What autonomous exception resolution actually means
Autonomous exception resolution means software identifies an invoice exception, pulls in the context needed to fix it, and closes the case itself, without a person opening a ticket, sending a chase email, or keying a manual correction. Most tools sold as AI-powered AP software stop a step short of that: they flag the mismatch and hand it to a queue, which is still manual resolution wearing a better front end.
Flagging an exception still leaves a human to resolve it
A system that spots a price variance or a missing receipt has done pattern matching. Nothing has been resolved yet. Someone still has to check the contract, call the supplier, or confirm the goods arrived before the invoice can move. Vendors count that flag as an automation win. AP teams count it as another line in the queue. Invoice exceptions only clear once the underlying question gets answered, and answering it is the part most "automated" tools skip.
An agent gathers the context and takes the action itself
Autonomous resolution means reading across the systems that hold the answer: the purchase order sitting in the ERP, the signed contract or price list, the receiving record, a supplier portal, or a past email thread nobody thought to check. An agent built for this reasons across that evidence the way an AP analyst would, then takes the next step itself: applying a tolerance, requesting a credit memo, correcting a GL code, releasing a payment hold. It logs the reason for the call it made, so the resolution holds up in an audit later. None of this requires a data migration or ripping out the ERP to get there; the agent works on top of what already exists.
Take a quantity mismatch where a shipment arrived in two partial deliveries. A flagging tool sees the receipt doesn't match the invoice quantity and stops. An agent checks whether a second receipt exists under a related PO line, confirms the split against the shipping documents, and closes the match without anyone opening the ERP.
Per-seat pricing rewards flagging over resolving
Most AP software is licensed per seat. A vendor whose revenue scales with the number of people touching invoices has a built-in reason to stop at detection: fewer touches means fewer seats renewed. Agentic software that actually closes exceptions cuts against that model, which is one reason genuinely autonomous resolution is rarer than the marketing around it suggests. Software billed by the seat is planning for people to keep doing the work.
Reserve people for judgment calls and automate the rest
Autonomous doesn't mean unsupervised. The agent handles the fact-finding and the mechanical steps; a person still decides the cases that require actual judgment: a disputed charge, a write-off above a set threshold, a policy exception with no precedent. That split is what manual vs. automated exception resolution should mean in practice: automation for anything with a knowable right answer, people for everything that doesn't.
Fragment builds agents that resolve invoice exceptions this way, reasoning across a company's existing ERP, MES, supplier portals, and inboxes without a migration first. See how the workflows work, or book a demo to watch one resolve a live exception.