Who has to say yes before an AI agent POC for banks can start? Six groups are frequently involved, from new product approval and the AI risk forum through to outsourcing risk, data protection, information security, and procurement. Only some of these need to give the green light before the AI agent touches customer data, and knowing which ones in advance lets the governance work run alongside the POC build instead of holding it up.
This guide covers the approvals that sit in front of a bank POC, how to scope one, what has to exist before day one, and what the exit decision needs to produce. For the work itself rather than the process around it, start from our guide to AI use cases in banking.
What an AI agent POC for banks actually involves
Buyers in financial services say POC far more often than pilot or proof of concept, and they mean something precise by it. The agent runs on real volume, inside your own systems, under your own controls, for long enough to produce evidence that the AI agent is ready to run at scale in your business.
What it proves | Who sets the terms | |
|---|---|---|
Demo | The agent demonstrates its capabilities | The vendor |
POC | The agent handles your case types, on your systems, under your controls | You |
Production rollout | The agent scales to full volume, with the approvals already granted | You, inside what the POC cleared |
A POC in a bank runs as two projects at once. One is technical: integration, procedures, test cases, and live traffic. The other is governance: risk assessments, data protection sign-off, and a procurement. Most plans resource the first and absorb the second as it arrives, which is the wrong way round, because the governance project is the one that sets the date.

The approvals that gate a bank POC before it starts
In a regulated firm the stakeholder calendar often decides your timeline, not the build. The most useful question to ask internally is not when the agent can be ready, but which committee has to see it and when that committee next sits.
These steps come up most commonly:
New product approval: Many banks route anything customer-facing through a new product approval process before launch. Where it applies, this is the real go-live dependency rather than the technical work, and it can sit weeks out from a working integration.
An AI or model risk forum: A cross-functional gate covering operations, compliance, IT, and legal, convened for anything built on a large language model. Find out whether yours wants to see the POC design in advance or only the result.
Third-party and outsourcing risk: Whether the arrangement counts as material outsourcing changes the documentation, exit planning, and audit rights you need. The EBA's third party risk management guidelines set the current EU expectation, and the UK's critical third parties regime covers designated providers while leaving each firm's own responsibility for managing its third parties exactly where it was.
Data protection: A data protection impact assessment before any customer data reaches the agent, alongside the sub-processor chain, retention terms, and where inference happens.
Information security: The due diligence questionnaire. Clearing it is the gate to piloting at all, and how cleanly a vendor moves through it tells you a great deal about how many regulated customers they already carry.
Procurement: Vendor onboarding is routinely the longest pole in a bank POC, and the onboarding cycle can outlast the POC itself. Ask on the first call whether a shorter POC contract and a reduced documentation set are available, because buyers who want to start before a full assessment almost always have to ask for that path by name.
Not all these steps have to finish before a POC can start. Product approval and the AI forum's final sign-off can run alongside configuration, and a POC contract lets procurement complete in parallel.
Data protection and information security cannot, because both gate the moment the agent first touches customer data, and the outsourcing classification needs settling before scoping because it changes what you have to document. The wider approval sequence, from stakeholder mapping through to the go-live decision, sits in our guide to evaluating AI agents in financial services.

How to scope an AI agent POC for banks so it proves something
Scope is usually treated as a question about the agent. In a bank, however, it is also a question about the approval path, because the case types you choose determine which controls apply and therefore which stakeholders you need.
Choosing the case type also decides the approval level. A POC on card freezes and balance queries touches a different control set from one on hardship assessment and forbearance, where vulnerability and FCA Consumer Duty obligations apply to every outcome. Both are reasonable places to start. They are not the same approval path, and picking the harder one by accident costs you a month.
Settle the customer-facing question early. A POC where the agent replies to real customers needs consumer-outcome sign-off that a back-office-only POC does not. Some banks deliberately run the first POC with the agent drafting rather than sending, to shorten the approval path. That trade is worth naming out loud: a draft-only POC shows you the agent's reasoning but never its unsupervised behaviour, so it cannot carry a production decision on its own. Our guide on AI copilots versus autonomous agents covers where that line sits in regulated work.
Size the volume against the risk, not the calendar. Agree a minimum number of cases per case type up front, weighted towards the types your risk function cares about most. A POC that runs the agreed six weeks but only sees eleven complaints has not tested complaints.
Write the data boundary into the scoping document. Which fields the agent reads, which systems it writes to, and what it is never permitted to touch. Your data protection assessment will ask for exactly this, so drafting it at scoping time saves the round trip.
What has to be in place before day one
Four things have to exist before the clock starts, and a fifth tends to surface in week one whether you planned for it or not.
A named owner per gate: One directly responsible person for each approval, not a committee and not a distribution list. Vendors stall on regulated accounts mostly because nobody can say who gives the answer.
System access, agreed and tested: The agent needs to read and act in the systems where the work actually happens. Confirm the integration path and the permissions model before the clock starts, because access requests sit in the same queues as everything else.
Your QA rubric and the people to apply it: Reviewed conversations have to be scored against your existing rubric by your own quality team. A vendor scoring its own output is not evidence, and your risk function will say so.
A baseline you captured before the agent arrived: Handling time, cost per case, and current quality scores for the same case types, measured now. You cannot reconstruct a credible baseline after the fact, and without one every number the POC produces floats free of any comparison.
That fifth thing is your procedures. The agent learns your work from them, so a POC exposes how out of date those procedures are inside the first week. Treating that as a finding rather than a delay is what separates a POC that produces a decision from one that produces an argument.
Plum reached a 52% day-one resolution rate and a 98.6% quality assurance score.
"Gradient's AI solution delivered impressive results with minimal effort on our part. The proof of concept made the decision clear, and the rollout was seamless. Seeing such a high CSAT and resolution rate validated our choice."
How the POC runs, and who owns each phase
Score only the period after configuration ends, because numbers from the weeks when the agent is still changing measure setup quality rather than agent quality. Which numbers belong on the scorecard, and which ones flatter a result, is covered in our guide to AI agent POC metrics. At Gradient Labs, customer support and back-office deployments at large regulated financial institutions typically take four to six weeks to reach production.
Phase | What happens | Your side owns | The vendor owns |
|---|---|---|---|
Scoping | Case types, volumes, data boundary, approval dates | The gates and their owners | The technical design |
Configuration | Procedures, systems, tone, test cases | Procedure accuracy and system access | Building and correcting the agent |
Scoring window | The agent stops changing and runs on live volume | QA scoring against your rubric | Reporting, unchanged from the agreed definitions |
Review | The evidence pack and the decision | The committee papers | The improvement plan after the POC |
Put the start date of the scoring window in the committee papers next to the approval dates, so the forum that signs off the result knows exactly which numbers it is reading. For how to compare the resulting numbers across vendors without being misled by differing definitions, see our AI resolution rate benchmark guide.
The exit decision, and what it has to produce
A POC ends in a decision, and in a bank, that decision needs to be in agreement across stakeholders. These criteria will get alignment:
The results, reported the way you defined them: Per case type, against the baseline you captured, scored by your own team.
Evidence that failure stayed bounded: Your risk forum will want this as data from the POC itself, not as a policy: which guardrails fired, why each handoff happened, and the case record behind any decision it chooses to check. A control framework described in a slide proves none of it. Our guide to secure AI agents for banking sets out what to ask a vendor for here.
The commercial terms, with the definitions attached: Write the definition of resolution into the contract itself, alongside the exit terms and the move from POC to production pricing, because contract length and flexibility is the commercial risk buyers raise most often once price is settled.
The production path: Which approvals the POC already cleared, which remain, and what the rollout sequence looks like across case types. Our guide on how to deploy AI agents in banking covers the stage after this one.
Rollout approval does not close the governance work. Name who owns the improvement work after the POC, on what cadence, and which committee sees the agent again before each new case type goes live, because the gap between a POC that stalls and one that reaches production is usually resourcing rather than technology. Our guide on moving AI agents from pilot to production in financial services covers where that commonly breaks down.
Scope a POC around your hardest queue
The quickest way to find out whether a POC will clear your gates is to scope one against the work your team finds hardest. Book a demo and we will map it with you, including the approvals.
Elizabeth Shew leads Brand and Advocacy at Gradient Labs, where AI agents handle customer support and back-office work for banks, lenders, and fintechs. Before that, she led customer marketing at Mastercard and built Dynamic Yield's customer marketing programme from the ground up, a decade spent turning customer results into industry-shaping stories. She writes about how support and operations teams actually put AI and technology to work. Before tech, she was a professional dancer in NYC.

