Conclusions and the executive agenda.
The most impactful and likely largest single workload is in bank-side preparation, governance, and oversight.
By Phil Hatch · Akholi · First Edition, July 2026 · DOI 10.6084/m9.figshare.33005630
An infant industry, and the fix sits with the institution.
The primary finding of this study is that the agentic AI outsourcing industry is in its infancy. As with past industry transformations, the industry has yet to fully understand the context needed to achieve high success rates and strong project returns.
Failures that led to low customer satisfaction and high engagement cancellation rates were consistent in the data provided by respondents, and the level of customer preparation, governance, and oversight was the top callout by respondents.
Outsourcing vendors lack experience and an established playbook for providing agentic AI outsourcing services to any customer, not just banks, and are attempting to define that new playbook with a team that lacks any material experience with the technology itself.
As the customer and vendor resolve institutional and technical challenges, co-management becomes the primary success factor in production-stage engagements. The industry is applying best practices built over the past 20 years for issue and risk management, decision-making, and change control — all optimized for human-centric delivery, and none built for the speed and volume agentic AI delivery models produce. Every engagement in the study that achieved an ROI had invested in co-management planning before the engagement started, and every successful engagement kept co-management process improvement a senior executive priority throughout.
The bank’s core technologies and the broader 3rd-party ecosystem of cloud, SaaS, services, and data providers warrant intentional investigation and remediation, but the most impactful and likely largest single workload is in bank-side preparation, governance, and oversight. The study demonstrates that banks must immediately:
Thoroughly document, optimize for AI, and ensure consistency of all processes that will be affected by agentic AI.
Address data quality, consistency (normalization), and accessibility challenges that are routinely affecting agentic AI.
Optimize governance, oversight, risk, and compliance for the new era of agentic AI. If one has not yet been identified and communicated to outsourcing vendors, the bank must define a senior executive as the official owner for all agentic AI work.
Fix legacy systems access, allowing agentic AI to consistently access the tools and data it needs at the required speed.
The executive agenda
Six immediate actionsGate every agentic contract behind an institutional readiness standard
Assign a senior executive as the institution’s AI Tsar if the role does not yet exist, and gate every agentic contract behind a written readiness standard — covering data quality and accessibility, process documentation, legacy-system access, and expert availability — met in full before signature and reviewed against every existing engagement on a regular schedule.
Build co-management discipline before contract signature, and improve it throughout
Include a senior-executive-led co-management definition in every request for proposal and build the discipline jointly with the vendor before signature, covering risk management, issue resolution, decision-making, and change control optimized for agentic AI. Keep improving it throughout the engagement with transparent weekly reviews between named engagement managers, and make a vendor’s failure to give co-management material weight grounds for termination.
Audit the design and execution of vendor reporting, not its arithmetic
Commission a 3rd-party design audit of vendor reporting, repeated annually, that asks who defines each metric and its comparator, what counts as complete, what enters the sample, and whether human-corrected work is reported as automated success — and embed a bank employee in the vendor’s team until behavior changes if issues surface. Add a standing question set to every performance review asking the vendor to demonstrate how each dashboard and report is generated, with a bank technical and functional expert examining the mechanical process.
Restructure commercial terms around outcomes and exit
Price the work on measured, validated outcomes defined in the contract with the design-audit protections of Action 3. Write the exit while the institution still holds negotiating power — staged gates with evidence requirements, data and knowledge escrow, transition obligations, and termination terms that make abandoning a failing engagement a decision rather than a write-off.
Treat the human oversight layer as a control function
Every assurance the institution receives about supervised agentic delivery rests on people averaging under 1 year of experience with the technology they supervise, so treat the human oversight and development layer as the institution treats any other control function. Write in qualification standards for the people supervising agents, a retention strategy for key persons, visibility into turnover on the institution’s engagements, and verified human-capital development programs.
Listen directly to the vendor’s delivery team
The delivery team holds an undocumented depth of knowledge largely out of the customer’s sight, and listening to it directly gives the bank its best view of reality — begin immediate rotations between the bank’s team and the vendor’s delivery site, and hold joint retrospectives, including Individual Contributors, where the vendor’s team can speak freely. Take part in departing staff’s exit interviews to learn why they are leaving and how the engagement could improve.
None of the 6 actions requires a technology decision or the permission of a regulator. Each is a governance choice within the authority the institution already holds. The record already prices inaction: cancellation stands at 37.33% realized, with half the study set canceled or at high risk. Standards for this industry are being set now by whoever moves first, and that work belongs on the board’s agenda this quarter.
Continue the report
All parts →Why Engagements Fail
Five problem areas; the institutions on both sides come before the technology.
The Integrity of Performance Information
Reporting the producers rate more likely manipulated than accurate.
Study Information & AI Disclosure
Design, independence, statistics, limits, and the AI-assistance disclosure in full.