
AI Agents for Three-Way Match Exceptions
The clean matches are already automated. The exceptions are where the time goes, and where reasoning, not rules, is what you need.
- Three-way matching compares the purchase order, the goods receipt, and the invoice. When all three agree, posting is routine. The exceptions are the workload.
- Rule-based matching passes the clean cases and dumps the rest in a queue. It cannot reason about a partial delivery, a price change, or a unit-of-measure mismatch.
- A governed agent reads all three documents, explains why they disagree, and proposes the resolution, whether that is accept within tolerance, split, or hold.
- The agent prepares the action as a draft and a person approves it, so exceptions clear faster without giving up control.
- ChatFin runs this on the ERP you already have, writing back only through the official API after approval.
In a healthy AP process most invoices never need a human. The purchase order, the goods receipt, and the invoice line up, and the bill posts. That part is a solved problem.
The trouble is the tail. A supplier ships part of the order. A price moved between the PO and the invoice. Someone received in cases and got billed in units. These are the exceptions, and they are where an AP team quietly loses its week.
Rules cannot fix this, because the answer depends on reading the situation. This page is about handing that reading to a governed agent that proposes the fix and waits for a person to approve it.

What actually breaks a match
A three-way match fails for a handful of familiar reasons, and each one needs judgment rather than a rule.
A partial delivery means the receipt is smaller than the order and the invoice. A price variance means the invoice unit price differs from the PO. A tax or freight line appears on the invoice that was never on the order. A unit-of-measure mismatch makes quantities that are actually equal look different.
None of these is a data error to reject. Each is a real business situation with a correct resolution, and figuring out which resolution applies is the work.
Partial and over delivery
The receipt does not equal the ordered or billed quantity. The agent checks what was received and proposes matching to the receipt or holding the balance.
Price and tax variance
The invoice price or an added tax or freight line differs from the PO. The agent quantifies the gap and checks it against tolerance.
Why rule-based matching leaves a queue
Native ERP matching and older automation are rule engines. They apply a tolerance, pass anything inside it, and route everything else to an exception queue for a human.
That is fine until the queue is the whole problem. A rule can tell you the invoice is 40 dollars over the PO. It cannot tell you that the 40 dollars is a freight charge the buyer agreed to by email, so the right action is to accept and code it.
So the exceptions pile up, and clearing them is manual reading, one at a time. The automation handled the cases that never needed help.
How a governed agent resolves the exception
An agent approaches the exception the way an experienced AP clerk does. It reads the purchase order, the receipt, and the invoice together, and it states, in plain language, why they do not agree.
Then it proposes a resolution. Accept within tolerance and code the difference. Match to the received quantity and hold the balance for the next delivery. Flag a genuine overbill back to the supplier. It shows its reasoning so the decision is easy to check.
Crucially, it does not post anything. It prepares the action as a draft for a person to approve.
Read all three documents
The agent pulls the PO, the goods receipt, and the invoice and lines them up field by field.
Explain the mismatch
It states the reason in plain language: short shipment, price change, added freight, unit mismatch.
Propose a resolution
Accept within tolerance, match to receipt and hold, or flag back to the supplier, with the reasoning shown.
Prepare for approval
It drafts the posting or hold action. A person approves, and it posts through the official API.
Governed, not autonomous
Speeding up exceptions is only valuable if you keep control of them. That is why the agent prepares and a person approves.
The reviewer sees the three documents, the explanation, and the proposed action in one place. Approving is a glance, not an investigation, because the reading has already been done.
The action then posts through the ERP's own API, so the same validations that apply to a manual entry apply here. Nothing is written to the database directly.
The clean matches never needed a person. Give the exceptions to an agent that reasons, and give the approval to a human.
ChatFinWhat changes for the AP team
The exception queue stops being a daily grind and becomes a review of prepared decisions. The same number of exceptions arrive, but each one comes with the reading already done and a resolution proposed.
That shortens the time from receiving an invoice to clearing it, which protects early-payment discounts and keeps suppliers current. It also frees the AP team from the part of the job nobody enjoys.
And because the agent explains itself, the team learns from the pattern rather than just clearing tickets.

How to start on your own ERP
This runs on the ERP you already have. The agent reads the PO, receipt, and invoice data through the feeds you expose, and writes the resolution back through the official API only after approval.
A sensible start is read-only for a couple of weeks: let the agent explain and propose while the team keeps posting. Once the proposals are trusted, turn on governed write-back with approval required.
From there the same connection and approval pattern extend to the rest of AP and beyond.
The one-line version
Your three-way match is already automated for the cases that never needed help. The value now is in the exceptions, and exceptions need reasoning rather than another rule.
Hand the reading to a governed agent and keep the approval with your team. The queue stops being a grind, the invoices clear faster, and nobody gives up control of what posts.
Frequently Asked Questions
What is a three-way match exception?
It is an invoice that does not agree with its purchase order and goods receipt, for a real reason such as a partial delivery, a price change, an added freight or tax line, or a unit-of-measure mismatch. The clean matches post automatically; the exceptions need judgment.
Why can rule-based matching not handle exceptions?
A rule engine applies a tolerance and routes anything outside it to a queue. It can measure the gap but cannot reason about the cause, so a human still reads each exception by hand. An agent reads the documents, explains the cause, and proposes the resolution.
Does the agent post the correction on its own?
No. It prepares the posting or hold as a draft with its reasoning shown. A person approves, and only then does it post through the ERP's official API, so every validation a manual entry faces still applies.
Give your three-way match exceptions to a governed agent.
ChatFin reads the PO, receipt, and invoice, explains the mismatch, and proposes the fix on the ERP you already run.
Every resolution is a draft a person approves, posted through the official API with a full audit trail.
Book a DemoWatch an agent resolve a real three-way match exception.
Bring a messy exception from your own AP. We connect through the interfaces your ERP exposes and show the agent explain it and prepare the fix for approval.