Credit where credit's due: Agents inform, humans decide

4 Minutes · Oct 12, 2026

Credit where credit's due: Agents inform, humans decide
Chuck Paiusi

Chuck Paiusi

Director of Product

Capital markets run on judgment applied to information. For most of the markets' history, the binding constraint has been the information.

A credit team could read only so many filings, reconcile so many positions, and trace so many counterparties before the deal had to close or the market moved on. So analysts triaged. They read what they had time for, and experience covered the rest. Every credit desk has lived inside that arithmetic.

That constraint is easing. A team that puts AI to work can read more of the record on each decision, test more assumptions against the data, and spend more time on the judgment itself.

We lend and manage assets onchain, where every collateral balance, loan term, and repayment is reported onchain, and our credit team uses AI agents to work through that data and the documents around it.

The operating framework

An analyst sets an agent to work, the agent returns a sourced review, and the analyst decides. No agent begins work on its own, and none of them issues a decision. A person on the credit team reads each review, challenges it, corrects it where it is wrong, and signs off before anything moves.

Maple's operational AI fleet

We made an early decision not to build one general assistant and ask it to do everything. Credit work rewards specialization, and so does the tooling around it. We scope each agent to a single job, give it the sources that job requires, and hold it to the standard we already expect of the Maple team. Analysts use them as a research and reference resource: They read documents, check figures, and raise questions, and the analyst decides what to do with the answer.

The onboarding agent helps review a new borrower, checking the entity and its owners and flagging what is missing or needs a closer look. The team signs off on every new borrower onboarding after reviewing the full file.

The creditworthiness agent reads a new borrower’s documents and recalculates the figures the borrower reports, so an analyst can see where the numbers hold and where they don’t.

The deal underwriting agent shows how a proposed loan would sit in the existing book, including concentration and headroom against risk committee limits, for an analyst to review.

The collateral review and onchain data agents help when a borrower proposes collateral we haven't accepted before. They research the collateral risk profile independently of the deal that raised the request. One reads the asset's own documents and summarizes custody and exit routes under stress. The other checks how much liquidity the market shows at the sizes we would need to sell.

The risk agent reads documentation from across the operations, credit, and trading teams and flags gaps and processes issues to the people accountable for them. A person reviews what it finds.

What we don't hand over

Humans at Maple still approve every loan, against limits the risk committee sets, and no agent has authority over them. An agent's output reaches the committee as a sourced review with a named analyst's sign-off, and that analyst owns what they signed. We retain every agent review, every correction an analyst made, and every sign-off as an internal audit trail.

None of this is new in principle. A credit committee, stated limits, and a record of who approved what are how credit has always been controlled. The change is the speed and quality of information discovered by the agents which is surfaced to the committee.

What got harder

Agents surface more than a person can follow up on, so the work has moved from gathering information to sorting and ranking it. The review process has also become more complex. A fully sourced 40-page report takes more work to check than a short memo, and an analyst who signs off on one without testing its weakest claim adds nothing to the process.

For most of the history of credit, the binding constraint was information. As agents take on more of the reading, more of the weight shifts to the analyst’s judgment and to testing a review that already looks finished. That is the point of the fleet: Our analysts spend less time collecting the record and more time questioning it, so each credit decision draws on more of the evidence.

Next in this series: How we build, on what the AI shift did to product and engineering.

Chuck Paiusi

Chuck Paiusi

Director of Product

Chuck leads product at Maple, bringing years of experience building and scaling products to millions of users across big tech and VC-backed startups. He previously worked in investment banking, which gave him a deep understanding of the institutional finance Maple brings onchain.

More Insights