Engagement Outcomes and Their Correlates

The engagements in the study were performing poorly on every core study measure. By the study date, 56 of the 150 engagements had been canceled (37.33%), and the delivery teams rated cancellation at least moderately likely on 58 of the 94 that remained active. Customer satisfaction had a mean of 43.85% across the 150 engagements. The engagement indicator mean, the average rating of the circumstances each engagement was delivered under, was 4.02 on the 10-point scale (n = 150 engagements). Intent to leave was high, with 430 of 798 respondents (53.88%) rating their likelihood of leaving within 12 months at 7 or above.

“We are four months into this engagement and customer satisfaction has collapsed. My team corrects client data almost every cycle before it can be used, legal research inputs arrive incomplete and inconsistent, and by the time it reaches the model it is not fit to work from.”Engagement Manager, KPO, Regulatory, audit, tax and legal

The four measures did not vary independently. An engagement delivered under poor conditions was, with few exceptions, an engagement that was canceled or at risk, rated low in satisfaction and staffed by people who intended to leave. One component accounted for 88.23% of the variance of the four measures, and every pair was correlated at 0.72 or more in absolute value across the 150 engagements. Among the 30 engagements delivered under the poorest conditions, 28 had been canceled. None of the 60 engagements delivered under the best conditions had been canceled.

The structural attributes of an engagement did not separate good outcomes from poor ones. Seven structural attributes were each unrelated to the outcomes. They were the type of financial institution, the outsourcing model, the class of work, the vendor’s type and size, the current size of the delivery team and the age of the engagement. Most of the results were equivalent to zero. What separated the engagements was the set of circumstances the delivery team worked within, recorded in the 100 rated indicators and grouped in the five failure categories (Part 2).

Agentic AI outsourcing is a new form of delivery, and the engagements in the study were young. Their mean age at the time of data collection was 5.69 months, and the oldest had run for 22 months (n = 150 engagements). Production had been reached by 30 of the 150 engagements, and the remaining 120 were in Build-Out or Transition/Cut-Over.

The vendors were new to agentic AI outsourcing as well. Every respondent who answered placed the vendor’s experience of delivering agentic AI outsourcing at two years or less, with a mean of 1.0 year (n = 571 respondents). No respondent reported more than one year of personal experience with agentic AI (n = 793 respondents). Engagement managers who had run earlier outsourcing programs for the same employer described what that inexperience meant on the engagement.

“The last program I ran for this employer was a contact center transition, and we were good at that, because we had done forty of them and there was a playbook that actually worked. Here there was no playbook, and the people who wrote the proposal had not run one of these either…”Engagement Manager, CXM, Client servicing and communications

“…we were sold into this as if the firm had done it before. It had not. There was no playbook for taking a screening agent from pilot into governed production, so we wrote the runbook ourselves at night while the alert backlog grew.”Engagement Manager, ITO, Financial crime and client onboarding

The markers of maturity did not separate good outcomes from poor ones. The age of the engagement, the time it had spent in its current stage and the number of agents it had in production were each unrelated to the four core study measures. All 12 associations were equivalent to zero. The vendor’s reported experience of agentic delivery was likewise equivalent to zero against all four measures, a result its narrow range of answers made unavoidable. The study set offers no evidence that an engagement improved with time alone.

Immaturity is the setting in which every other finding in the report arose. The vendors and their delivery teams were building the method of agentic delivery on live engagements for which the financial institution was paying, and the financial institution’s own staff were new to agents as well. The agentic AI knowledge and skills of the financial institution’s staff on the engagement were distressed for 418 of 793 respondents (52.71%).

“We are the first team in our unit to run agents on an investment platform build, so we are working out the method as we go, that is the honest position. Leadership are supportive and the people we are given are capable enough, but ownership when an agent puts a wrong assumption into the pricing module is still not settled anywhere above me.”Engagement Manager, ITO, Technology engineering and operations

“I have been in the industry a long time. The last six months have been rough. … We are creating a new reality every day. AI behaviour changes without explanation, data feeds are not consistent. Our team has to make updates to data, the process, and the AI rules every day.”Portfolio Director, BPO, Payments, cards and disputes

1.1 Lifecycle Position of the Engagements

Live operation had not been reached by 120 of the 150 engagements. Build-Out and Transition/Cut-Over held 60 engagements each, and Production held 30. The stage an engagement had reached did not mark its performance. Cancellation occurred in every stage, and none of the four study measures differed across the stages.

Cancellation was most frequent in Build-Out, at 29 of 60 engagements (48.33%), and least frequent in Transition/Cut-Over, at 17 of 60 (28.33%). In Production, 10 of 30 engagements had been canceled (33.33%), a rate between the other two stages. The difference across the stages did not reach significance (chi-square 5.39, p = 0.068, n = 150 engagements). Among the active engagements, customer satisfaction, the engagement indicator mean and intent to leave were each equivalent to zero across the stages. Reaching Production did not distinguish the engagements that survived from those that did not.

Engagements were canceled early as well as late. Respondents on some described an engagement that ended within weeks of starting.

“We were canceled at eight weeks, which means most of what this section asks about was never actually tested…”Engagement Manager, CXM, Client servicing and communications

“On the Tuesday we agreed how vendor queries should route between their team and mine, by the Friday the client had pulled the work.”Engagement Manager, CXM, Corporate and back-office functions

1.1.1 Lifecycle Stage Associated With the Study Measures Only Through Cancellation

Respondents were asked what stage their engagement was in. The most frequent answer was Engagement Canceled, given by 309 of 810 respondents (38.15%). Where both a respondent’s answer and the Engagements sheet recorded a stage, the two agreed in 501 of 501 cases.

The stage respondents reported was associated with the study measures through one answer, cancellation. The engagement indicator mean differed across the reported stages (Welch F 40.50, p < 0.001, n = 141 engagements), as did intent to leave (Welch F 71.51, p < 0.001, n = 141 engagements). Both differences rested on the canceled engagements, whose engagement indicator mean was 2.6 and intent to leave 9.3 (n = 56 engagements). Among active engagements alone, neither measure differed across the stages, and customer satisfaction was equivalent to zero.

The written responses that addressed the stage numbered 162, from 139 respondents on 69 engagements, and 139 came from canceled engagements. Neutral scores were given to 124 of the 162 responses (76.54%), and the neutral responses recorded what the respondent had valued in an engagement that had ended. Among 61 responses on what respondents loved most about the job, 57 referred to the end of the engagement, each from a canceled engagement. An individual contributor named the end of the engagement itself as what he or she loved most about the job.

“…the fact that it is finally over”Individual Contributor, ITO, Technology engineering and operations

“to be candid i felt relieved when they told us it is stopping. before this i was on a card disputes queue, small account, but the process was written down, one way to do the thing, and if you follow it the case closes. here it was never like that.”Individual Contributor, CXM, Client servicing and communications

1.1.2 Engagement Age Equivalent to Zero Against All Four Measures

The Engagements sheet gives the age of every engagement. The mean age was 5.69 months and the median 5.0 months, from 1.0 to 22.0 months (n = 150 engagements). Every respondent’s answer on the age of the engagement matched the sheet.

The age of the engagement showed no association with any of the four study measures, and each result was equivalent to zero. The correlation with cancellation was r = −0.08 (95% CI −0.24 to 0.08, p = 0.327, n = 150 engagements). With customer satisfaction among active engagements it was r = 0.09 (p = 0.374, n = 94 engagements). For the engagement indicator mean the result was r = 0.12 (p = 0.142, n = 150 engagements), and for intent to leave r = −0.08 (p = 0.360, n = 798 respondents). Age varied with the lifecycle stage, from a mean of 2.7 months among active engagements in Build-Out (n = 31 engagements) to 11.5 months among active engagements in Production (n = 20 engagements). It describes the stage an engagement had reached, not its performance.

The written responses that addressed age numbered 73, from 69 respondents on 48 engagements. In 15 of the 73, the engagement was too young for the survey’s questions to be answered as asked. In 14, all from canceled engagements, it had ended within weeks or months. Seven said a problem had persisted for six months or more without improving.

“We are two month in and there is no platform yet, so what am i rating… i send a request for a schema and it sit for eleven days, then the reply is ask someone else.”Individual Contributor, ITO, Technology engineering and operations

“Hard to answer these when engagement lasted four weeks. There was no joint model to speak of, only two kickoff calls and a governance pack nobody signed.”Engagement Manager, KPO, Corporate and back-office functions

“I feel these questions are asking about things which did not happen yet for us. We are two months in, nothing is in Production, so how do I answer for running many agents in live.”Individual Contributor, KPO, Corporate and back-office functions

“The scores look good because we are only one month in and the volume is small.”Individual Contributor, KPO, Credit, underwriting and suitability

“19 months and nobody from their side came to sit with the queue for one hour.”Individual Contributor, CXM, Client servicing and communications

1.1.3 Time in Current Stage Without Association With Any Study Measure

Respondents reported how many months the engagement had spent in its current stage. The mean was 3.1 months and the median 2.0 months (n = 794 respondents). None of the four study measures was associated with the time in stage, and every result was equivalent to zero. Against cancellation the result was r = −0.10 (p = 0.233, n = 150 engagements), and against customer satisfaction among active engagements r = 0.07 (p = 0.488, n = 94 engagements). Against the engagement indicator mean it was r = 0.15 (p = 0.069, n = 150 engagements), and against intent to leave r = −0.11 (p = 0.166, n = 745 respondents).

The time in stage varied with the lifecycle stage, as the age of the engagement did. It ranged from 1.7 months among respondents on active engagements in Build-Out (n = 167 respondents) to 6.5 months among those in Production (n = 98 respondents). Three written responses addressed it, each describing a stalled engagement or a request that went months without an adequate answer.

“We asked in March for role catalogue, we asked again in May, and what came back was a spreadsheet exported from an old attestation cycle, nine hundred groups, no description against most of them.”Engagement Manager, ITO, Technology engineering and operations

“Once the pilot stalled, the vendor kept the infrastructure-readiness dashboard green by reporting only the cloud environments that had passed setup, quietly excluding the legacy on-prem systems that never cleared readiness testing.”Individual Contributor, CXM, Client servicing and communications

1.1.4 Planned Time to Production Associated Only Through Canceled Engagements

Respondents reported how many months remained, under the current plan, until the engagement reached Production. The mean answer was 3.9 months (n = 545 respondents on 104 engagements), and 251 respondents answered “I do not know.” Respondents on canceled engagements, and on engagements already in Production, reported zero. Active engagements in Build-Out reported a mean of 7.7 months (n = 163 respondents) and those in Transition/Cut-Over 3.9 months (n = 230 respondents).

A longer planned time to Production was associated with less cancellation (r = −0.37, 95% CI −0.53 to −0.19, p < 0.001, n = 104 engagements). The association rests on the zero answers of canceled engagements and did not hold with the thin-base segments removed. Among active engagements, the plan was associated with none of the study measures, and its association with customer satisfaction was equivalent to zero (r = −0.04, p = 0.680, n = 94 engagements).

The single written response on the plan described two issue trackers, one on each side, and workflow steps for which neither party had named an owner.

“There are two issue trackers. We log in ours, they log in theirs, and when same issue is in both nobody merge them, so one problem has two statuses and two owners. I raise this in stand up in first month, nothing change.”Individual Contributor, ITO, Technology engineering and operations

1.2 Depth of Agentic Operation and Vendor Experience

Few engagements were running agents in live operation, and few vendors had delivered agentic work for long. The mean number of agents running in production was 1.2 and the median 0 (n = 787 respondents), and every respondent on an active engagement before Production reported none. The vendor’s experience of agentic AI outsourcing was placed at two years or less by every respondent who answered (n = 571 respondents).

“We are not having many live systems running yet, which is why the low marks from me… the teams here are still learning on this one.”Individual Contributor, KPO, Corporate and back-office functions

Neither the scale of live operation nor the vendor’s tenure separated good outcomes from poor ones. Both were equivalent to zero against all four study measures. Within the study set, the result is consistent with vendors and financial institutions that had not yet accumulated a working method. Vendors with two years of agentic delivery, and engagements with a dozen agents in production, did not record better outcomes than engagements just beginning.

“We sold a data platform build with agents and we did not have anyone in the firm who had actually stood one up end to end, so my engineers were learning the orchestration layer at the same time as they were meant to be delivering it.”Engagement Manager, ITO, Technology engineering and operations

“for me personally it has been good learning, i knew nothing about custody six month back and now i can read a statement properly. only thing, we are the first team doing this so every mistake is ours to find… that is tiring some weeks”Individual Contributor, BPO, Trade, settlement and reconciliation

1.2.1 Agents in Production Concentrated in the Production Stage

Respondents reported the number of AI agents running in production on the engagement. The mean was 1.2 and the median 0 (n = 787 respondents). Every respondent on an active engagement in Build-Out or Transition/Cut-Over reported none. Respondents on active engagements in Production reported a mean of 6.1 agents (n = 97 respondents), and those on canceled engagements 1.2 (n = 304 respondents).

Every association of the agent count with the four study measures was equivalent to zero. The count recorded r = −0.01 against cancellation (p = 0.922, n = 150 engagements) and r = 0.06 against customer satisfaction among active engagements (p = 0.573, n = 94 engagements). Against the engagement indicator mean it recorded r = 0.03 (p = 0.746, n = 150 engagements), and against intent to leave r = 0.01 (p = 0.913, n = 742 respondents). With a median answer of 0, the count marks whether an engagement had reached Production more than how far it had scaled.

Four written responses addressed the number of agents, each in comment on the respondent’s employer. They described counts of five to about 40 agents and treated the capacity to run more as a gap in the employer’s readiness.

“What has not moved is running agents at any real number… we are fine with twelve, I would not want to be asked for two hundred tomorrow.”Engagement Manager, KPO, Credit, underwriting and suitability

“Running many agents at once is not there yet, we have five and it is already manual”Individual Contributor, ITO, Technology engineering and operations

1.2.2 Vendor Tenure in Agentic Delivery Confined to Two Years or Less

Respondents reported how many years their employer had been delivering agentic AI outsourcing. Every answer was zero, one or two years, with a mean of 1.0 year and a median of 1.0 year (n = 571 respondents). Another 220 respondents answered “I do not know.” No respondent reported a vendor with a longer history of agentic delivery.

Vendor tenure was equivalent to zero against each of the four study measures. Its correlation was r = −0.12 with cancellation (p = 0.136, n = 150 engagements) and r = −0.09 with customer satisfaction among active engagements (p = 0.407, n = 94 engagements). With the engagement indicator mean it was r = 0.04 (p = 0.633, n = 150 engagements), and with intent to leave r = −0.02 (p = 0.726, n = 543 respondents). With every answer confined to three values, the result could not have been otherwise, and it is not evidence that experience is unimportant. The finding lies in the answers themselves. No respondent placed the vendor’s experience of agentic AI outsourcing beyond two years.

No written response addressed the vendor’s tenure directly. Responses on the employer described firms delivering agentic work for the first time.

“But nobody here has done this before, there is no playbook. Every time we ask how a agent goes to production the answer is different from each person.”Individual Contributor, ITO, Technology engineering and operations

Cancellation was the most visible failure in the study. By the study date, 56 of the 150 engagements had been canceled (37.33%). Cancellation was not concentrated in any part of the study set. Its rate did not differ by type of financial institution, outsourcing model or type of vendor.

The delivery teams expected more to follow. On 19 of the 94 active engagements, the mean rating of cancellation likelihood was 7 or above, and on a further 39 it was 5 or above and under 7. Combined with the 56 canceled engagements, these make 114 of the 150 engagements (76.00%) canceled or rated at least moderately likely to be canceled. The canceled-or-at-risk population is the most consequential single figure in the report.

Canceled engagements were rated after they had been canceled. Their respondents described the end of the engagement and the conditions that preceded it in the same answers. The associations of cancellation with the other measures therefore describe how canceled engagements were rated, not what came before the decision to cancel. The active engagements are not subject to that limit, and on them the delivery team’s estimate of cancellation likelihood was associated with every study measure.

“If I am honest we were flying blind. We asked for custodian access in week two and still didn’t have it when the plug was pulled.”Engagement Manager, BPO, Trade, settlement and reconciliation

2.1 Realized Cancellation

Realized cancellation is a fact recorded for every engagement, stating whether the engagement had been canceled by the study date. It is one of the four core study measures and the only one recorded independently of any respondent’s rating. Its association with the engagement indicator mean and with intent to leave shows how the conditions and the workforce of canceled engagements differed from those of the engagements that survived.

2.1.1 Realized Cancellation and Its Association With Indicator Conditions and Intent to Leave

The Engagements sheet records 56 of the 150 engagements as canceled (37.33%). Canceled engagements had a markedly lower engagement indicator mean (r = −0.72, 95% CI −0.79 to −0.63, p < 0.001, n = 150 engagements). Respondents on canceled engagements reported markedly higher intent to leave (r = 0.76, cluster bootstrap 95% CI 0.70 to 0.81, p < 0.001, n = 798 respondents). Both associations held in every robustness check computed.

No segment of the study set differed in its rate of cancellation after Holm correction. Cancellation was not concentrated in any type of financial institution, outsourcing model, process family, class of work or type of vendor. Customer satisfaction has no primary estimate against cancellation. The satisfaction recorded for canceled and active engagements lay in two ranges that did not overlap.

The ratings of canceled engagements were given after the cancellation, and the associations describe how those engagements were rated, not what preceded the decision. The written responses from canceled engagements describe requests that went unanswered until the engagement ended.

“I flagged forty odd vendor records with two different bank details back in March, no answer, chased it twice, still nothing when the account got pulled. There is no one on their side who is allowed to say yes to anything, so the queue just grows and we work round it.”Individual Contributor, BPO, Corporate and back-office functions

“Before that I would put a surveillance rule change in our tracker and it would just sit. … We do not have a shared risk list either. We keep ours in jira and they keep theirs in whatever they use, and when an alert set went wrong last month it took two days to even agree it was a problem.”Individual Contributor, ITO, Regulatory, audit, tax and legal

“The vendor’s monthly report aggregated error rates across all payroll cycles run that month, diluting the visibility of the specific cycle where a paycheck error triggered the cancellation, so the dashboard never flagged the real risk building underneath.”Individual Contributor, BPO, Corporate and back-office functions

2.2 Forward Cancellation Risk on Active Engagements

For the 94 engagements still running, the delivery team’s own estimate of cancellation likelihood is the only view the study holds of what lay ahead. On 19 of the 94 the mean rating was 7 or above, and on 39 of the 94 it was 5 or above and under 7. Those 58 engagements, with the 56 already canceled, form the canceled-or-at-risk majority. The estimate was associated with every study measure, and most closely with the circumstances the engagement was delivered under.

2.2.1 Delivery Team Estimates of Cancellation Likelihood

Respondents rated how likely the financial institution was to cancel the engagement. On the goodness view, on which a higher figure marks an engagement judged less likely to be canceled, the mean rating was 5.0 and the median 5.0 (n = 328 respondents). A further 269 respondents answered “I do not know.”

The estimate of cancellation likelihood was associated with all four study measures, and each association held under the 22 robustness checks. Engagements judged less likely to be canceled had a lower cancellation rate (r = −0.54, 95% CI −0.65 to −0.42, n = 150 engagements). They recorded higher customer satisfaction among active engagements (r = 0.66, 95% CI 0.52 to 0.76, n = 94 engagements). Their engagement indicator mean was also higher (r = 0.69, 95% CI 0.59 to 0.76, n = 150 engagements). Their respondents reported lower intent to leave (r = −0.64, cluster bootstrap 95% CI −0.72 to −0.55, n = 316 respondents). The estimate was more closely associated with customer satisfaction, the engagement indicator mean and intent to leave than with realized cancellation itself.

Individual contributors judged cancellation more likely than portfolio directors, at a mean of 4.7 (n = 145 respondents) against 6.3 (n = 31 respondents) on the goodness view. Nine written responses addressed the likelihood of cancellation. Seven, all from canceled engagements, described an end the respondent had expected.

“…this whole engagement already feel like it is ending badly so one more issue doesnt change much”Individual Contributor, KPO, Regulatory, audit, tax and legal

“i dont know if anyone above me even knows, happens almost every day, i just override it and keep moving because nobody will fix this before it gets canceled”Individual Contributor, CXM, Client servicing and communications

Customer satisfaction separated the canceled engagements from the active ones completely. The 56 canceled engagements recorded satisfaction of 20.00% or less, with a mean of 9.88% (n = 56 engagements). The 94 active engagements recorded 40.00% or more, with a mean of 64.10% (n = 94 engagements). No engagement lay between the two groups. The delivery team’s own view of the financial institution’s satisfaction was associated with the recorded figure, and more closely still with the engagement indicator mean.

The engagements’ delivery performance was weak on every count the survey asked about. Respondents rated delivery against dates at a mean of 4.8 (n = 798 respondents) and delivery within budget at 4.9 (n = 554 respondents). Progress toward the financial institution’s key performance indicators received a mean of 4.8 on the goodness view (n = 782 respondents). Each rating was associated with all four study measures, and each was associated most closely with the engagement indicator mean.

“Our team is full of freshers. We need senior workers with some agentic AI experience. We’ve had open reqs for months, and can’t find anyone. I will have to tell the customer we will miss targets if we can’t find them by the end of the month.”Portfolio Director, BPO, Financial crime and client onboarding

The written responses describe how the reported delivery results were achieved. Dates were met by working around approvals that arrived after the deadline, and budgets absorbed effort that the plan had never priced. Missed targets appeared in the reports to the financial institution under other headings. The delivery results reported to the financial institution rested in part on manual work and on reporting choices made by the vendor (Part 5).

“When the agent mis-cleared a set of cash breaks in March, our own side took two weeks to decide who owned the fix. The platform is good, the answer on accountability was slower than the client’s patience.”Engagement Manager, BPO, Trade, settlement and reconciliation

“A composite construction decision sat waiting on client sign-off for eleven days — the change board only meet monthly, as such we made the call ourselves and moved it through.”Engagement Manager, KPO, Trade, settlement and reconciliation

“Overrides on NAV breaks sit for days because sign off belong to nobody, not the client, not us, and my team clear them anyway so the pack go out on time.”Engagement Manager, BPO, Trade, settlement and reconciliation

Measured return could be assessed only for the 30 engagements in Production. A return was reported by 17 of the 30, with a mean of 24.71% (n = 17 engagements), and every engagement that reported a return was active. Within those 17, a higher return was associated with higher customer satisfaction and lower intent to leave. A base of 17 engagements is thin, and the study’s evidence on return is correspondingly limited.

3.1 Recorded and Perceived Customer Satisfaction

The study holds two views of the financial institution’s satisfaction. The first is the figure recorded against each engagement on a 100-point scale. The second is the delivery team’s rating of whether the financial institution was happy with the engagement’s performance. The recorded figure divided the study set into two groups with an empty band between them, shown in Figure 1.3.

Among active engagements, the delivery team’s rating of the financial institution’s satisfaction was associated with the recorded figure. It was associated more closely still with the engagement indicator mean, the summary of the circumstances the delivery team worked within.

3.1.1 Non-Overlapping Satisfaction of Canceled and Active Engagements

Customer satisfaction recorded against the 150 engagements had a mean of 43.85% and a median of 53.50% (n = 150 engagements). The 94 active engagements recorded a mean of 64.10%, from 40.00% to 89.00% (n = 94 engagements). The 56 canceled engagements recorded a mean of 9.88%, from 1.00% to 20.00% (n = 56 engagements). Every canceled engagement recorded lower satisfaction than every active engagement.

Satisfaction was associated with every other study measure, and each association held in each of the 16 robustness checks. Its association with cancellation (r = −0.94, 95% CI −0.96 to −0.92, n = 150 engagements) measures the separation between the two groups. Higher satisfaction was associated with a higher engagement indicator mean (r = 0.85, 95% CI 0.80 to 0.89, n = 150 engagements). It was also associated with lower intent to leave (r = −0.84, cluster bootstrap 95% CI −0.87 to −0.81, n = 798 respondents).

The associations with the engagement indicator mean and with intent to leave did not depend on the canceled engagements. Among the 94 active engagements alone, they held at r = 0.84 and r = −0.63 respectively. Among the engagements still running, those delivered under better circumstances recorded higher satisfaction, and their delivery teams were less inclined to leave.

3.1.2 Delivery Team Perception of Financial Institution Satisfaction

Respondents rated the statement that the financial institution was happy with the performance of the engagement at its current stage. The mean rating was 5.0 and the median 5.0 (n = 379 respondents), and 244 respondents answered “I do not know.” Individual contributors gave lower ratings than portfolio directors, and canceled engagements rated lowest.

The perceived satisfaction of the financial institution was associated with all four study measures, and every association held under the 22 robustness checks. Engagements whose respondents judged the financial institution happier had a lower cancellation rate (r = −0.57, 95% CI −0.67 to −0.45, n = 150 engagements). Their recorded customer satisfaction among active engagements was higher (r = 0.53, 95% CI 0.36 to 0.66, n = 94 engagements). The rating was most closely associated with the engagement indicator mean (r = 0.74, 95% CI 0.65 to 0.80, n = 150 engagements). The delivery team’s view of the financial institution’s satisfaction was associated more closely with the engagement indicator mean than with the recorded figure.

The written responses that addressed the financial institution’s satisfaction numbered 13, from 13 respondents on 12 engagements, and 9 of the 13 were negative. They placed the dissatisfaction in data the financial institution did not own, expectations it did not state and accountability the vendor was slow to assign.

“Their data is the worst I have seen in nine years, one prospect sits four times in the CRM with four spellings and no client id, so the agent makes four leads and the relationship manager sees the same family twice in a week and stops trusting the list. … There is no name against AI anywhere in that organisation, that is the real gap.”Engagement Manager, KPO, Client servicing and communications

3.2 Delivery Against Dates, Budget, and Key Performance Indicators

Four indicators recorded the delivery team’s view of the engagement’s delivery against its dates, its budget and its stated key performance indicators, and the team’s own knowledge of the key performance indicators. All four were rated near the middle of the scale, with means between 4.8 and 4.9 on the goodness view (n = 554 to 798 respondents). Each was associated with all four study measures, held under every robustness check, and was associated most closely with the engagement indicator mean.

The written responses on delivery describe three methods. Respondents met dates by deciding without approvals that arrived late, absorbed rework and checking within a fee that had not priced them, and reported missed targets to the financial institution under other headings. The reporting choices are examined in full with manual alteration (Part 5).

“sign off arrive after the filing deadline it was even meant for, the Review Board sit too rarely to help, so i decide myself and kindly hope it hold up later”Individual Contributor, KPO, Regulatory, audit, tax and legal

3.2.1 Delivery Dates Met Through Unsanctioned Workarounds

Respondents rated the statement that the engagement was meeting its delivery dates. The mean rating was 4.8 and the median 5.0 (n = 798 respondents). Canceled engagements rated lowest, at 3.4 (n = 304 respondents), and individual contributors rated lower than portfolio directors.

Ratings of delivery against dates were associated with all four study measures, each association holding under the 22 robustness checks. Engagements rated as meeting their dates had a lower cancellation rate (r = −0.60, 95% CI −0.70 to −0.49, n = 150 engagements). They also recorded higher customer satisfaction among active engagements (r = 0.59, 95% CI 0.44 to 0.71, n = 94 engagements). The closest association was with the engagement indicator mean (r = 0.78, 95% CI 0.71 to 0.84, n = 150 engagements). Respondents who rated delivery higher reported lower intent to leave (r = −0.61, n = 751 respondents).

The written responses that addressed delivery dates numbered 57, from 54 respondents on 42 engagements, and 42 of the 57 were negative. In 14 responses, approvals arrived after the deadline and the respondent decided without them or let the date slip.

“I just made the call myself, the change approval group meets barely at all and the decision always lands after the deadline its needed for, so I tweak the account data to match what the workflow expects and push it through myself before anyone signs off.”Individual Contributor, CXM, Client servicing and communications

“In June we had a release window on a Friday evening, and the change record needed an approver from their platform team, who had gone on leave without delegating, so the deploy sat until Monday even though everything else was ready.”Engagement Manager, ITO, Technology engineering and operations

Where dates were met, the written responses described how. In another 14 responses, respondents kept the date by working around a process or an approval. Seven described reporting to the financial institution that hid the slippage.

“…change process is too slow for how fast we work, so I catch it myself and keep report on time”Individual Contributor, CXM, Regulatory, audit, tax and legal

“The delivery methodology gaps got hidden by scoping every missed milestone as a change request, which reset the baseline schedule each time and kept the reported on-time delivery rate looking healthy.”Individual Contributor, ITO, Technology engineering and operations

“And do note it happen most weeks, i tell my lead about the same issue every time, nothing change so i just fix it and the payroll goes out on time”Individual Contributor, BPO, Corporate and back-office functions

“The vendor reported fleet-scaling milestones as met by counting agents deployed to a test environment rather than production, so the status report showed steady progress for months while zero tickets were actually being routed live.”Individual Contributor, CXM, Technology engineering and operations

3.2.2 Budget Pressure From Effort the Plan Did Not Price

Respondents rated the statement that the engagement was within budget. The mean rating was 4.9 and the median 5.0 (n = 554 respondents), and 242 respondents answered “I do not know.” Delivery within budget was associated with all four study measures, and none of the 22 robustness checks overturned an association. Engagements rated within budget had a lower cancellation rate (r = −0.54, 95% CI −0.64 to −0.42, n = 150 engagements). Their customer satisfaction among active engagements was higher (r = 0.62, 95% CI 0.48 to 0.73, n = 94 engagements). The closest association was with the engagement indicator mean (r = 0.76, 95% CI 0.69 to 0.82, n = 150 engagements).

Seven written responses addressed the budget, each negative. Six described effort or cost that the engagement absorbed outside its plan or fee, in the form of rework, staff added under a fixed fee, repeated runs of the AI tools and manual checks.

“In June we were asked to add two more allocation reviewers and were told the fee is fixed on the numbers agreed last year, which nobody has revisited since agents took half the run. So we absorbed it.”Engagement Manager, BPO, Trade, settlement and reconciliation

“on my last project the bot was a script, you knew the cost. this one burns through budget re-runing the same patch check four times, nobody warned us about that part”Individual Contributor, ITO, Client servicing and communications

“We check output every run now, it move on its own sometimes, feels like extra work nobody planned for but we do it anyway.”Individual Contributor, BPO, Corporate and back-office functions

“Scope-creep tickets tied to mandate changes were logged under a generic ‘change request’ bucket, which kept the reported schedule variance small while the real rework hours piled up underneath it.”Individual Contributor, ITO, Payments, cards and disputes

3.2.3 Missed Key Performance Indicators Presented Under Other Headings

Respondents rated the statement that the engagement was not achieving its stated key performance indicators. On the goodness view, on which a higher figure marks an engagement closer to its targets, the mean rating was 4.8 (n = 782 respondents). Distressed ratings came from 241 of the 782 answers (30.82%), and the engagement mean was distressed on 32 of the 150 engagements. Individual contributors rated the engagement furthest from its targets, at 4.7 (n = 608 respondents), and portfolio directors closest, at 5.8 (n = 24 respondents).

Progress toward the key performance indicators was associated with all four study measures, and the associations held under every one of the 22 robustness checks. Engagements rated further from their targets had a higher cancellation rate (r = −0.59, 95% CI −0.69 to −0.48, n = 150 engagements). Engagements rated closer to their targets recorded higher customer satisfaction among active engagements (r = 0.62, 95% CI 0.48 to 0.73, n = 94 engagements). They also had a higher engagement indicator mean (r = 0.79, 95% CI 0.72 to 0.84, n = 150 engagements).

The written responses that addressed the indicator numbered 14, from 13 respondents on 12 engagements, and 12 of the 14 were negative. Four, written in answer to the question on how the vendor altered its reports, described missed targets presented to the financial institution under other headings. Other responses described service levels that declined once agents took on the tasks.

“Missed KPI targets get reclassified as data-quality exceptions rather than performance misses, which keeps the official miss rate lower than what the team actually experiences month to month.”Individual Contributor, CXM, Client servicing and communications

“SLAs collapsed once agents touched the work”Engagement Manager, ITO, Client servicing and communications

“KPIs are missed almost every cycle and nobody above me seems alarmed enough”Engagement Manager, CXM, Corporate and back-office functions

“SLA and KPI targets collapsed the moment agent output stopped matching human pace”Individual Contributor, BPO, Trade, settlement and reconciliation

3.2.4 Delivery Team Knowledge of Financial Institution Key Performance Indicators

Respondents rated the statement that they knew the key performance indicators the financial institution used to measure success. Its mean rating was 4.8 and its median 5.0 (n = 787 respondents). Canceled engagements rated lowest, and individual contributors rated lower than portfolio directors.

Knowledge of the targets was associated with all four study measures, with no association overturned in the 22 robustness checks. Engagements whose respondents knew the targets better had a lower cancellation rate (r = −0.64, 95% CI −0.72 to −0.53, n = 150 engagements). They recorded higher customer satisfaction among active engagements (r = 0.65, 95% CI 0.51 to 0.75, n = 94 engagements). Its association with the engagement indicator mean (r = 0.81, 95% CI 0.75 to 0.86, n = 150 engagements) was the closest of the four delivery indicators. Delivery teams with greater knowledge of what the financial institution was measuring were on engagements delivered under better circumstances.

The single written response on the indicator described expectations the financial institution had not stated to the delivery team.

“We asked three times for a methodology note and got a call instead, forty minutes, two of their people disagreeing with each other while we listened. So we built to what we heard, and then the review came back saying the outputs did not look like what they expected – which is fair, except nobody had told us what they expected.”Engagement Manager, KPO, Corporate and back-office functions

3.3 Measured Return in Production

Measured return is recorded only for the 30 engagements that had reached Production, the only engagements far enough along to produce one. It is never computed across the whole study set, where 120 engagements could not yet have generated a return and a figure across all 150 would not be interpretable. The evidence it offers rests on 17 engagements and is reported with that limit stated.

3.3.1 Measured Return Confined to Active Production Engagements

A return is recorded for each of the 30 engagements in Production, and 10 of the 30 had been canceled by the study date. A return was reported by 17 engagements, with a mean of 24.71% and a range from 6.00% to 47.00%. The other 13, the 10 canceled engagements and three active ones, had not generated a return. Every engagement that reported a return was active.

Within the 17 engagements that reported a return, a higher return was associated with higher customer satisfaction (r = 0.71, 95% CI 0.35 to 0.89, p = 0.001, n = 17 engagements). A higher return was also associated with lower intent to leave (r = −0.71, 95% CI −0.89 to −0.34, p = 0.002, n = 17 engagements). The association with the engagement indicator mean was positive and inconclusive (r = 0.47, 95% CI −0.01 to 0.78, p = 0.056, n = 17 engagements), a test with power of 0.48 at this base. Every result rests on 17 engagements, and the intervals are wide.

Return, like customer satisfaction, is the figure the vendor’s team reported for the engagement, and the study holds no record from the financial institution against which to confirm it. Two engagement managers, writing on the vendor’s reporting, described return calculations that overstated the return. Individual contributors described efficiency figures computed without the hours staff spent checking or completing the agents’ work.

“The vendor’s ROI tracker recalculated the baseline savings assumption each quarter to match whatever the actual result was, so the reported ROI never fell short even as the underlying business case kept eroding.”Engagement Manager, KPO, Corporate and back-office functions

“The vendor’s monthly value report counted every hour the agent spent on forecasting as fully productive, with no line for the hours controllers then spent reconciling those forecasts against the ledger, which made the ROI case look real when it was not.”Engagement Manager, KPO, Corporate and back-office functions

“The vendor’s efficiency report calculated hours saved against a fully manual baseline without netting out the additional hours analysts spent verifying agent-flagged outputs, inflating the apparent productivity gain shown to the desk.”Individual Contributor, KPO, Corporate and back-office functions

“The vendor reported time saved based on tasks routed to the agent, without netting out the extra time staff spent manually completing the same tasks after the agent hit an API wall, which overstated the actual efficiency gain.”Individual Contributor, CXM, Corporate and back-office functions

The engagement indicator mean is the delivery team’s average rating of the circumstances an engagement was delivered under. It spans the five failure categories of financial institution readiness, vendor capability, co-management, AI technology and the third-party ecosystem. Across the 150 engagements its mean was 4.02 on the 10-point scale, from 1.2 on the engagement rated worst to 7.3 on the engagement rated best. It was associated with each of the other three core measures, and the associations held among the active engagements alone.

The four core measures form one connected outcome rather than four separate results. Every pair was correlated at 0.72 or more in absolute value, and a single component accounted for 88.23% of their variance across the 150 engagements. An engagement delivered under good conditions was, in the study set, an engagement that survived, recorded higher customer satisfaction and had a delivery team less inclined to leave. An engagement delivered under poor conditions was, with few exceptions, the reverse.

The gradient across the indicator quintiles shows the relationship in five steps. Among the 30 engagements in the lowest band of conditions, 28 had been canceled (93.33%). The cancellation rate was lower in each band that followed, at 19 of 30 and then 9 of 30, and none of the 60 engagements in the two highest bands had been canceled. Customer satisfaction was higher and intent to leave lower in each band than in the band before it, with no reversal.

No single indicator was associated with outcomes beyond the engagement indicator mean. Each of the 100 rated indicators was tested against cancellation, customer satisfaction and intent to leave with the mean removed, and none of the 300 tests held after correction. The conditions recorded by the indicators matter together. The reader of Parts 2 to 5 should take each indicator as one view of the conditions an engagement was delivered under, not as a separate influence on its outcome.

4.1 The Engagement Indicator Mean as a Summary Measure

The engagement indicator mean draws on the 100 rated indicators, 20 in each failure category, and on every respondent who rated any of the 100. It is the study’s summary of the conditions each engagement was delivered under. Its distribution, and its association with the other three core measures, establish the frame within which every indicator in Parts 2 to 5 is reported.

4.1.1 Distribution of the Engagement Indicator Mean and Its Association With Outcomes

Across the 150 engagements, the engagement indicator mean was 4.02 and the median 3.8, from 1.2 to 7.3. The rating behind it was given by all 850 respondents, each of whom rated at least one of the 100 indicators. Canceled engagements recorded a mean of 2.6 (n = 56 engagements), against 4.8 to 4.9 in the three active stages.

A higher engagement indicator mean was associated with less cancellation (r = −0.72, 95% CI −0.79 to −0.63, n = 150 engagements). It was associated with higher customer satisfaction among the active engagements (r = 0.84, 95% CI 0.77 to 0.89, n = 94 engagements). The mean was also associated with lower intent to leave (r = −0.78, cluster bootstrap 95% CI −0.82 to −0.74, n = 798 respondents). Every association held in each of the 15 robustness checks.

The associations with customer satisfaction and intent to leave held among active engagements alone. They do not rest on the gap between canceled and active engagements. Within the engagements still running, better conditions were associated with higher customer satisfaction and a delivery team more inclined to stay.

4.2 Shared Structure Across the Four Measures

The four core measures share one structure. They were correlated in every pair, they varied together across the five bands of engagement conditions, and their association held among the active engagements alone. Figure 1.4 shows cancellation across the indicator quintiles, the gradient that governs the interpretation of the whole report.

The structure means that a result on one measure cannot be separated from the others. Because the four measures share most of their variance, an indicator associated with one of the four in this study set was, as a rule, associated with the other three as well.

4.2.1 A Single Dominant Component Across the Four Study Measures

The four core measures were correlated with one another in every pair across the 150 engagements. The absolute correlations ranged from 0.72 to 0.94. The strongest pair was cancellation with customer satisfaction (r = −0.94, 95% CI −0.96 to −0.92, p < 0.001, n = 150 engagements). The weakest was cancellation with the engagement indicator mean (r = −0.72, 95% CI −0.79 to −0.63, p < 0.001, n = 150 engagements).

A principal component analysis of the four measures found one dominant component. It accounted for 88.23% of their variance across the 150 engagements, with an eigenvalue of 3.53, and the second component had an eigenvalue of 0.30. Every measure loaded on the first component at 0.91 or more in absolute value, with cancellation and intent to leave loading negatively. An engagement that stood high on one measure, oriented so that a higher value is the better outcome, stood high on the other three.

4.2.2 Monotonic Improvement in Every Measure Across the Indicator Quintiles

The 150 engagements were ranked by engagement indicator mean and divided into five bands of 30. No study measure was worse in any band than in the band before it.

The bands differed in cancellation (chi-square 85.32, df = 4, p < 0.001, n = 150 engagements). They differed in customer satisfaction (F = 84.65, df = 4 and 145, p < 0.001, omega squared 0.690, n = 150 engagements). The bands also differed in intent to leave (F = 76.29, df = 4 and 145, p < 0.001, omega squared 0.668, n = 150 engagements). The rise of the engagement indicator mean from band to band follows from how the bands were built and is not a finding. The behavior of the other three measures is. Engagements delivered under the poorest conditions were almost all canceled, at 28 of 30, and none of the 60 delivered under the best conditions had been canceled.

4.2.3 Persistence of the Associations Among Active Engagements

Among the 94 engagements not canceled by the study date, the three remaining core measures were correlated in every pair. Higher customer satisfaction was associated with a higher engagement indicator mean (r = 0.84, 95% CI 0.77 to 0.89, p < 0.001, n = 94 engagements). It was associated with lower intent to leave (r = −0.73, 95% CI −0.81 to −0.62, p < 0.001, n = 94 engagements). A higher engagement indicator mean was associated with lower intent to leave (r = −0.69, 95% CI −0.78 to −0.57, p < 0.001, n = 94 engagements).

The delivery teams rated cancellation at least moderately likely on 58 of the 94 active engagements, and those engagements recorded lower customer satisfaction. Across the active engagements, a higher rated likelihood of cancellation was associated with lower customer satisfaction (r = −0.66, 95% CI −0.76 to −0.52, p < 0.001, n = 94 engagements). The shared structure of the four measures does not depend on the canceled engagements. It holds among the engagements still running, where no cancellation had occurred for a respondent to react to.

4.3 Indicators Net of the Engagement Indicator Mean

The engagement indicator mean governs how every indicator in Parts 2 to 5 is reported. Once the mean is taken into account, no single indicator was associated with outcomes. The indicators matter together, and the report does not rank one indicator over another on the strength of its association with cancellation, satisfaction or intent to leave. The differences between indicators lie in their written evidence and in what each reveals about the conditions of the engagement.

4.3.1 No Indicator-Specific Association Net of the Engagement Indicator Mean

Each of the 100 rated indicators was correlated with cancellation, customer satisfaction and the engagement mean of intent to leave, with the engagement indicator mean held constant. The mean was recomputed without the indicator under test. In 33 of the 300 tests, p was under 0.05, against 15 expected by chance alone.

After the false discovery rate was controlled across all 300 tests, none held. The smallest adjusted value, q, was 0.323. The largest partial correlation in absolute value was 0.25, under the study’s bound for a negligible association.

The result sets the rule for how every indicator in the report is interpreted. An indicator’s association with outcomes is an association it shares with all the conditions the engagement indicator mean summarizes. No indicator is presented as associated with outcomes beyond the mean, and none is ranked over another on that basis. Where indicators differ, they differ in what their written evidence shows.

A reader looking for an explanation of poor outcomes would turn first to the structural attributes of an engagement. The attributes are the type of financial institution, the outsourcing model, the process outsourced, the commercial terms, the vendor and the size of the delivery team. None of the attributes separated good outcomes from poor ones. Equivalence to zero against all four core measures held for 10 of the 19 attributes, and for those the study can rule out any association large enough to matter. The null results have the same standing as any finding in the report. They rule out the explanations a reader would consider first.

Each attribute’s result against each core measure is stated in one of the four states of a reported result.

Two attributes are exceptions, each with a limit. Engagements whose respondents agreed that the financial institution was shifting more work to its own global capability centers recorded worse outcomes on every measure, and the associations held among active engagements alone. The rating records the respondents’ view of the financial institution’s plans, and the study holds no evidence on which came first. Engagements that had reached a larger peak team had a higher cancellation rate. Two of the three associations of peak team size did not hold among active engagements, and the data do not show whether teams grew in response to problems.

Large parts of the delivery team did not know the commercial and institutional facts of their own engagement. The answer “I do not know” was given by 228 of the 575 respondents who answered on the pricing model and 258 of the 597 who answered on the contract term. The same answer was given by 243 of the 597 who answered on how long the vendor had served the financial institution. All 150 engagement managers and all 31 portfolio directors said whether the financial institution operated its own capability centers, against 168 of the 667 individual contributors (25.19%). The individual contributors who performed the engagement’s tasks knew least about the terms on which it had been sold.

“The playbook that exists is the old staffed-delivery playbook with the word agent inserted. Pricing was still constructed on headcount and hours.”Portfolio Director, KPO, Corporate and back-office functions

Where the written responses addressed the structural attributes at all, they placed the difficulty elsewhere. Respondents described employers whose training did not cover the type of financial institution served, and pricing built on the human effort the agents were meant to replace. What separated good outcomes from poor ones was the set of conditions inside each engagement, recorded in the 100 rated indicators (Part 2).

“The vendor priced the expense agent’s throughput against the old FTE headcount baseline instead of actual hours saved, so the billed ‘efficiency gain’ stayed flat no matter how many expense reports the agent cleared on its own.”Individual Contributor, BPO, Corporate and back-office functions

“Pricing model still assumes FTE hours nobody is spending anymore”Individual Contributor, CXM, Technology engineering and operations

5.1 The Financial Institution

The delivery team’s account of the financial institution’s attributes covers its type, its size, its global capability centers and its plans for agentic AI outsourcing. None of these attributes separated good outcomes from poor ones, with one exception. Engagements whose respondents agreed that the financial institution was shifting more work to its own capability centers recorded worse outcomes on every measure.

Much of what the survey asked about the financial institution was unknown to the delivery team. More respondents answered “I do not know” than gave a rating on the growth in the financial institution’s agentic engagements. Its asset size was unknown to 334 of the 576 respondents who answered (57.99%). The individual contributors who performed the engagement’s tasks knew least.

5.1.1 Institution Type Without Association With the Study Measures

The Engagements sheet classifies the financial institution served by each engagement into 10 types, from 24 engagements in asset management to nine in private equity and private capital. The type of financial institution was not associated with cancellation (chi-square 5.71, df = 8, p = 0.693, n = 141 engagements). It was not associated with the engagement indicator mean (Welch F 1.03, p = 0.425, n = 141 engagements) or with intent to leave (Welch F 0.83, p = 0.583, n = 141 engagements). Each of the three results was equivalent to zero. The test of customer satisfaction, on the four types with 10 or more active engagements, did not reach significance and was inconclusive (Welch F 2.39, p = 0.097, n = 46 engagements).

Investment banking engagements recorded the lowest customer satisfaction among active engagements (mean 56.40%, n = 10 engagements). They also recorded the lowest engagement indicator mean (3.4, n = 19 engagements) and the highest intent to leave (7.7, n = 19 engagements). None of the three differences held under test.

Six written responses addressed the type of financial institution, and four were negative. Each negative response placed the difficulty in the vendor’s preparation for that type of institution, not in the type itself.

“The spend on training is real, I will grant that, but the courses are generic and nothing in them touches an asset manager’s book.”Portfolio Director, KPO, Trade, settlement and reconciliation

“On the estate they are no better or worse than most asset managers we deal with at this stage. Their people answered when we raised things, and the business case never wobbled.”Portfolio Director, CXM, Corporate and back-office functions

5.1.2 Financial Institution Asset Size Largely Unknown to the Delivery Team

Respondents were asked the approximate total asset size of the financial institution. The question was answered by 576 of the 850 respondents, and 334 of the 576 (57.99%) answered “I do not know.” Those answers came from 141 of the 150 engagements. Among the 242 respondents who gave a listed size, the most frequent answer was $250 billion to under $1 trillion (90 respondents, 37.19%).

Knowledge of the asset size was greater in each more senior role. A listed size was given by 146 of the 667 individual contributors (21.89%), 64 of the 150 engagement managers (42.67%) and 30 of the 31 portfolio directors (96.77%). With “I do not know” included as an answer, the asset size did not separate the respondents on any of the four study measures.

5.1.3 Financial Institution Capability Centers as Reported by the Delivery Team

Respondents were asked whether the financial institution operated its own global capability centers. The question was answered by 596 of the 850 respondents, and 247 of the 596 (41.44%) answered “I do not know.” The financial institution operated its own capability centers on the account of 190 of the 349 respondents who answered yes or no (54.44%).

All 150 engagement managers and all 31 portfolio directors answered yes or no, against 168 of the 667 individual contributors (25.19%). The answer was not associated with cancellation, the engagement indicator mean or intent to leave, and each of the three results was equivalent to zero. The test of customer satisfaction did not reach significance and was inconclusive. Whether the financial institution operated its own capability centers did not separate the engagements. Whether it was shifting work to them was associated with every outcome measure.

5.1.4 Reported Scaling of Agentic Outsourcing Without Association With Outcomes

Respondents rated the statement that the financial institution was scaling its agentic AI outsourcing engagements, on a scale with neither a favorable nor an unfavorable end. The mean rating was 5.5 and the median 5.0, at the middle of the scale (n = 364 respondents), and 256 respondents answered “I do not know.”

Reported scaling was equivalent to zero against all four study measures. Its correlations were r = −0.04 with cancellation (p = 0.640, n = 150 engagements) and r = 0.02 with customer satisfaction among active engagements (p = 0.846, n = 94 engagements). With the engagement indicator mean the correlation was r = 0.06 (p = 0.476, n = 150 engagements), and with intent to leave r = 0.00 (p = 0.990, n = 352 respondents). Whether the delivery team saw the financial institution as expanding its agentic outsourcing was not associated with the outcomes of the engagement the team was delivering.

5.1.5 Reported Growth in the Number of Engagements Unknown to Most Respondents

The statement that the financial institution was increasing the number of its agentic AI outsourcing engagements was rated on a scale without a favorable or an unfavorable end. More respondents answered “I do not know” than gave a rating, 345 against 247. The mean rating was 5.3 and the median 5.0 (n = 247 respondents).

No association of reported growth with the four study measures differed from zero, and each was equivalent to zero. Against cancellation the result was r = 0.04 (p = 0.635, n = 127 engagements), and against customer satisfaction among active engagements r = −0.07 (p = 0.563, n = 77 engagements). Against the engagement indicator mean it was r = −0.08 (p = 0.392, n = 127 engagements), and against intent to leave r = 0.05 (p = 0.442, n = 235 respondents). The financial institution’s wider program of agentic outsourcing lay largely outside the delivery team’s view.

5.1.6 Shift of Work to Financial Institution Capability Centers and Uniformly Weaker Outcomes

Respondents rated the statement that the financial institution was shifting more work to its own global capability centers, on a scale with neither a favorable nor an unfavorable end. The mean rating was 6.1 and the median 6.0 (n = 341 respondents), and 246 respondents answered “I do not know.”

Stronger agreement was associated with worse outcomes on every measure, beginning with more cancellation (r = 0.54, 95% CI 0.42 to 0.65, n = 150 engagements). Agreement was associated with lower customer satisfaction among active engagements (r = −0.66, 95% CI −0.76 to −0.53, n = 94 engagements). It was also associated with a lower engagement indicator mean (r = −0.76, 95% CI −0.82 to −0.68, n = 150 engagements). Respondents who agreed more strongly reported higher intent to leave (r = 0.63, cluster bootstrap 95% CI 0.54 to 0.71, n = 323 respondents). Each association held among active engagements alone and in each of the 22 robustness checks. Agreement was strongest on canceled engagements and among individual contributors, and weakest among portfolio directors.

The finding has a clear limit. The rating records how respondents perceived the financial institution’s plans, not a measured movement of work. The study holds no evidence on which came first. Work may have moved in-house before the engagement weakened, or after it. No written response addressed the indicator.

5.2 The Relationship Between the Two Firms

Two questions described the wider relationship between the vendor and the financial institution. They asked how long the vendor had served the financial institution across all its services, and how many agentic engagements the two firms had under way at the same time. Neither the length nor the breadth of the relationship separated good outcomes from poor ones. Engagements within a long relationship built on earlier outsourcing recorded no lower cancellation rate, and a portfolio of concurrent engagements did not distinguish the engagements that held.

5.2.1 Relationship Length Without Association With Cancellation or Indicator Conditions

Respondents typed the number of months their employer had served the financial institution across all its services. The mean answer was 32.6 months and the median 33.0, from 3.0 to 67.0 months (n = 354 respondents), and 243 respondents answered “I do not know.”

The length of the relationship was not associated with cancellation (r = 0.05, p = 0.524, n = 150 engagements). The engagement indicator mean showed no association with it either (r = 0.05, p = 0.576, n = 150 engagements). Intent to leave showed none (r = −0.06, p = 0.467, n = 340 respondents), and each of the three results was equivalent to zero. The association with customer satisfaction among active engagements was positive and did not reach significance (r = 0.20, 95% CI 0.00 to 0.39, p = 0.055, n = 94 engagements). Two robustness checks did not hold, and the result is inconclusive.

Five written responses addressed the relationship. Three gave its length as 14 years, 15 years and two decades, each longer than any typed answer.

“Fourteen years of relationship equity gone in one incident”Portfolio Director, CXM, Client servicing and communications

“Fifteen years of a dependable client relationship”Engagement Manager, CXM, Corporate and back-office functions

“two decades of relationship history to draw on”Engagement Manager, KPO, Trade, settlement and reconciliation

5.2.2 Concurrent Agentic Engagements With the Same Financial Institution

Respondents typed the number of agentic AI outsourcing engagements under way at the same time between their employer and the financial institution. More respondents answered “I do not know” than gave a number, 362 against 254. Among those who gave a number, the mean was 17.2 and the median 18.0, from 2.0 to 30.0 (n = 254 respondents on 129 engagements).

The count of concurrent engagements was equivalent to zero against each of the four study measures. It recorded r = 0.10 with cancellation (p = 0.267, n = 129 engagements) and r = 0.03 with customer satisfaction among active engagements (p = 0.804, n = 83 engagements). With the engagement indicator mean it recorded r = 0.03 (p = 0.716, n = 129 engagements), and with intent to leave r = 0.04 (p = 0.691, n = 243 respondents). The scale of the two firms’ agentic work together did not distinguish the engagements that held from those that were canceled.

5.3 The Work Outsourced

The services outsourced varied widely across the study set, from contact-center queues to trade settlement, identity management, financial crime screening and the drafting of investor letters. Each engagement was classified by outsourcing model, class of work, process family and delivery location. None of the three classifications that could be tested separated good outcomes from poor ones. The type of service was associated with which processes an engagement served and which vendors delivered it, and not with how the engagement performed.

The written responses described the tasks of the engagement as the most rewarding part of the job. Asked what they loved most about the job, respondents named the use case, or the problem it addressed, and 27 of the 33 responses on the use case were positive (81.82%).

5.3.1 Outsourcing Model Equivalent to Zero Against All Four Measures

The Engagements sheet classifies each engagement by outsourcing model, with CXM on 42 engagements, ITO on 40, BPO on 37 and KPO on 31. Every respondent’s answer matched the sheet. None of the four study measures differed by outsourcing model, and each result was equivalent to zero. Cancellation recorded chi-square 0.08 (p = 0.994, n = 150 engagements), and customer satisfaction among active engagements Welch F 0.78 (p = 0.510, n = 94 engagements). The engagement indicator mean recorded Welch F 0.90 (p = 0.444, n = 150 engagements), and intent to leave Welch F 0.47 (p = 0.703, n = 150 engagements).

The four models differed in the vendors that delivered them and the processes they served, not in the outcomes the study measured. The single written response on the model came from an engagement manager on a canceled engagement, who described an engagement sold under one model and staffed under another.

“Training on funds and NAV work is decent, I will give them that. On agents the knowledge is thin. We were sold as an AI delivery and staffed as an ordinary BPO.”Engagement Manager, BPO, Trade, settlement and reconciliation

5.3.2 Class of Work Equivalent to Zero Against All Four Measures

The Engagements sheet classifies each engagement as serving a line-of-business operation (76 engagements) or a corporate function (74 engagements). The respondents’ answers on the class of work agreed with the sheet in every case. The two classes did not differ on any of the four study measures, and every result was equivalent to zero. The test of cancellation gave chi-square 0.30 (p = 0.583, n = 150 engagements), and the test of customer satisfaction among active engagements Welch F 0.05 (p = 0.828, n = 94 engagements). The engagement indicator mean gave Welch F 0.07 (p = 0.795, n = 150 engagements), and intent to leave Welch F 0.17 (p = 0.678, n = 150 engagements).

The class followed the process served. Corporate and back-office and technology engagements were mostly corporate functions, while servicing, trade and payments engagements were mostly line-of-business operations. Whether an engagement served the business directly or a function behind it was not associated with its outcomes.

5.3.3 Process Family as the Most Valued Element of the Work

Each engagement’s use case is assigned to one of eight process families, from 47 engagements in corporate and back-office functions to seven in financial crime and client onboarding. The process family was not associated with customer satisfaction among active engagements (Welch F 0.16, p = 0.850, n = 63 engagements). It was not associated with the engagement indicator mean either (Welch F 1.20, p = 0.325, n = 125 engagements), and both results were equivalent to zero. The test of cancellation (chi-square 7.48, df = 4, p = 0.113, n = 125 engagements) did not reach significance and was inconclusive. The test of intent to leave (Welch F 1.77, p = 0.155, n = 125 engagements) was inconclusive as well.

The written responses that addressed the use case numbered 33, and 27 of the 33 responses were positive (81.82%), among the most favorable written evidence in the study. Most answered the question on what the respondent loved most about the job and named the use case or the problem it addressed. Respondents on canceled engagements valued the problem they had been asked to solve as well.

“The fraud detection problem itself was worth solving”Engagement Manager, ITO, Financial crime and client onboarding

“The reconciliation problem itself is genuinely interesting work”Individual Contributor, BPO, Trade, settlement and reconciliation

“Solving identity sprawl at this scale is real work worth doing”Individual Contributor, ITO, Technology engineering and operations

“Genuinely hard problem, keeps me sharp”Individual Contributor, ITO, Technology engineering and operations

5.3.4 Delivery Location

Delivery was located in India on 92 of the 150 engagements and in the United States on 19, with the remaining 39 delivered from other locations. Delivery location has no companion analysis, and the report does not test it against the four study measures. It is reported to describe the composition of the study set.

5.4 Commercial Terms

The commercial terms of an engagement, its pricing model, contract term and annual value, did not separate good outcomes from poor ones. Large parts of the delivery team could not state those terms, and the share who could was greater in each more senior role. On one point the written responses were unanimous. All 33 responses on pricing were negative, and 21 of the 33 stated that the engagement was still priced on the human effort the agents were meant to replace.

5.4.1 Pricing on Human Effort for Work Performed by Agents

The question on the pricing model of the engagement was answered by 575 of the 850 respondents, of whom 228 answered “I do not know.” Among the 347 who named a model, the most frequent answer was FTE-based pricing, priced on full-time equivalent staff, given by 140 (40.35%). None of the study measures differed by pricing model. Cancellation recorded chi-square 2.97 (simulated p = 0.413, n = 71 engagements), and the engagement indicator mean Welch F 1.31 (p = 0.290, n = 71 engagements). Intent to leave recorded Welch F 0.79 (p = 0.511, n = 71 engagements). Only the result for intent to leave was equivalent to zero, and customer satisfaction could not be tested.

The written responses gave a different and unanimous account. All 33 written responses that addressed pricing were negative (100.00%). In 21 of the 33, the respondent stated that the engagement was still priced or billed on headcount, seats or hours while agents performed its tasks. At least three responses described the commercial consequence for the vendor. Under pricing on headcount, volume absorbed by agents left the vendor’s revenue flat, and respondents saw no incentive in the vendor’s own chain to let the agents succeed.

“As long as this is billed on heads, every hour an agent saves is an hour I am asked to explain, and on that basis there is no incentive anywhere in my chain to let the thing work properly.”Engagement Manager, BPO, Trade, settlement and reconciliation

“We are still sold and staffed as a headcount deal, forty seats on a monitoring queue, while the deck that won the work promises the agents do the first pass. So when the agents do less than that, the only lever I have is more analysts, which I cannot get, and when they do more, we are billing for seats nobody needs.”Engagement Manager, BPO, Financial crime and client onboarding

“The vendor billed the engagement on a per-FTE-equivalent basis calculated from pre-agent staffing levels, so the invoice looked justified on paper even though the agent had already absorbed most of the manual expense-processing work months earlier.”Individual Contributor, BPO, Corporate and back-office functions

“Merchant onboarding volume went up nearly double after client launched new acquiring product, and our agents absorbed it without adding headcount. Good news, except that our invoice to client is still built on seats. So we did more work and billed less, and i had to explain to my own director why revenue on account went flat in a month where we handled forty percent more cases.”Engagement Manager, CXM, Payments, cards and disputes

“We must stop pricing this as if it were bodies on a floor.”Portfolio Director, ITO, Technology engineering and operations

5.4.2 Contract Term Without Association With the Study Measures

Respondents typed the contract term of the engagement in months. The mean answer was 30.9 months and the median 36.0, from 12.0 to 60.0 months (n = 339 respondents on 150 engagements), and 258 respondents answered “I do not know.” The typical term lay between two and three years.

Each association of the contract term with the four study measures was equivalent to zero. Its correlation with cancellation was r = −0.10 (p = 0.223, n = 150 engagements), and with customer satisfaction among active engagements r = 0.03 (p = 0.744, n = 94 engagements). With the engagement indicator mean it was r = 0.08 (p = 0.302, n = 150 engagements), and with intent to leave r = −0.02 (p = 0.781, n = 327 respondents). A longer contract term was not associated with less cancellation.

5.4.3 Contract Value Unknown to Most of the Delivery Team

Respondents were asked the annual contract value of the engagement. The question was answered by 601 of the 850 respondents, and 353 of the 601 (58.74%) answered “I do not know.” Among the 248 who gave a listed value, 106 (42.74%) answered $1 million to under $5 million and 102 (41.13%) $5 million to under $10 million. The share who gave a value was greater in each more senior role, from 152 of the 667 individual contributors (22.79%) to 30 of the 31 portfolio directors (96.77%).

At engagement level, the contract value was not associated with any study measure, on a base of 31 engagements and with most results inconclusive. At respondent level, with “I do not know” counted as an answer, the answers differed in cancellation (F = 3.74, p = 0.006, n = 601 respondents). They also differed in the engagement indicator mean (F = 3.06, p = 0.018, n = 601 respondents). Respondents who gave the largest value range were the most likely to be on canceled engagements, but the groups at the two ends of the range held fewer than 30 respondents each. The first finding is the more important of the two. The value of the engagement was unknown to most of the delivery team who answered.

5.5 The Vendor

Three survey questions described the firm that delivered each engagement. They asked for the type of company, the number of its employees worldwide and the number of agentic AI engagements it runs for financial institutions. None of the three separated the engagements that held from those that were canceled. Each was tested against cancellation, customer satisfaction among active engagements, the engagement indicator mean and intent to leave, and all 12 results were equivalent to zero. Outcomes did not differ between an IT services major and a pure-play agentic AI firm. They were not associated with the size of the vendor or with the breadth of its agentic work for financial institutions.

The two typed questions also showed how little of its employer’s profile the delivery team could state. On the size of the employer, 260 of the 798 respondents who answered (32.58%) answered “I do not know.” On the number of agentic AI engagements the employer runs for financial institutions, 328 of the 850 respondents (38.59%) gave a number. For a financial institution choosing a vendor, the type and scale of the firm did not distinguish the engagements that held in this study.

5.5.1 Vendor Type Equivalent to Zero Against All Four Measures

Respondents were asked what type of company employed them. An IT services major was named in 316 of the 793 answers (39.85%) and a BPO specialist in 198 (24.97%). A consulting firm was named in 117 of the 793 answers (14.75%), a boutique or specialist provider in 87 (10.97%) and a pure-play agentic AI firm in 75 (9.46%). The respondents on each engagement gave the same answer, and the question therefore classifies the 150 engagements as well. IT services majors delivered 59 of the 150 engagements (39.33%) and pure-play agentic AI firms 14 of the 150 (9.33%).

No study measure differed by vendor type, and each result was equivalent to zero. The cancellation rate did not differ (chi-square 2.02, p = 0.731, n = 150 engagements), and neither did customer satisfaction among active engagements (Welch F 0.76, p = 0.524, n = 85 engagements). The engagement indicator mean (Welch F 0.74, p = 0.573, n = 150 engagements) and intent to leave (Welch F 0.31, p = 0.872, n = 150 engagements) did not differ either. The vendor types differed in the outsourcing models and process families they served. They did not differ in the outcomes the study measures.

5.5.2 Vendor Size Without Association With Outcomes

Respondents typed the approximate number of employees their employer has worldwide. The mean answer was 110,954.9 employees and the median 90,854.0, from 321.0 to 340,114.0 (n = 538 respondents on 150 engagements). A further 260 respondents answered “I do not know,” and they were spread across 127 of the 150 engagements (84.67%). On each of the 127 engagements, at least one respondent did not know the size of his or her own employer.

Employer size recorded results equivalent to zero on all four study measures. Against cancellation it recorded r = −0.06 (95% CI −0.22 to 0.10, p = 0.471, n = 150 engagements). Against customer satisfaction among active engagements it recorded r = −0.07 (95% CI −0.27 to 0.13, p = 0.500, n = 94 engagements). Against the engagement indicator mean it recorded r = −0.01 (95% CI −0.17 to 0.15, p = 0.901, n = 150 engagements). Against intent to leave it recorded r = −0.01 (cluster bootstrap 95% CI −0.17 to 0.15, p = 0.901, n = 508 respondents). The typed sizes differed markedly by vendor type, from the IT services majors to the pure-play agentic AI firms, consistent with the vendor types respondents named.

5.5.3 Breadth of Vendor Agentic Work for Financial Institutions

Respondents typed the number of agentic AI engagements their employer runs for financial institutions. Fewer than half gave a number. A number was typed by 328 of the 850 respondents (38.59%), 242 (28.47%) answered “I do not know,” and 280 (32.94%) left the question blank. Among those who gave a number, the mean was 24.6 engagements and the median 25.0, from 5.0 to 40.0 (n = 328 respondents on 150 engagements).

The breadth of the vendor’s agentic work recorded no association with the four study measures, each result equivalent to zero. The correlation with cancellation was r = 0.02 (95% CI −0.14 to 0.18, p = 0.833, n = 150 engagements). With customer satisfaction among active engagements it was r = 0.09 (95% CI −0.11 to 0.29, p = 0.390, n = 94 engagements). With the engagement indicator mean it was r = 0.03 (95% CI −0.13 to 0.19, p = 0.743, n = 150 engagements). With intent to leave it was r = −0.05 (cluster bootstrap 95% CI −0.21 to 0.10, p = 0.501, n = 312 respondents).

The count records the breadth of a vendor’s agentic work for financial institutions. It does not record whether the vendor had delivered the same service before. An engagement manager described a vendor that had sold a service it had never delivered.

“…this is not about the engineers… they were good, they were simply sent in without a pipeline of our own to deploy from. We sold release automation to a global bank and we do not have a working reference of it anywhere.”Engagement Manager, ITO, Technology engineering and operations

5.6 Size of the Delivery Team

Three survey questions measured the vendor team at three points in time. They asked for its size today, its largest size to date and its planned size 12 months ahead. The three measures behaved differently against the outcomes. The current team was equivalent to zero against all four study measures. The peak team was larger on canceled engagements. The planned team followed the status of the engagement and, among active engagements, was associated with higher customer satisfaction.

The written responses addressed a different question from the survey. They described whether the team was large enough for its tasks, and the most common statement, in 17 of the 33 responses, was that it was not. Respondents also described members who had left the team, had been moved off it or were looking to leave. Those accounts of understaffing return in the examination of the delivery workforce (Part 4).

5.6.1 Current Team Size Equivalent to Zero Against the Study Measures

The Engagements sheet records the headcount of the vendor team on every engagement. The mean was 8.4 people and the median 8.0, from 3.0 to 16.0 (n = 150 engagements). Respondents were also asked how many people were on the vendor team today, and every answer equaled the headcount recorded for the engagement (n = 789 respondents on 150 engagements).

Against the four study measures, every association of the headcount was equivalent to zero. Its correlation with cancellation was r = 0.12 (95% CI −0.05 to 0.27, p = 0.159, n = 150 engagements). With customer satisfaction among active engagements it was r = 0.02 (95% CI −0.18 to 0.22, p = 0.851, n = 94 engagements). With the engagement indicator mean it was r = −0.10 (95% CI −0.26 to 0.06, p = 0.226, n = 150 engagements). With intent to leave it was r = 0.09 (cluster bootstrap 95% CI −0.07 to 0.24, p = 0.289, n = 798 respondents).

The headcount followed the type of engagement. After Holm correction it differed by process family (Welch F 4.28, Holm p = 0.022, n = 150 engagements) and by class of work (Welch F 12.32, Holm p = 0.003, n = 150 engagements). Corporate and back-office engagements and corporate functional engagements had the smaller teams.

The written responses addressed whether the team was large enough for its tasks. There were 33 written responses on the size of the team, from 33 respondents on 29 engagements, and 23 of the 33 (69.70%) were negative. The most common statement, made in 17 of the 33, was that the team was too small for its tasks. Six described members who had left, had been moved or were looking to leave.

“If you only read one line of this: we don’t have the people. I was given three analysts for a diligence workstream that needs six, and two of them had never seen a data room before this.”Engagement Manager, KPO, Corporate and back-office functions

“…my team is only four people and two already looking for other job.”Individual Contributor, CXM, Corporate and back-office functions

“My lead is asking for more people every week, nobody give him…”Individual Contributor, KPO, Corporate and back-office functions

“Talent shortage means we’re always short-staffed”Individual Contributor, BPO, Corporate and back-office functions

The written responses were voluntary, and they record the experience of the respondents who wrote them. The headcount records the size of each team, and the written responses record whether, in the view of the respondents who wrote, the team was sufficient for its tasks.

5.6.2 Larger Peak Teams on Canceled Engagements

Respondents typed the largest headcount the vendor team had reached on the engagement to date. The mean answer was 10.5 people and the median 10.0, from 3.0 to 18.0 (n = 787 respondents on 150 engagements), and no respondent answered “I do not know.” At engagement level, the mean peak team was 10.0 people against a mean current headcount of 8.4 (n = 150 engagements).

Engagements with a larger peak team had a higher cancellation rate (r = 0.24, 95% CI 0.09 to 0.39, p = 0.003, n = 150 engagements). They also had a lower engagement indicator mean (r = −0.25, 95% CI −0.39 to −0.09, p = 0.002, n = 150 engagements). Respondents on those engagements reported higher intent to leave (r = 0.23, cluster bootstrap 95% CI 0.09 to 0.37, p = 0.002 with standard errors clustered by engagement, n = 739 respondents). The peak team was not associated with customer satisfaction among active engagements, and that association was equivalent to zero (r = −0.10, 95% CI −0.30 to 0.10, p = 0.322, n = 94 engagements). Canceled engagements reported the largest peak teams.

Two of the three associations did not hold among active engagements alone. Within the 94 active engagements, the association with the engagement indicator mean was r = −0.15 (p = 0.136, n = 94 engagements). Among respondents on those engagements, the association with intent to leave was r = 0.09 (p = 0.358). Those two associations therefore reflect the difference between canceled and active engagements more than a gradient within active work.

The data record the largest size the team reached. They do not record the course of the team before or after that point, or whether it grew in response to problems on the engagement. One individual contributor described a team reduced in a single day, partway through a task.

“in March they moved two people off my team to another account in one day, no handover. And we were mid way through tuning the screening rules. My lead found out from the rota, not from anyone above him. After that backlog went to nine days and the same managers sat in review asking why disposition quality dropped… nobody joined those two things.”Individual Contributor, ITO, Financial crime and client onboarding

5.6.3 Planned Team Size as a Marker of Continuation

Respondents typed the number of people the current plan puts on the vendor team 12 months ahead. A number was given by 339 of the 850 respondents (39.88%). The question was answered by 589 respondents, and 250 of the 589 (42.44%) answered “I do not know.” On 131 of the 150 engagements (87.33%), at least one respondent did not know the plan for the team. Among the respondents who gave a number, the mean was 5.6 people and the median 6.0 (n = 339 respondents).

On the canceled engagements the answer was fixed. All 135 respondents on canceled engagements who gave a number typed zero. A canceled engagement has no team planned. The association between the planned team and cancellation therefore restates the status of the engagement and is not a separate finding (r = −0.86, p < 0.001, n = 150 engagements). The stronger associations across all engagements include the canceled engagements and are subject to the same limitation. They are the association with the engagement indicator mean (r = 0.73, n = 150 engagements) and with intent to leave (r = −0.78, n = 323 respondents).

Among active engagements, where the plan is a forecast, the 204 respondents who gave a number typed from 2.0 to 19.0 people. A larger planned team was associated with higher customer satisfaction (r = 0.47, 95% CI 0.30 to 0.62, p < 0.001, n = 94 engagements). It was associated with a higher engagement indicator mean (r = 0.35, 95% CI 0.16 to 0.52, p < 0.001, n = 94 engagements). Among respondents on the active engagements, it was associated with lower intent to leave (r = −0.35, 95% CI −0.47 to −0.22, p < 0.001). The association is descriptive. The data do not show whether the plan for the team followed the state of the engagement or anticipated it.

The engagements in the study were young, and 114 of the 150 engagements had been canceled or were at risk of cancellation (76.00%). The four study measures were closely associated with one another, and no attribute of an engagement’s structure separated those that held from those that were canceled. The attributes tested ran from the type of financial institution and the process family to the vendor, the commercial terms and the current size of the team. The attributes that did mark the outcome, the peak and the planned size of the team, followed the status of each engagement more than they distinguished its design. What separated the engagements lay in the circumstances under which the delivery teams worked, recorded in the five failure categories (Part 2).