
A PO invoice matching agent reads each supplier invoice, finds the purchase order and goods receipt it belongs to, and checks price and quantity against your tolerances. Inside tolerance, it drafts the entry. Outside, it stops and hands the invoice to a named person with the evidence attached. It never pays, and it never changes who gets paid.
In one sentence: let the agent find the match, let your rules decide what a match is, and keep the payment with a person.
What PO invoice matching is
Matching is old. ERP systems have done it for decades. Microsoft's Dynamics 365 documentation describes the two common types1:
- Two-way matching compares the price on the invoice with the price on the purchase order.
- Three-way matching does that and also compares the invoiced quantity with the quantity on the goods receipt. You only pay for what arrived.
Differences are checked against tolerances you set. Microsoft's example: a PO for 1,000 batteries at 1.00 each, a 5% price tolerance, and an invoice at 1.10. A price of 1.05 would pass. 1.10 does not, so the invoice waits until someone resolves the difference1.
The rule engine was never the hard part. The hard part is everything before it: an invoice that arrives as a scanned PDF, quotes no PO number, splits one delivery across two lines, or bills three receipts at once. That is where AP time goes, and that is the part an agent is for.
What the agent adds
The agent's job is finding the match, not deciding what a match is. Microsoft's own Payables Agent for Business Central shows the split clearly. It "uses AI to intelligently find matching orders and order lines that aren't identified by a unique reference on the vendor's invoice"2. Then it proposes PO lines, and Business Central itself validates them, allocates quantities and raises the order-match warnings3. A supervisor reviews the draft before it becomes a purchase invoice3.
That is the right shape. The agent proposes. Your ERP's matching rules check. A person decides the hard cases. In practice, a good matching agent does five things:
- /01
Reads the invoice in any layout: PDF, scan or email body, without a template per supplier.
- /02
Finds the vendor and the PO, even when the invoice gives no PO number.
- /03
Matches line by line: price against the PO, quantity against the receipt.
- /04
Explains each variance: the PO price, the invoice price, the tolerance it broke, with the documents attached.
- /05
Drafts the entry, unposted, ready for your normal approval workflow.
Where it must stop
Six kinds of invoice always go to a person. None of them should depend on how confident the model feels.
- A variance outside tolerance. The agent explains it. A person accepts it, disputes it or asks for a credit note.
- No PO, or no receipt yet. Matching against nothing is guessing.
- A possible duplicate. Same supplier, same amount, a slightly different invoice number. These are cheap to catch and expensive to pay twice.
- A new vendor. Business Central blocks a vendor the agent created until a person reviews and unblocks it3. Copy that rule whatever system you run.
- Changed bank details. Never the agent's call. Business email compromise cost US businesses $2.77 billion in reported losses in 20244, and new bank details are how it gets paid.
- Anything over a money limit. Set it per invoice and per day. How to set AI agent spend limits covers how to pick the numbers.
Tolerances are your number, not the agent's
Vendors sell matching agents on their touchless rate: the share of invoices that go through with no person involved. It is a fair number to track. It is also an incentive. The fastest way to raise it is to widen the tolerance.
So the tolerance cannot live where the agent can change it. Keep it in your ERP or in signed rules outside the agent. Change it on purpose, with a name and a date on the record. If a vendor's agent "learns" its own tolerances, ask who approved the last change.
Text in an invoice is data, not instructions
An agent that reads invoices reads everything on them, including text a stranger put there to steer it. A line in small print can ask it to update the remittance address or mark the invoice urgent. The fix is not a smarter filter. It is a rule the invoice cannot change: the agent may match and draft, and nothing it reads can widen that. Prompt injection in accounts payable has a one-hour test you can run on your own agent.
Microsoft's setup advice for its own agent points the same way. It recommends an internal shared mailbox that is not exposed to vendors, so your employees are the first to review incoming invoices and spot possible fraud3.
What to ask a vendor before you buy one
- /01
Show me a held invoice. What does the approver see, and is the evidence attached?
- /02
Where do tolerances live, and who can change them? The answer should be a person, on the record.
- /03
Can the agent post or release a payment? The answer you want is no.
- /04
What happens with a new vendor or changed bank details? Both should stop, every time.
- /05
How are duplicates caught? Ask for the rule, not the accuracy rate.
- /06
Can I export every match, hold and decision, with the name of who decided? That export is your audit trail.
If you already run RPA in AP, keep it for the fixed steps and put the agent on the exceptions. RPA vs AI agents for accounts payable shows how to split the work.
Gatehouse fit
Gatehouse, Surehand's control plane, sits in front of a matching agent. The agent's scope is a signed manifest: read invoices, POs and receipts, match, draft. Never post a payment. Tolerances and spend limits live in the rules, not in the agent. Variances, duplicates, new vendors and changed bank details hold for a named approver, and every match and hold lands in a sealed record. The composite invoice matching deployment shows one AP run end to end.
At a glance
| Category | Operations |
|---|---|
| Also asked as | AI invoice matching agent, three-way match agent, AP matching agent |
| What it does | Finds the PO and receipt, matches price and quantity, drafts the entry |
| Where the rules live | Your ERP and signed rules, outside the agent |
| Always held for a person | Out-of-tolerance variance, no PO or receipt, duplicate, new vendor, bank change, over limit |
| Never the agent's call | Posting, paying, creating vendors, changing bank details |
| The one test | Can anything written on an invoice change who gets paid? |
Common questions
What is a PO invoice matching agent?
An AI agent that reads each supplier invoice, finds the purchase order and goods receipt it belongs to, and checks price and quantity against your tolerances. Inside tolerance it drafts the entry for posting. Outside tolerance, or when something is missing, it holds the invoice for a named person with the evidence attached.
What is the difference between two-way and three-way invoice matching?
Two-way matching checks the invoice price against the purchase order. Three-way matching also checks the invoiced quantity against what was actually received, from the goods receipt. Three-way catches being billed for goods that never arrived, so it is the usual choice for physical goods.
Can an AI agent approve or pay invoices on its own?
It should not. Let it match, explain the variance and draft the entry. Posting, releasing payment, creating vendors and changing bank details stay with people your policy names, so the agent prepares and someone else approves.
Sources
- [1]Microsoft Learn, Accounts payable invoice matching overview (Dynamics 365 Finance)learn.microsoft.com In text
- [2]Microsoft Learn, Match purchase invoices to orders with Payables Agent (Business Central release plan, 2025 wave 2)learn.microsoft.com In text
- [3]Microsoft Learn, Payables Agent overview in Business Centrallearn.microsoft.com In text
- [4]FBI Internet Crime Complaint Center, 2024 IC3 Reportic3.gov In text
Read next

RPA vs AI agents for accounts payable
In accounts payable, RPA handles the steps that never change and an agent handles the invoices that do not fit. Which part of the AP process each one should own, and the controls an agent needs before it touches a payment.

Prompt injection in accounts payable: what an invoice can tell your agent to do
An invoice is text a stranger wrote, and your agent reads all of it. Prompt injection turns that text into instructions. You can't filter it out. You can limit what it can do.

What is an AI audit trail?
Seven fields. And why logs don't count. What your risk team will ask for, and what most systems can't produce.