The Immaturity of Agentic AI Outsourcing

Agentic AI outsourcing is a new form of delivery, and the engagements in the study were young. The mean is the arithmetic average, and n identifies the number of observations included in a result. 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 vendor 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 runbook 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 associations of engagement age, time in the current stage, and production-agent count with the four core measures met the study's equivalence criteria. Equivalence testing assesses whether a relationship lies within prespecified bounds treated as small under the study's criterion. It does not establish a zero relationship. Vendor experience also met those criteria, although its restricted range limits interpretation. These singular cross-sectional results do not demonstrate whether individual engagements improved over time.

Limited experience provides context for the findings. Written accounts described vendors developing delivery practices during paid engagements. Respondents also rated the agentic AI knowledge and skills of financial institution staff poorly: 418 of 793 ratings (52.71%) were distressed. These accounts identify perceived capability gaps without establishing their causal contribution to outcomes.

“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

Eight months in, this is better than it was. When we started, nobody had written down which of the feeds into the repository actually mattered, and we found out the hard way when one of them changed its date format over a weekend. Since then we've mapped the ones that touch classification and we put explicit checks where the old team used to just eyeball it and fix it themselves… that part I'd call solved. What hasn't moved is warning. We still find out that a provider changed something because the queue looks wrong on Monday morning, not because anyone told us. And the model retirement notices go to our platform group, not to us, so we get maybe a week. For a POC that's survivable. If this goes to production on the same footing I'd be nervous about it.”

ENGAGEMENT MANAGER · BPO, Corporate and back-office functions

1 Lifecycle Position of the Engagements

Cancellation occurred at every measured lifecycle stage. The sample included 60 engagements in Build-Out, 60 in Transition/Cut-Over, and 30 in Production. The observed rate differences must be distinguished from statistically supported differences and estimates meeting the study's equivalence criteria.

Cancellation was recorded for 29 of 60 Build-Out engagements (48.33%), 17 of 60 Transition/Cut-Over engagements (28.33%), and 10 of 30 Production engagements (33.33%). The chi-square test compares cancellation frequencies across stages. The stage difference did not reach statistical significance (chi-square 5.39, p = 0.068, n = 150). The p value is the probability, under the null model, of a result at least as extreme as the observed result. Among active engagements, the stage contrasts for customer satisfaction, engagement indicator mean, and intent to leave met the study's equivalence criteria. These results do not establish that the stage is irrelevant outside the observed sample.

Engagements were canceled early as well as late. Respondents 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 Lifecycle Stage and Outcome Comparisons

Indicator C036 · What stage is the engagement in? · Full statistical analysis ›

Cancellation was the most frequent respondent-reported lifecycle answer, although cancellation is recorded separately from the engagement's measured stage. 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.

Reported stage was associated with the engagement indicator mean (Welch F 40.50, p < 0.001, n = 141 engagements) and intent to leave (Welch F 71.51, p < 0.001, n = 141). The canceled category had a mean indicator score of 2.6 and a mean intent to leave of 9.3 (n = 56 engagements). Among active engagements, neither stage comparison reached statistical significance, and the customer-satisfaction contrast met the equivalence criterion. The full-sample result reflects the inclusion of cancellation as a response category.

The written responses that addressed the stage numbered 162, from 139 respondents on 69 engagements. 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 they 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

1.2 Engagement Age and the Four Study Measures

Indicator C035 · How long ago did this specific engagement start, in months? · Full statistical analysis ›

The engagements were generally recent. The mean age was 5.69 months and the median 5.0 months, from 1.0 to 22.0 months (n = 150 engagements). The median is the middle value in an ordered distribution. Every respondent's answer on the age of the engagement matched in the study.

The estimated associations of engagement age with all four measures met the equivalence criterion. Correlation, denoted r, expresses the direction and strength of a linear association. A confidence interval (CI) expresses uncertainty around an estimate under the stated method. The correlation with cancellation was −0.08 (95% CI −0.24 to 0.08, p = 0.327, n = 150 engagements). The correlation with customer satisfaction among active engagements was 0.09 (p = 0.374, n = 94 engagements). It was 0.12 with the engagement indicator mean (p = 0.142, n = 150 engagements). At respondent level, the correlation with intent to leave was −0.08 (p = 0.360, n = 798 respondents). Mean age ranged from 2.7 months in active Build-Out engagements (n = 31) to 11.5 months in active Production engagements (n = 20). Age varied with stage without a detected outcome association under these tests.

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

1.3 Time in the Current Stage and Outcome Associations

Indicator C037 · How many months has the engagement been in its current stage? · Full statistical analysis ›

Reported time in the current stage averaged 3.1 months, with a median of 2.0 (n = 794 respondents). Its estimated associations met the equivalence criterion for cancellation (r = −0.10, p = 0.233, n = 150 engagements), active-engagement customer satisfaction (r = 0.07, p = 0.488, n = 94), engagement indicator mean (r = 0.15, p = 0.069, n = 150), and 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

1.4 Planned Time to Production

Indicator C038 · Per the current plan, how many months from today until the engagement reaches Production/Delivery? · Full statistical analysis ›

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 reported 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, none of the tested associations reached statistical significance. The customer-satisfaction estimate was within the study’s equivalence bounds (r = −0.04, p = 0.680, n = 94 engagements).

“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

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 was 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

Production-agent count and reported vendor experience had estimated associations within the study's equivalence bounds for all four measures. These findings do not establish the adequacy of delivery methods or the effect of acquiring more experience. Both questions describe a sample with limited exposure to sustained agentic operation.

“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

2.1 Agents in Production

Indicator C043 · How many AI agents are running in production on this engagement today? · Full statistical analysis ›

Respondents on active Production engagements reported more agents operating in production than respondents on canceled engagements. Respondents on active engagements in Production reported a mean of 6.1 agents (n = 97 respondents) and a median of 0 (n = 787 respondents). Canceled engagements reported a mean of 1.2 (n = 304 respondents).

The estimated associations of production-agent count met the equivalence criterion for cancellation (r = −0.01, p = 0.922, n = 150 engagements), active-engagement customer satisfaction (r = 0.06, p = 0.573, n = 94), engagement indicator mean (r = 0.03, p = 0.746, n = 150), and intent to leave (r = 0.01, p = 0.913, n = 742 respondents). The median count of zero limits conclusions about engagements operating larger numbers of agents.

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

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

Indicator C078 · How many years has your employer been delivering agentic AI outsourcing? · Full statistical analysis ›

No respondent reported that the vendor had delivered agentic AI outsourcing for more than two years. 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.”

Estimated associations of vendor experience met the equivalence criterion for all four core measures. The correlation with cancellation was r = −0.12 (p = 0.136, n = 150 engagements). It was r = −0.09 with customer satisfaction among active engagements (p = 0.407, n = 94 engagements) and r = 0.04 with the engagement indicator mean (p = 0.633, n = 150 engagements). At respondent level, the correlation with intent to leave was r = −0.02 (p = 0.726, n = 543 respondents). The restricted range of zero to two years limits inference about longer experience. It does not force a null result or establish that experience is unimportant.

Written responses 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

“Playbooks are better than three months back, hence i escalate less. But nobody senior here really knows the agent side, we learn it on shift.” -Individual Contributor, ITO, Technology engineering and operations

“…nobody in my batch knew anything about agents when we joined, and the company did put us on a course. Two weeks. It was not enough for AML work but it was something, and the trainer did answer questions after also. The problem is nobody above the trainer has done this before either.” -Individual Contributor, ITO, Financial crime and client onboarding