were associated with a lower cancellation rate (r = −0.66, 95% CI −0.74 to −0.56, n = 150 engagements), higher active-engagement customer satisfaction (r = 0.61, 95% CI 0.47 to 0.73, n = 94 engagements), and a higher engagement indicator mean calculated without this indicator (r = 0.87, 95% CI 0.82 to 0.90, n = 150 engagements). They were also associated with lower intent to leave (r = −0.66, cluster-bootstrap 95% CI −0.71 to −0.60, n = 746 respondents). The engagement mean was distressed on 39 of the 150 engagements.

Of the 21 indicator-specific written answers, 20 had negative sentiment. The accounts described undocumented corrections, departures of people who knew them, and repeated rediscovery of the same correction when a third-party problem recurred. The existing written-sentiment analysis detected no statistically significant relationship with the core measures. The accounts illustrate the knowledge gap without measuring how often it occurred across all engagements.

### 5.5 Replacement of Informal Corrections With Explicit Controls

Replacement of informal third-party corrections with explicit checks received 267 distressed ratings among 796 answers (33.54%). The mean rating was 4.58 and the median was 5.0. Respondents on canceled engagements reported a mean of 3.1 (n = 302). The indicator distinguishes knowing about a correction from incorporating it into a control when agents assumed the work.

Higher replacement ratings were associated with a lower cancellation rate (r = −0.64, 95% CI −0.73 to −0.54, n = 150 engagements), higher active-engagement customer satisfaction (r = 0.71, 95% CI 0.59 to 0.80, n = 94 engagements), and a higher engagement indicator mean calculated without this indicator (r = 0.88, 95% CI 0.84 to 0.91, n = 150 engagements). They were also associated with lower intent to leave (r = −0.65, cluster-bootstrap 95% CI −0.71 to −0.59, n = 748 respondents). The engagement mean was distressed on 40 of the 150 engagements.

Nine indicator-specific written answers addressed control replacement: seven had negative sentiment, one positive, and one neutral. They described corrections not incorporated into explicit checks when agents assumed tasks. The small base supports illustration rather than a frequency estimate of the incidents described. The structured ratings remain respondents’ assessments of control replacement rather than an independent control test.

← Part 5 overviewManual Alteration by Cancellation Status and Engagement Indicator Mean →


## Links to related study evidence

- [Part 5](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/part-5-manual-alteration/)
- [Manual Alteration](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/part-5-manual-alteration/)
- [customer satisfaction](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-1-measure-customer-satisfaction/)
- [engagement indicator mean](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-1-measure-engagement-indicator-mean/)
- [intent to leave](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-4-c019/)
- [C199](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-5-c199/)
- [C205](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-5-c205/)
- [C211](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-5-c211/)
- [Part 3](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/part-3-governance-oversight-control/)
- [governance, oversight, and control](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/part-3-governance-oversight-control/)
- [third-party ecosystem](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/part-2-five-failure-categories/third-party-ecosystem/)


## Statistical analyses referenced in this paper

- [Regular Manual Alteration of Financial-Institution-Provided Data by the Wider Delivery Team](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-5-c203/)
- [Regular Manual Alteration of Financial Institution Processes by the Wider Delivery Team](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-5-c209/)
- [Regular Manual Alteration of AI Outputs by the Wider Delivery Team](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-5-c215/)
- [Accuracy of Dashboards and Performance Reports](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-5-c052/)
- [Reporting Alteration in Dashboards and Performance Reports](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-5-c053/)
- [Retrospective Comparison of Reporting Alteration](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-5-c054/)
- [Retention of the Ability to Perform Manual Workarounds](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-2-c190/)
- [Knowledge of Informal Third-Party Corrections](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-2-c191/)
- [Replacement of Informal Corrections With Explicit Controls](https://akholi.com/outsourcing-2026-agentic-ai-tier-1-banks/study-data/part-2-c192/)