2 Reconciliation AI Agents, Live in Production.

Bank reconciliation is the process most vulnerable to a well-intentioned shortcut. The objective as everyone experiences it is to make the reconciliation close, and there are several ways to do that which are not the same as reconciling: match on amount alone and accept a coincidence, plug a small residual to a difference account, or classify an unclear transaction to sundry and move on. Each is a rounding error in isolation. Repeated monthly across a large transaction population, they produce a set of accounts where nobody can explain a specific balance, and the investigation to unwind it costs far more than the original resolution would have. Classification has the same failure mode: the narrative on a bank line is often genuinely ambiguous, and the fastest resolution is a default rather than a question.

These agents are built around refusing the shortcut. Bank Transaction Matching matches transactions to ledger entries and leaves the genuinely unmatched genuinely unmatched — no forcing, no plugging the difference to make the reconciliation close. The output is therefore sometimes less tidy than a manual one, and that is the point: the untidiness is information. Transaction Classification classifies bank transactions to the right account and cash flow category from what the narrative actually supports, and routes the ambiguous ones to a person instead of defaulting them to sundry. The volume of judgement calls a human has to make does not go to zero; what changes is that each one arrives already isolated, with the evidence attached, rather than hidden inside a population of two thousand rows.