AI Agents for Bank Credit Risk: What They Do and How Banks Control Them

An AI agent is software that can carry out a task made of several steps, not just answer a question. Given a goal such as "prepare the annual review for this borrower", it can find the documents, read them, pull out the figures, run the calculations, write a draft and hand it to a person. Along the way it uses tools: it opens files, queries systems and writes results back.

That last part is what separates an agent from a chat assistant. A chat assistant produces text for someone to read. An agent takes actions. When people talk about risk agents for banks, they usually mean agents that do the preparation work of a credit analyst or portfolio manager: gathering, reading, calculating, drafting and flagging. Because agents act, the question for a bank is not only whether the output is good, but what the agent is allowed to do and who signs off.

The best uses today are tasks with a lot of reading and arithmetic, clear rules and a person reviewing the result. The examples below cover commercial and retail credit at banks, credit unions and lenders. In each one, the agent prepares the work and a person with the right authority confirms or decides.

Preparing a decision and making a decision are different jobs. An agent can assemble the facts for a loan approval, but the approval belongs to the officer or committee that holds the authority under the bank's delegated authority matrix. The same applies to assigning or changing a risk grade, granting a covenant waiver or amendment, approving a policy exception, changing a customer's credit line and choosing a collections action that affects a customer.

There are practical reasons for this, not only cautious ones. These decisions carry accountability that sits with named people. They depend on judgment about things that are not in the file, such as the quality of management or a conversation with the borrower. And they are the decisions that loan review and examiners will ask a person to explain. A sensible design lets the agent propose and lets a person dispose, with the proposal and the final decision both kept on record.

Agents share the familiar risks of any analytical tool, and add some of their own because they act across several steps without a person watching each one. A credit team should be able to name these risks before it puts an agent to work, since each one needs a specific control and not just a general promise of care.

Instructions alone are not a control. Telling an agent to follow credit policy is like telling a new analyst to follow it: helpful, but not something you can prove afterward. The stronger approach is to check every action the agent wants to take against the bank's own credit rules before the action runs. The check sits outside the agent, so it works the same way whichever AI model is doing the work.

The rules are the bank's, written from its credit policy and delegated authority: who may approve what, which figures need a source, which data may leave which system, which steps need a second person. Each proposed action gets one of four outcomes. This is what AI agent policy enforcement means in practice.

AI agent governance depends on the record as much as on the rules. After the fact, a credit officer, an internal auditor or an examiner should be able to pick any figure in a memo or any step in a review and see how it came to be. If the only evidence is the finished document, the bank cannot show that its controls worked. A useful record answers the questions below for every action.

Banks already have a way of thinking about tools that produce numbers and recommendations. Model risk management guidance generally asks an institution to know which models and tools it uses, understand what they are for and where they are weak, test them before relying on them, monitor them over time, and keep people accountable for the decisions they inform. AI agents fit into that way of thinking, though they add questions about actions as well as outputs. Whether a particular AI tool is treated as a model under your policy is a question for your own model risk and compliance teams.

In general terms, examiners tend to care about the same things for AI as for any other part of the credit process: that the bank understands what the tool does, that credit decisions are made by people with authority, that the work can be traced to its sources, that exceptions are documented, and that the bank manages its vendors. A design where the agent prepares, a person decides, policy is checked before each action and every step is recorded gives a bank clear answers to those questions. It also makes it easier to start small, with one workflow and a defined set of loans, and widen the use as evidence builds.

Pick one workflow where the work is heavy, the rules are clear and the output is easy to check, such as spreading, annual review drafts or covenant testing. Run the agent alongside the current process for a set period and compare the results. Measure the things that matter to a credit team: whether the figures match the source documents, how many drafts needed correction and of what kind, how much time was saved, and whether every held action reached the right signer.

Write the rules before the agent runs, not after. Decide which actions are allowed, which need a signer and which are blocked, and agree who can stop all activity. Bring model risk, compliance, information security and loan review into the plan early. They will ask for the record anyway, and they often spot a missing control that the credit team did not think of.

Vendors describe AI agents in very different ways, and the word covers everything from a drafting assistant to software that acts on its own. The questions below help a chief credit officer or a risk team find out what a product really does and how it is controlled. Ask to see each answer working on your own documents, not on a slide.

In Aarvion Risk OS, AI spreads financial statements and drafts memos with figures traceable to source documents. Every AI action is checked against the bank's own credit rules, which can allow it, hold it for the right signer, block it or stop all activity, and it is recorded with who proposed it, which rule applied, who approved it and when. Decisions such as waive, amend, approve and grade stay with the bank's people.

On the covenant side, Aarvion tests every covenant on every facility and shows the required level, last actual, headroom and trend. It tracks breaches, waivers, compliance certificates and reporting ticklers, with facility, payment and past-due data coming from the bank's core system. Covenant results feed portfolio monitoring, annual review drafts and risk grade staleness. Customers start with a 90-day pilot on one workflow.