Agentic AI Outsourcing at Tier 1 Financial Institutions

This paper draws on 150 agentic AI outsourcing engagements serving Tier 1 financial institutions. Data were collected between January and June 2026 from 850 members of the vendors’ delivery teams. Those respondents provided 150,889 structured data points and 7,456 free-form comments.

All data for this study came directly from the delivery teams. The study did not approach any vendor or financial institution for data, validation, or commentary. The study collected no information that could identify a financial institution, bank customers, or the vendor.

The engagements were selected with partial intent to cover a range of financial institution types, outsourcing models, and use cases. The weighting of any financial institution type, outsourcing model, or use case in this study does not reflect real-world frequency.

The study used a 10-point scale for structured ratings. A rating of 3 or below is called distressed in this paper.

Phil Hatch · info@akholi.com · DOI: 10.6084/m9.figshare.34036890

Agentic AI outsourcing for Tier 1 financial institutions is performing poorly. At the time of the study, 56 of 150 engagements (37%) had been canceled. Of the remaining engagements, 58 (39%) were categorized as having a moderate-to-high risk of cancellation. Respondents estimated a mean customer satisfaction of 44% across 150 engagements.

Only 30 out of the 150 engagements had reached the production stage at the time of data collection. 17 of the 30 engagements (57%) reported a measured return, with an estimated mean of 25%. Only 11% of 782 respondents estimated their engagement was meeting its customer-defined KPIs throughout any phase of the outsourcing lifecycle.

Key findings from this study include:

1. All engagements in this study (150 out of 150) have numerous instances of the delivery team manually altering bank-provided data, processes, AI outputs, customer-facing reports, and 3rd party data feeds. The study included questions asking respondents whether they disclosed manual alterations to the bank or if the bank was involved in key alteration decision-making. Only 13 out of 1,004 data points (1.3%) indicated the financial institution was involved in decision-making or informed of any manual alteration.
2. AI control, oversight, and governance functions are struggling. Respondents regularly noted a material gap between operational realities and the initially designed function, driven primarily by slow response rates by both parties to the speed and volume of agentic AI output.
3. Agentic AI outsourcing is in its infancy. All findings in this study are underpinned by the fact that all parties involved lack experience with agentic AI.
4. Financial institutions have yet to fully understand the potential of agentic AI and have not optimized direct or secondary processes touching the outsourcing engagements in this study. In many cases, AI directly replaced a human step-by-step, with no gain in volume, speed, or quality.

Every engagement in the study recorded numerous instances of the delivery team making material alterations to financial institution data, processes, AI outputs, customer-facing reporting, and 3rd party ecosystem data feeds. Respondents noted that they involved the bank in key alteration decision-making or discussed the change with the financial institution in 13 of 1,004 collected data points (1.3%).

Most of the quotes respondents provided did not indicate malicious intent. Instead, respondents’ quotes indicated they genuinely had the best intentions, acted under duress, and did not know the full legal or regulatory obligation that may have been violated.

Quotes provided indicated that catching errors in data or processes before an agent failed had become a significant source of professional pride.

Altering Financial Institution-Provided Data

Delivery staff reported manually altering data supplied by the financial institution in 127 of 150 engagements (85%). Data arrived incomplete, error-ridden, and in inconsistent formats. Respondents also noted manual alterations to data due to accessibility issues with the financial institution’s legacy systems.

“Rows go missing from their extract and nobody flags it on their end, so I rebuild missing lines myself before run starts, I dont wait on feedbacks from them because feedbacks never come anyway, not in time for payroll cutoff…”Individual Contributor, BPO, Corporate and back-office functions

“…reference data behind the AML screen is old by years, i kindly asked more than once for someone to refresh it properly and nothing comes back, so i chnage it myself before it feeds into anything, as such the model is working off numbers i already touched and nobody upstream really know that in detail, i dont raise it as formal thing internally, its just what i do quiet between cases.”Individual Contributor, ITO, Financial crime and client onboarding

“So I work off an export thats a day old and I reconcile the duplicates by hand before anything runs. Every morning. Its about forty minutes. Nobody has ever asked me how long it takes because nobody has ever asked whether it happens. … the whole point of the engagement was to take manual effort out of procurement, and there is a person doing forty minutes of manual reconciliation every morning so the automation can run.”Individual Contributor, CXM, Corporate and back-office functions

Altering Financial Institution Processes

Respondents reported manually altering financial institutions’ processes in 132 of 150 engagements (88%). Delivery staff changed steps, rerouted work, optimized for AI, and wrote their own versions when no written process existed. Respondents routinely noted making key process decisions when the financial institution did not respond or responded late to clarifying questions. Respondents frequently cited conflicting clarifications from the bank as a root cause for manually altering processes without informing the bank.

Critically, quotes indicated that many processes the financial institution provided to the delivery team differed materially from final-state processes activated in production.

“It was frustrating. There is no written process for de-provisioning, so we watched what the client team did on a call and copied it, because the person who knows is always in another meeting and nobody wants to own the call on scope.”Individual Contributor, ITO, Technology engineering and operations

“Every desk contact on their side confirms trades a different way. Some use the Portal. Some just email a PDF. Some call and read numbers off a screen. And there is no single process document that says which one is correct for which counterparty, so we made our own version months ago and just follow that now.”Individual Contributor, BPO, Trade, settlement and reconciliation

“I have three different people on their side who each believe they own the servicing queue, and none of them will overrule the other two, so what we get back is silence, and silence in this setup means my team decides. Which they should not be doing. They are servicing analysts, they are not there to make policy for a private bank at eleven at night.”Engagement Manager, CXM, Client servicing and communications

Resolving Non-Deterministic AI Outputs

Non-deterministic AI output was a major challenge, with respondents noting frequent manual changes to agent output in 108 of 150 engagements (72%).

“Model output shifts between runs, nobody on the client side is willing to own the call, and change control cannot keep pace, so I sign off on adjusted numbers myself rather than let the deadline slip again. I know it is not strictly how this is meant to work, and I have said so to both sides without anything changing.”Engagement Manager, KPO, Corporate and back-office functions

“the problem is same case, same data, different result next run, so i correct it manually before it goes to QA, this happen more than i like to admit on dispute side, you rerun same chargeback case and get different recommendation back, no warning, no explanation, and dispute case cannot have mistake in it because Reg E deadline does not care that model changed its mind, so i sit there and go through informations again myself before submitting …”Individual Contributor, BPO, Payments, cards and disputes

Altering Customer-Facing Reporting

Respondents noted manually altering customer-facing reporting in 127 of 150 engagements (85%). Respondents selectively filtered KPIs, used data from prior reporting periods, and employed other methods to improve the vendor’s claimed performance.

Respondents regularly altered reporting to credit AI for manual work completed in parallel with or instead of agentic AI.

Respondent quotes hinted at internal moral conflict and significant pressure from the vendor to prove the viability of the outsourcing firm’s agentic AI core offering.

“Reconciliation-break reports showed a shrinking break count each week, but the drop reflected a human logging in manually to clear breaks the agent could not touch, not the agent resolving more on its own. Settlement-funding reports never separated agent-initiated actions from human keyboard fixes, so the dashboard read as steady progress right up to cancellation.”Individual Contributor, BPO, Trade, settlement and reconciliation

“…reported QA pass rates against agent outputs that had already been silently corrected by a human reviewer before scoring… the dashboard never reflected the raw error rate the client was actually exposed to.”Individual Contributor, CXM, Client servicing and communications

Manually Resolving 3rd Party Ecosystem Challenges

Written answers from 63 of 150 engagements (42%) described delivery staff manually resolving third-party ecosystem access issues, service outages, and data problems. Providers changed data schemas without notice, and staff rebuilt mappings or keyed in data themselves.

Respondents often described their ingenuity in collecting data from 3rd parties when no programmatic access was available to agents.

“The market data feed changes shape maybe once a quarter officially and about once a month actually, and the once a month ones nobody announces. So you find out because something breaks. And then you rebuild. I have a set of scripts now for the four most common breakages which is either good engineering or a sign that something is badly wrong depending how you look at it.”Individual Contributor, ITO, Technology engineering and operations

“…the problem is the Rates feed changed its field names on a wednesday and nobody told us. every model that pulled from it came back blank and we lost two days keying rates in manually. i asked who owns that supplier, nobody could say. there is no list “…the problem is funding file comes in short rows most days, nobody flag it on their end, i rebuild what is missing before i can even discuss about it with lead because cutoff does not wait for a conversation to happen first. i tell lead after, not before, thus its logged somewhere but not before the fix already went in, and client never hear about any of this, we dont discuss about it with them at all…”Individual Contributor, BPO, Trade, settlement and reconciliation

All engagements in the study had difficulties with defined AI governance, oversight, and control functions. Respondents submitted comments in every engagement, noting that such control functions were ineffective and that the delivery team often put forth only token effort in support.

Two Control Functions of Note

Respondents submitted quotes highlighting shortcomings in manual review and change control functions in 136 of 150 engagements (91%).

“There was a joint operating forum on paper — I have the terms of reference still, names on it, a weekly slot — and in eight months it met four times, twice with only my side in the room. … the review that was supposed to be joint became my team checking their work and their work checking nothing.”Engagement Manager, CXM, Corporate and back-office functions

“What we do now is we keep a group chat and that is the real record. Anything that needs a decision goes there, somebody screenshots it, and that becomes the approval. Nobody signed anything. … The other thing is the review. It is me reviewing my teammate and him reviewing mine, we are both tired and we both want to go home, so the second look is not really a second look.”Individual Contributor, ITO, Regulatory, audit, tax and legal

Manual Review of Agentic AI Products (Human in the Loop)

Written answers from 119 of 150 engagements (79%) faulted joint human review and quality assurance. 48% of 787 respondents rated the reliability of the joint quality assurance loop in catching agent errors before they reached live operations as distressed. The delivery team described the overall volume of agentic AI as overwhelming for both the bank and the vendor.

Respondents also noted that much of the manual review occurred on the final AI output, without the ability to validate internal processing or incremental product changes. The implication is that the only review that occurred may have validated that the AI output was consistent, but could not validate accuracy.

“We set a target of reviewing a sample of guided filings each day and we are not hitting it. In practice the reviewer opens the transcript, checks the loss date and the policy number match, and signs it. That is not scrutiny, that is a tick. Nobody has time for more than that at current volume, and volume is going up when the pilot widens in the fall.”Engagement Manager, CXM, Client servicing and communications

“The checker on our side signs off two hundred fee lines a day and the approval on their side is a same-day formality, so a wrong tier rate can pass both and only surface when an advisor complains a month later.”Engagement Manager, BPO, Payments, cards and disputes

Change Control

Written answers from 99 of 150 engagements (66%) were highly critical of formal change control functions. 48% of 806 respondents rated how well the two firms jointly governed change as distressed. Most respondent quotes indicated that the implemented change control functions were designed in a pre-AI context and that any change control board met weekly or monthly. However, respondents frequently noted that agents needed behavioral tuning far more often, and no defined rulebook specified which changes required a formal change control review.

Further, respondents highlighted that many of the agents operated by the outsourcing vendor were part of a larger ecosystem of agents provided by the bank and third-party vendors. Respondents noted that change control within the broader agent ecosystem was ineffective. The delivery team did not receive timely or effective communication about changes to bank- or third-party-owned agents.

“In February the agents began handling the interaction effect differently after a platform update on our side — nobody had asked for that. It took eleven days to establish whose change board owned the question, by which time three monthly attribution packs had gone to the client’s investment committee. Neither firm could show the other what had actually been done in that window. We closed it by recalculating the quarter by hand.”Portfolio Director, KPO, Trade, settlement and reconciliation

“What we did was stop using the change board for anything that mattered. It met fortnightly and the agents moved faster than that, so we built a parallel list, one sheet, updated every evening, of every case where the model had done something we had not agreed to and what we had done about it. Their underwriting lead and I read it at nine each morning and cleared it between us, and whatever survived that went to the formal forum a fortnight later as a note rather than a decision. It worked, in the sense that files kept moving. It also means the actual governance of this engagement was two people and a spreadsheet. Nobody asked us to do it and nobody stopped us, which I think tells you where the accountability sat.”Engagement Manager, BPO, Credit, underwriting and suitability

Cause

Respondents consistently cited three primary challenges limiting the use or effectiveness of defined control, oversight, and control functions:

1. Agentic AI Production Speed and Volume

The volume and speed of agentic AI output far exceeded the design of existing control functions. 48% of all indicators related to the speed of manual review across all control, governance, and oversight functions were rated as distressed. Manual review sessions were not scheduled frequently enough. Even when manual review sessions occurred more frequently, the volume of agentic AI output exceeded the capacity of the review team — 47% of all respondents rated manual review capacity issues as distressed.

“The human review point on arrangement offers has quietly become an approval step rather than a scrutiny step. The reviewer opens a queue of forty and releases thirty-eight, which is not review in any sense a supervisor would accept. That is a capacity design problem and not a people problem. The governance we jointly operate was drawn up for a world in which a person made each offer and could be asked afterwards why, and it does not sit comfortably over work of this shape, however well both sides chair the meetings.”Portfolio Director, CXM, Payments, cards and disputes

“Their change advisory board meets Tuesday and Thursday, and the agents do not know that. A remediation we approve on Monday sits nine days before it is allowed anywhere near live, and in nine days the estate has already moved on. Nobody on either side is behaving badly here, the calendar simply belongs to an older way of working.”Engagement Manager, ITO, Technology engineering and operations

2. Bank Decision-Making Speed

In 134 out of 150 engagements, respondents noted difficulties with the bank’s decision-making and response rates. The delivery team struggled with competing decision-makers or no decision-maker at all. When the bank assigned a single decision-maker, the delivery team often waited for responses to key questions. In some cases, the delivery team spent more time waiting for the bank to respond to questions than doing production work.

“The problem is nobody owns the sign off. That is the whole thing really. What would have to change. Give one person on their side the authority to approve a validation change on the day. One person. Not a board that meets Thursday. Because by Thursday i have already patched the test suite and moved on, and then i am sitting in a call explaining a fix from last week that they are approving now. It is backwards.”Individual Contributor, ITO, Trade, settlement and reconciliation

“we changed it because the client sign off takes eleven days and the deal timeline does not have eleven days in it. so we make the call and the model goes out and it is right, it has always been right, that is not the issue. the issue is if anyone ever asks how the number was arrived at, the answer involves a decision that i made on a tuesday afternoon because nobody else would.”Individual Contributor, KPO, Corporate and back-office functions

3. Outsourcing Vendor Internal Functions

Written commentary faulted the vendor for challenges with bank-mandated control, oversight, and governance functions in 105 of the 150 engagements (70%). Of note, 47% of structured indicators related to the vendor’s internal quality control and compliance functions were rated as distressed. Comments indicated that outsourcing vendors were new to agentic AI delivery and still struggling to implement basic, yet effective, functions required by all customers.

“Last week the agent matched a cash item onto the wrong value date and it went out with the pack, nobody had looked at it. For my manual clears there is a checker, it has always been that way, but whatever the agent matches goes straight through. My employer has not built any checking for that side of it …”Individual Contributor, BPO, Trade, settlement and reconciliation

“Review load fell on my team. Both sides agreed a checkpoint at the end of each pipeline build, which sounds fine on paper, but in practice it is four of my engineers reading generated code until nine at night. Two asked to move off in month two. i had to discuss about it with my manager and there was nothing to offer them.”Engagement Manager, ITO, Technology engineering and operations

“My honest feeling is that we are selling ahead of our own playbook. … the route from a promising cardholder use case to something a bank’s risk function will sign off is still being invented on each engagement rather than followed. The training we give does not close that gap; it teaches tooling, not how to walk an agent through a bank’s approval gates.”Portfolio Director, CXM, Payments, cards and disputes

Control, Oversight, and Governance Functions as Showmanship

This study did not provide enough depth to accurately capture specific regulatory or legal risk. However, the structured data and commentary respondents provided are alarming. The study indicated that many of the core control, oversight, and governance functions required of financial institutions (or companies in any highly regulated industry) may be ineffective. Delivery staff described approvals being clicked through as unread and sign-offs written long after decisions were made. In many cases, respondents described bank-mandated control, oversight, and governance functions as performance theater.

“the problem is nobody ever agreed where a person actually looks at the output. Agent does gross to net and then it just goes, there is no point in the run where somebody on their side puts a name against it. On our side my Lead checks what he can, but he has four people and one cycle is eleven thousand records, so he checks sample and that is it.”Individual Contributor, BPO, Corporate and back-office functions

“review is mostly quick approve, i will be honest… when queue is long the checker just click through. but this is not because anyone is careless, the volume of claimant messages is very high and both side are only small team.”Individual Contributor, CXM, Client servicing and communications

Engagements in this study averaged 5.69 months in length at the time of data collection. No vendor, respondent, or financial institution had material experience with agentic AI in general. The overall lack of experience and understanding of agentic AI underpinned every finding in this study.

Vendors Are New to Agentic AI

Of those able to answer (571 out of 850 respondents), 82% noted the vendor had a year or less of experience delivering agentic AI outsourcing services. No vendor in the study had over two years’ experience. Without practical experience delivering agentic AI outsourcing services, vendors struggled to develop an effective delivery playbook and to understand the full commercial realities beyond the previous outsourcing era.

“Not one person at my firm has done this before. Nobody. They sold agent desk support to a trading bank and then went looking for who knows how to run it, and the answer was nobody … That is not malice. That is a firm out of its depth that will not say so out loud…”Individual Contributor, ITO, Technology engineering and operations

The Delivery Team Lacks Agentic AI Experience

No respondent reported over one year of personal agentic AI experience. For 46% of 799 respondents, the current engagement was their first agentic AI commercial project.

“They put me on this with a half day of onboarding and a link to a recorded webinar. No one on my team had done agentic before. When the Agent started closing tickets wrong the answer was that its the clients data. Maybe it is. We still had no plan for it…”Individual Contributor, CXM, Corporate and back-office functions

The Delivery Team is Young, Inexperienced, and Does Not Know the Financial Industry

Individual contributors, the delivery team’s front line, were young, inexperienced, and lacked a solid understanding of the financial industry. Individual contributors reported a mean career length of 1.9 years. Average financial industry experience was 1.1 years. 40% of 615 individual contributors could not identify any legal or regulatory obligations governing their work.

“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

“I don’t know if it is a Gen Z issue or if it is because they have never had a job before. They may be smart and have some AI skills, but I can’t get them to do basics like responding to a client email or showing up to meetings on time. I spend more time teaching them how to be an employee than I do about the actual work.”Engagement Manager, KPO, Regulatory, audit, tax and legal

Outsourcing Vendor Leaders Lack Material Agentic AI Understanding

Respondents frequently noted that their employer and senior leadership lacked a basic understanding of agentic AI and were struggling to modernize the commercial model from the previous headcount-based outsourcing era.

“Put it plainly, we were sold as an agentic engagement and then priced and staffed like an ordinary ITO ramp. Nobody above me could explain what we bill for once the agent clears an access request itself.”Engagement Manager, ITO, Technology engineering and operations

“the company keep saying we are AI-first in the town halls… i do not see it on my screen. what i see is one person, me, sat checking policy answers at eleven at night because nothing tells us when the thing has gone wrong.”Individual Contributor, CXM, Corporate and back-office functions

“Im a fresher and know more about AI than anyone in our executive team.”Individual Contributor, KPO, Corporate and back-office functions

Banks Are Also New to Agentic AI

Respondents noted that the financial institution (their customer) lacked material agentic AI experience and basic understanding. 53% of 793 respondents rated the financial institution’s staff’s agentic AI knowledge on the engagement as distressed.

“Their people are not bad. When i escalate to relationship desk they reply same day mostly, they know their clients well. But they do not understand what agent is doing. They think it reads a mail the way a person reads it. So feedback we get is route better. That is not something i can put into a rule.”Individual Contributor, CXM, Client servicing and communications

“And their platform team are good people, they know the deal flow better than we do, but nobody there has worked with an agent before, so when I ask what the expected output should be I get an opinion, not a spec.”Individual Contributor, ITO, Technology engineering and operations

All Parties are Learning Real-Time on Paid Engagements

Outsourcing vendors and the delivery team learned in real time on paid engagements. Written answers from 71 of 150 engagements (47%) described defining best practices, creating the vendor-side delivery playbook, and collecting functional knowledge during the delivery of commercial services to the financial institution.

“We were given two days of Training on the agent console and then put on live Tickets. When the routing broke on the second week nobody in my company had done this before, consequently we called a person from another account who also didnt know. There is no playbook, we invented one on a whiteboard.”Individual Contributor, CXM, Corporate and back-office functions

“When the agent closed a batch of alerts wrong in March, my lead asked me what the process is, and i asked him the same thing. There is no process. So we made one up on the call and wrote it after.”Individual Contributor, ITO, Financial crime and client onboarding

The study routinely found financial institutions attempting to replace a component of a larger process with agentic AI, with little effort to optimize the entire process to match agentic AI’s capabilities. An analogy given applies:

“Companies are attempting to replace the horses in a horse-drawn carriage with cars and fail to see the potential of the car itself. There is no longer a need for the carriage.”

53% of 804 respondents rated the financial institution’s willingness to redesign its processes for agentic AI as distressed. Respondents frequently wrote about processes still built for people rather than agents.

“What is missing is they want the old manual test process copied exactly as it is, nothing changed, and the people we sit with have never worked with agents before.”Individual Contributor, ITO, Technology engineering and operations

“their process leans on a person catching the miss, the agent has no such instinct, so i step in myself before the model runs. that is the whole thing really. the workflow they documented assumes an analyst who notices when a comparable looks wrong, and the agent does not notice, it just uses it. so somebody has to be the analyst. that is me most days. … i think the difficulty is that rewriting it means somebody admits the original design was wrong, and the original design has names on it.”Individual Contributor, KPO, Corporate and back-office functions

In the previous outsourcing era, people absorbed the inconsistencies or shortfalls in processes, data, and systems. Delivery staff described colleagues who knew which values to swap or which records needed fixing — checks that were never written down. Agents do not absorb inconsistencies. They stop, pass the error along, and retrain on these inconsistencies.

“We spent months with the bank mapping processes…. No one in the bank consistently followed the plan. Every due diligence packet we got was different. AI was self-training based on these variations. We lost control of the quality of the output.”Engagement Manager, BPO, Corporate and back-office functions

“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.”Engagement Manager, KPO, Client servicing and communications

The study examined five agentic AI outsourcing performance categories through 100 structured indicators. None of the five categories are performing well. Financial institution indicators averaged 3.55 out of 10 across 150 engagements, the lowest of the five. Co-management averaged 3.65. Vendor-specific indicators averaged 3.67 out of 10.

The two technology categories rated higher, though still below the middle of the ten-point scale. AI technology averaged 4.60 out of 10 across 150 engagements. The third-party ecosystem averaged 4.64.

Challenges specific to financial institutions produced the lowest indicator mean across the study. 100% of all engagements in the study recorded written quotes faulting the customer’s data, processes, governance, engagement management, decision-making speed, and AI ownership.

“They handed us a model template in week two and told us it was the house standard – it wasn’t. Every banker had their own version of the comps build, and the differences were not cosmetic, they sat in how the peer set got screened and how non-recurring items were treated. None of it was written down. 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.”Engagement Manager, KPO, Corporate and back-office functions

“their spec docs were written for a person doing this by hand, field definitions dont match between the two systems they gave us and half of it is in a pdf from years ago. nobody on their side knew what an agent even needs to be handed.”Individual Contributor, ITO, Payments, cards and disputes

AI Ownership and Response Rates

52% of 802 respondents rated clear ownership of AI inside the financial institution as distressed. 51% of 789 respondents rated the institution’s response speed when the engagement raised an issue as distressed. Without clear ownership or timely answers to critical questions, the delivery team acted without guidance or awareness from the financial institution.

“My last engagement was a smaller bank and their finance team gives us one owner for the model, one person to ask. Here nobody owns it. I send a question on the entity heirarchy and it goes round three teams and comes back to me.”Individual Contributor, KPO, Corporate and back-office functions

“What is missing is one person with authority to say this source is approved, go ahead. Ownership for AI sits nowhere in particular, it sits with a committee, and committee meets monthly.”Engagement Manager, KPO, Corporate and back-office functions

Financial Institution Data Readiness

Written answers from 142 of 150 engagements (95%) faulted the institution’s data. 51% of 809 respondents rated the consistency of formats and definitions across the institution’s source systems as distressed. Delivery staff described duplicate data, schema variations, inconsistent formatting, accessibility issues, and numerous other data challenges.

“What did not move at all in four months is the supplier master. The same vendor sits in there three times with three tax numbers, spend categories are free text in one region and coded in another, and nobody owns that record. We are asking agents to reason over a file that no human has agreed on.”Engagement Manager, BPO, Corporate and back-office functions

“Same customer attribute means one thing in card platform and another thing in deposits, and nobody inside client organisation holds authority to say which definition wins for a regulatory submission. … What they could not give us was one agreed version of a number, and in reporting work that single gap poisons everything downstream of it …”Engagement Manager, KPO, Regulatory, audit, tax and legal

Financial Institution Process Readiness

54% of 795 respondents rated how completely the institution’s processes were mapped and documented to the depth required by agentic AI as distressed—the highest share of distressed ratings among the study’s 100 root cause indicators. Delivery staff described processes held in people’s heads and answered from memory.

51% of 801 respondents rated the consistency with which the institution performed the same processes across its business units and regions as distressed. Respondents regularly noted competing processes within the customer, without clarity. In most cases, the delivery team standardized on one process without consulting the bank.

“… the client is four companies that never merged their HR rules. Leave entitlement, asset return, even the shift codes are different in each one, also nobody there could tell us which version was correct. We were asked to automate a process that has four versions and no owner.”Engagement Manager, CXM, Corporate and back-office functions

“We asked their compliance group for the field that flags a client as high risk. Three people, three different fields, and two of them were in a system nobody could get me a login for. We built against the wrong one for five weeks. And none of it was written down anywhere. Then a manager on their side asked us why the alert counts looked off.”Individual Contributor, ITO, Regulatory, audit, tax and legal

Financial Institution People and Skills

53% of 793 respondents rated the agentic AI knowledge and skills of financial institution staff working with the engagement as distressed. Delivery staff described customer employees who had never worked beside an agent and who treated its output as a search box or a report to argue with. Numerous quotes noted that the bank’s internal team struggled to determine whether the vendor’s delivery team or any given agent produced a volume of work.

Agentic AI has fundamentally changed the outsourcing delivery model, and best practices from the previous outsourcing era no longer apply. 49% of 811 respondents rated the institution’s management of its side of the engagement as distressed. 50% of 794 respondents rated the institution’s readiness to absorb the change agents introduced into daily work as distressed.

“And their side does not know what an agent is. Not their fault, nobody trained them … they have never seen it run.”Individual Contributor, CXM, Payments, cards and disputes

“… i think the truth is nobody above them has decided this is anyone’s job. they told us in the kickoff there would be a dedicated owner from their side, that person moved to another department in month two and was not replaced. so we carry on. … they are just not ready and nobody told them they had to be.”Individual Contributor, CXM, Technology engineering and operations

The co-management best practices and functions from the prior outsourcing era no longer apply in an agentic AI context. The fit of the joint governance model to work done by agents was rated distressed by 49% of 803 respondents.

“On my last engagement we had a standing call twice a week with the client’s ops lead, and if something broke you knew who to go to and it got decided that day. Here i never even learned who on their side owned the assistant. My lead would ask and it would come back as, they’re checking. For two months.”Individual Contributor, CXM, Client servicing and communications

“We were so busy fixing pre-prod issues that we never thought about how to do the actual work.”Engagement Manager, KPO, Corporate and back-office functions

Decision-Making Speed and Ownership

Issues took too long to resolve because ownership was unclear. 49% of 809 respondents rated the gap between the response rate required by AI and reality as distressed. Delivery staff described two firms keeping separate issue lists and arguing over ownership rather than resolving issues promptly.

“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. … So when something is wrong the first hour goes to argument about whose step it was, not to the repair.”Individual Contributor, ITO, Technology engineering and operations

“When the assistant sends a HR case to wrong queue and it sits four days, we have meeting about it and both sides explain why it is not their part. And then next week same thing. My lead says we cannot change the routing rules ourself because it is joint decision, but joint decision takes longer than the SLA it is breaking.”Individual Contributor, CXM, Corporate and back-office functions

Change Control

Change control cannot keep pace with the speed and volume of agent work. 50% of 798 respondents rated how well joint change control keeps up with changes in agent behavior as distressed. Change boards met weekly or monthly, while agents needed daily adjustments. In some engagements, changes made by one firm were never disclosed to the other. The change only surfaced after something broke.

“One thing i want understood is that the client is not unwilling. They are quite open to changing how the work is done. The problem is change board meets weekly and our runbooks need adjusting more often than that.”Individual Contributor, ITO, Technology engineering and operations

Growing Distance Between the Delivery Team and the Customer

The delivery team has less visibility inside the financial institution than in the prior outsourcing era. Given the delivery team’s inexperience and the overall newness of agentic AI in outsourcing, this gap is magnifying delivery challenges and increasing risk to the financial institution. Delivery staff described months on an engagement without speaking to anyone at the bank. Questions went up through their own management and came back second-hand, if at all. Those doing the financial institution’s work, and making its decisions by default, had no direct line to the people who owned it.

“I never speak to anyone from the client side, not once in eight months. So these questions are asking me to rate people I only know from the file names on a deal folder.”Individual Contributor, KPO, Corporate and back-office functions

“There is no together. There is us, and there is a name on an email chain who i have never met.”Individual Contributor, KPO, Regulatory, audit, tax and legal

Vendor-side indicators averaged 3.67 out of 10, the third-lowest among the five categories. Respondents noted internal vendor challenges with evolving the basic outsourcing commercial model to agentic AI realities in 119 of the 150 engagements. Data indicated significant challenges in finding qualified talent. Respondents discussed a fundamental lack of knowledge about agentic AI within the executive leadership team. Notably, respondents frequently highlighted a lack of standard operational procedures for delivery within an agentic AI context.

Vendor Leadership Still Learning AI

Delivery staff described senior leaders who had never run agentic AI work and did not understand it. The depth of the vendor’s own knowledge of agentic AI was rated distressed by 52% of 807 respondents. Some accounts had leaders promising what the firm had not yet built. Respondents wrote about vendor leadership writing proposals, making key delivery decisions, and building internal delivery playbooks without understanding agentic AI.

“…the people above me have never run this kind of work before, you can tell, they ask us how long something takes and then put a diffrent number in the deck. On the old account my manager had done the job himself for years, so he knew when I said a batch was bad. Here they dont know, so they just push.”Individual Contributor, BPO, Corporate and back-office functions

“The sales team pitched our dev kit we built as being optimized for AI. No one in sales checked if it was. We panicked for a couple of weeks as we validated AI compatibility after contracts were signed.”Portfolio Director, ITO, Technology engineering and operations

The Outsourcing Commercial Model is Still in Transition

Respondents regularly noted that their employer still struggled to transition the basic commercial model from the pre-AI outsourcing era. The vendor’s ability to price agentic work profitably was rated distressed by 48% of 795 respondents. Managers described internal pressure to account for a reduction in hourly billings due to the increased use of AI.

“First, the pricing. 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 still sell this account on a rate card built for bodies. When the pipeline agent absorbs work that used to take two engineers, the commercial conversation that follows is awkward and nobody at my level is given a line to take.”Engagement Manager, ITO, Technology engineering and operations

Outsourcing Firms Lack a Viable Agentic AI Delivery Playbook

Staff described vendors without standard operating procedures for agentic delivery. The maturity of the vendor’s playbook for taking agentic work from concept to governed production was rated distressed by 49% of 807 respondents. Delivery staff described playbooks originally designed for staffed call centers, runbooks never written down, and slide decks offered in place of a method. Without an agreed method, each team invented its own.

“There is no playbook. People kept telling me there is a playbook and every time I asked for the same, what came back was a slide deck from a conference.”Engagement Manager, ITO, Technology engineering and operations

“We sold release automation to a global bank and we do not have a working reference of it anywhere. When it comes apart, nobody above me stands up and owns the same.”Engagement Manager, ITO, Technology engineering and operations

The Agentic AI Talent Shortage is Affecting Engagements

Vendors are struggling to find qualified people. 48% of 802 respondents gave the supply of qualified agentic AI talent a distressed rating. Managers said they asked for experienced specialists but received whoever was available, including team members with little functional or financial industry experience. Time meant for delivery went to teaching new staff the basics.

“We went back to leadership twice over the summer and asked for two senior drafters with capital-markets background, and both times we were given people who were available rather than people who were right …”Portfolio Director, CXM, Client servicing and communications

“In week three I asked for two analysts who had done entitlement reviews before. What arrived were two people who had completed the internal AI course and had never seen a bank access review. I spent a fortnight teaching segregation of duties instead of running the engagement.”Engagement Manager, ITO, Technology engineering and operations

Non-Deterministic Output

AI is not ready to deliver at the standards the financial industry demands. Agent output is non-deterministic. Delivery staff in 143 of 150 engagements (95%) described agent output they could not rely on. 32% of 800 respondents rated the consistency of the tools’ output, given the same input more than once, as distressed.

“i run it twice and get two different answers each time, and it’s not just twice, i’ve run the same case five times in one afternoon and gotten five different numbers back, so nothing here holds still long enough to actually sign off on, and that’s the part that gets to me, not that it’s wrong sometimes, but that it’s never the same wrong twice, i can’t even build a pattern around it.”Individual Contributor, KPO, Corporate and back-office functions

“run the same case twice and get two different answers, and there’s no fast way to get anyone to say which one’s actually right, so i just pick the one that looks closer to what i’d expect and go with that, which honestly is basically the QA process at this point.”Individual Contributor, CXM, Client servicing and communications

Black-Box Processing and the Lack of a Processing Audit Trail

33% of 800 respondents rated the tools’ ability to explain and evidence each output, at the depth an audit requires, as distressed. Logs recorded the result but not the steps behind it. Some tools reported success when the task had failed. Compliance teams asked how a figure was reached and found only the final number.

“AI invented a husband for a widow on her loan application. I looked at the processing logs. I don’t know why it thought this was important, or how it created this new person.”Portfolio Director, BPO, Corporate and back-office functions

“A remittance narrative on an incoming file contained a line of text addressed, in effect, to the [AI] reader, instructing that the payment be treated as previously cleared. The tool obliged, and it recorded the case as closed with high confidence in its own log, which is the part that troubled me most: not that it was fooled, but that its account of what it had done was confident and wrong, and our reviewer had no way to see the instruction that caused it. … But I would not yet say we could evidence to a regulator how any single disposition was reached.”Portfolio Director, KPO, Financial crime and client onboarding

Although scoring higher than organizational performance categories, the broader ecosystem of third-party data, systems, and service providers averaged 4.64 out of 10 across 150 engagements.

Third-party Systems are Not Optimized for AI

In many cases, third-party systems had no programmatic access, or programmatic access that was not optimized for AI. Respondents wrote of manually pulling data from third-party systems, or writing scripts to scrape needed data from various websites.

“Their certification portal is a set of screens built for a person with a token in their hand, so every run needed one of my engineers sitting there anyway, and nothing on this form has a place for that. The other piece is licensing. Every seat we bought is a named human seat, and nobody at the processor could tell us what that means when the thing logging in is an agent running fifty times a night. We never got an answer before the plug was pulled”Engagement Manager, ITO, Payments, cards and disputes

“… most of them give us no proper interface either, so half the time we are driving a screen built for a person because there is no other way in. when that screen moves one field, the agent carries on regardless and produces something that looks correct. that is the part that worries me. it doesnt fail loudly. it fails quiet and the run comes back green.”Individual Contributor, ITO, Technology engineering and operations

Service Disruptions Not Communicated to the Delivery Team

Service disruptions reached the delivery team after the fact, if at all. 32% of 798 respondents rated how reliably word of an upcoming third-party change reached the people operating the affected agents as distressed. Delivery staff described service third-party system outages without any prior notice.

“We got no notice at all, not once. One morning the sandbox key stop working and i spend whole day finding out why. Nobody in my team knew, and person who knows is not in our Meeting anyway.”Individual Contributor, ITO, Technology engineering and operations

“…the agent kept reporting success with zero documents pulled for two days, we only noticed when the counts were low. the notice went to a mailbox my Lead does not check”Individual Contributor, BPO, Corporate and back-office functions

Third-Party System Changes Not Communicated to the Delivery Teams

Providers changed file layouts, field names, and data formats without notice reaching the delivery team. 31% of 801 respondents rated advanced notice from third-party providers before such changes as distressed.

“In the second month the HR feed provider altered the layout of the leavers file over a weekend, no notice reached anyone here, and the deprovisioning agents processed nothing for four days without raising a thing. We found it because a line manager asked about somebody who had left and still had access. Previously a person would have opened that file on the Monday and seen it looked wrong. That habit was never turned into a check.”Portfolio Director, ITO, Technology engineering and operations

“the benefits provider changed their file layout in March and nobody told the team operating it. two days of payroll went out with the wrong deduction column.”Individual Contributor, BPO, Corporate and back-office functions

All three categories rated weakest in this study, the financial institution’s own readiness, co-management with the vendor, and vendor capability, are within the institution’s control or reach. The study indicated that most of the manual alteration sustaining engagements and regulatory compliance challenges stem from those three weaknesses.

Within 12 weeks, the board must know operational realities across all priority engagements, contain the critical issues, and approve a funded plan to address bank-side, vendor-side, and co-management critical issues.

In the first week, the board appoints one senior executive to lead an audit of every agentic AI outsourcing engagement. The appointee holds full autonomy and authority over the audit, is independent of every executive who sponsors or runs an engagement, and reports to the board’s risk committee. A written mandate grants access to every record, every person on either side, and every vendor site, and the authority to impose interim controls. Any refused request goes to the committee chair. A second executive, named by the CEO, is accountable for the corrections the audit will identify.

Four instructions issued in the same week.

1. Counsel orders preservation of every engagement record on the institution’s side, including agent configurations, changes to rules and prompts, the agents’ outputs before and after correction, manual work records, and the calculations behind every report.
2. Management freezes every change to an agent’s rules or configuration without recorded dual approval and names a person on each side with authority to stop each agent.
3. Legal and Compliance determine the supervisory notifications required in each jurisdiction and the handling of privilege.
4. The CEO states to all staff the purpose of the review, the duty to cooperate, and the protection of anyone who reports in good faith.

Every week, the audit lead reports in writing to the committee chair, and at the end of the fourth, eighth, and twelfth weeks to the committee. Any critical finding reaches the chair on confirmation. By the end of the first week, the committee holds the signed mandate, both appointments, the preservation order, and the interim controls.

In the second week, the audit lead forms the team from Compliance, governance, and regulatory affairs, Legal, Human Resources, Procurement, outsourcing management, Risk, Technology, and Finance. An independent third-party firm joins as advocate for the vendors and operator of the confidential data collection channel. Conflicts of interest are cleared before work begins.

Each vendor receives formal notice of a comprehensive performance and compliance audit. The notice orders preservation of all work materials, invokes the institution’s audit and access rights, and requests the engagement’s contracts, change logs, agent configurations, manual work records, staffing, and the roster of the delivery team. The notice states that the audit gives the vendor a formal mechanism to raise bank-side (customer) obstacles with evidence. Each vendor provides written notification that no member of its staff will suffer repercussions for full disclosure.

A list of all outsourcing engagements is compiled the same week from the procurement and outsourcing registers, the contracts, the payment records, and the technology and model inventories. Pilots, transition work, subcontractors, and third-party tools are included.

In the third and fourth weeks, the team reviews every engagement from the institution’s own records and ranks each on financial exposure and the level of legal and regulatory obligation it carries. Impact on the institution’s customers, the autonomy and volume of the agents, and the manual work already visible weigh in the ranking. The ranking method is recorded and given to the committee. From the ranking, the audit lead defines the core group of high-risk engagements that receive the immediate audit and, where warranted, a corrective action plan. Interim restrictions apply to any core engagement whose exposure cannot wait.

Through the third-party firm, the team opens a confidential interview channel to every member of every delivery team, whatever their seniority or role, and to the institution’s own staff on the engagements. Participation is confidential and carries no consequence. Identities remain with the third party, and findings reach the institution in aggregate.

Questions cover what was manually altered, when, why, by whom, and whether the institution was told. A second set covers legal and regulatory compliance and the reports the institution receives. Improvements to the engagement are sought across the institution’s side, the vendor’s side, co-management, the AI tools, and the third-party ecosystem. Each roster arrives in the third week, and answers are due by the end of the fourth.

In Weeks five to eight, the team audits each core engagement in a face-to-face walkthrough with the vendor, across the full lifecycle from the institution’s preparation to production. Four areas are reviewed in full:

1. Review of the institution’s side covers its data and processes, its organization, its secondary and dependent processes, and its ownership and response rates.
2. Review of the vendor side delivery team, its playbook, its controls over manual alteration and regulatory compliance, and customer-facing reporting.
3. Co-management is reviewed for how the two firms decide, change, escalate, and resolve challenges throughout the engagement.
4. Technology is reviewed for the agents, their configurations and change history, their monitoring, and the third-party ecosystem.

Controls are tested as they operate. For the audit period, the manual alteration register is reconstructed, and each entry is traced to its cause. Reported performance is reconciled to the work performed, including the human work behind it. Skills on both teams are observed in the work. In parallel, every other engagement is reviewed against the questionnaire returns, and any engagement whose staff report undisclosed manual alteration or compliance concerns joins the core group.

From this evidence, the audit identifies the critical issues across high-risk engagements and prepares an emergency action plan. Each critical issue carries its containment, its owner on each side, its date, the assessment of harm to the financial institution, the notification decision, and the contract remedies. By the end of the eighth week, the committee holds the evidenced findings and the emergency action plan.

In Weeks 9 to 12, management corrects the critical issues found in Weeks 5 to 8. Each correction is retested against production evidence before Internal Audit accepts closure. A revised policy, contract, or attestation is not a closure. Where a critical issue cannot be corrected in these weeks, the accountable executive and the CRO restrict or suspend the activity or begin an exit. No affected use expands while a critical issue remains open.

Weeks 9 to 12 also produce the first-year plan for the board’s approval at Week 13. Four methods make up the plan:

1. Remaining engagements are reviewed in ranked order under the same walkthrough, evidence standard, and confidential channel, and every engagement is decided by the end of the year.
2. Co-management is rebuilt on every continuing engagement. Each side names a decision owner, one joint issue list carries agreed response times, agent changes pass through joint control, the manual alteration register is shared, and reports are reconciled to the work performed.
3. On the institution’s side, every engagement and material workflow receives a named owner with measured response times, dependent processes are redesigned for agents, and data is reconciled before an agent uses it. Engagement owners, process owners, and their overseers are trained in the preparation, active management, and co-management of an agentic AI engagement.
4. Vendor delivery teams are trained in the financial industry function, the law and regulation that govern it, and the institution’s processes, and demonstrate that competence before holding unsupervised responsibility. Vendor staffing is stabilized, playbooks are rewritten, and controls over manual alteration and disclosure are written into the contract.

A standing oversight, control, and compliance function for the agents is also created, under a single executive accountable for AI and reporting to the CRO. Under it sit the inventory of agents, the change control, the monitoring, review capacity sized to the volume of output, the authority to stop an agent, and the obligations mapped to each workflow. Every item in the plan carries an owner, a date, a budget, and a closure test applied by Internal Audit. A communication plan for the institution’s staff and the delivery teams accompanies it. By the end of Week 12 every critical issue is corrected or contained, and the funded plan is before the committee.

At Week 13, the board approves the first-year plan and decides whether each core engagement continues, continues under added controls, is restructured, is transferred, or exits. From Week 14, management completes the plan as approved. Each quarter, the committee receives a report with the manual alteration register, the closure record, the response-time measures, the competence results, and the remaining exposure. Internal Audit validates each closure before it is reported. By Week 52, the board holds a decision on every engagement, a complete and disclosed manual alteration register, controls that pass retesting, and reporting that matches the work performed. Lessons are written into the outsourcing policy and the contracting standards, and the review becomes a standing function for the life of every engagement.

This study found an immature delivery model sustained by manual work that, by the delivery teams’ own account, the financial institution was rarely told about. Every weakness behind it is within the institution’s reach. Within 12 weeks, the program above brings that work into the open, and within the year it corrects what it reveals. The board can make the first appointment this week.

Phil Hatch is a global outsourcing and workforce-development executive with more than 30 years of experience helping organizations and governments build, manage, and transform large-scale talent and service-delivery operations. He served as the CEO of ManpowerGroup/Experis global outsourcing business, where he led the development and growth of a global outsourcing organization serving clients across industries and geographies. His work spans multiple industries and geographies, with a particular focus on connecting business strategy, talent, and economic opportunity.

Beyond his corporate leadership, Hatch has served as a senior labor and economic development advisor to governments, helping public-sector leaders develop workforce strategies to attract investment, expand employment opportunities, and build globally competitive talent ecosystems. His experience sits at the intersection of outsourcing, workforce policy, economic development, and international business—giving him a practical perspective on how governments and companies can work together to create sustainable economic growth.

Following Russia’s invasion of Ukraine in 2022, Hatch was asked by the U.S. Department of State and USAID to serve as a formal advisor to Ukraine on workforce and economic development. In that role, he has focused on how technology, outsourcing, skills development, and international investment support Ukraine’s economic resilience and long-term recovery.

Today, through Akholi, Hatch continues to advise organizations and governments on the changing global outsourcing landscape, including the implications of AI and agentic AI for work, talent, and service delivery.