Five Questions That Decide If An AI Agent Works In Finance

Mitul Tiwari is co-founder and CTO of Numos AI, which builds AI agents for finance teams.

Finance systems record the state of a transaction: the general ledger balance, the posted journal entry or the “paid” flag on an invoice. They capture what happened, but not the reasoning that led from one state to the next.​ Why was that accrual booked at $40,000 instead of $52,000? Why did the controller override an automatic bank match?

That reasoning gets reconstructed in an email the night before the board deck is due. I call this the “fragmentation tax”: the cost of re-stitching context that the ERP, bank feed and procurement tool each hold a piece of, but none hold together. Now we’re asking AI agents to exercise judgment over the close with none of that precedent in scope.

If 2025 was the year of the agent, 2026 is the year of the harness: everything that wraps a model so it can do real work, from context and tools to verification and audit trails.​ The question has moved from “Can the model do it?” to “Can the system around the model do it reliably with a trail an auditor would accept?”

Finance makes that question unusually hard. Coding agents have fast feedback from compilers and tests. Finance has no off-the-shelf unit test for “this reconciliation is correct.” The cost of a hallucinated number is not a broken build, but a restated earnings release. That is why finance co-pilots without a harness often follow a predictable arc: a demo that impresses, a pilot that drifts and a quiet rollback.

I’ve worked on this problem as a co-author of TapeAgents—an open-source framework built around durable records of everything an agent does—and while building a harness that runs inside live month-end finance closes. Here are five questions I would ask before deploying an agent in finance.​