AI Agents for Accounts Payable: Where Invoices Actually Break

AI agents for accounts payable pay off in the exception queue, not on the clean invoices.

A clean invoice already flows through most modern AP tools. The expensive ones are the 10 to 20 percent that stop: the price is off, the PO is missing, the same invoice arrived twice, the vendor changed its bank account.

This post sorts those exceptions into types. It shows which ones an agent can clear, which it should only prepare, and which must stay with a person.

Why the straight-through number misleads

Every AP vendor leads with a touchless or straight-through rate. Take Ardent Partners’ AP Metrics that Matter in 2025, a February 2025 survey of 212 AP professionals. It reports 32.6% of invoices processed touchless on average. It puts the average cost to process one invoice at $9.40. And it says invoice exception rates dropped to 14% in 2024.

Three problems with leaning on a number like that:

• Definitions differ. Does “touchless” mean no human opened it, or no human changed it? Does an invoice that a person only approved count? Vendor pages quote exception rates from the low teens to near 20 percent. I could not tie them to one definition, so treat any single figure as a rough guide.
• It counts invoices, not effort. A 30-minute dispute over a freight charge counts the same as a five-second tick.
• It ignores rework. An invoice that flows straight through and is later reversed looks like a win in the rate and a loss in the ledger.

My recommendation: ignore the headline rate when you evaluate an agent. Ask what it does with the invoices that fall out. Ardent lists the usual causes: coding errors, missing information, approval bottlenecks, and lack of PO data. That is where staff time goes, and where an agent can help.

The exception taxonomy

Before you automate anything, name the exception types. Pull a month of held invoices and tag each one. You will probably find six or seven buckets cover most of the pile.

Here is the table I’d start from.

Exception	What happened	Agent can clear?	Who owns the final call
Price or quantity mismatch	Invoice differs from the PO or the receipt	Within a written tolerance, yes. Beyond it, route	Buyer or PO owner
Missing PO	Invoice arrives with no PO reference	Propose a likely PO or cost center. Do not approve	Requester or budget owner
Missing receipt	Goods or services not confirmed received	Chase the receiver. Do not release	Receiver or department
Duplicate invoice	Same vendor, same amount, similar number	Hold it and explain why. Do not release	AP supervisor
Vendor bank or address change	A request to change where money goes	Never. Log and route	Someone outside invoice processing
Coding or GL error	Wrong account, cost center, or tax code	Suggest the fix from history	AP reviewer under a set limit
Tax, freight, currency, credit memo	Extra lines or odd amounts	Explain the gap, draft the note	AP reviewer
Not an invoice	Statement, quote, pro forma, remittance	Classify and file	None needed

The “agent can clear” column is a policy choice, not a technical limit. A model can read an invoice and compare it to a PO. Whether it may close the exception is for finance to decide, in writing, before the pilot.

AI agents for accounts payable: what an agent can clear

The safe work has two traits. It is reversible, and it follows a rule you can write down.

Match within tolerance. A three-way match compares quantities and price across the purchase order, the receiving document, and the invoice. The Washington State Auditor calls it a standard best practice. Most tools do this already. An agent adds value on the near misses: a unit-of-measure mismatch, a split shipment, a PO line that was closed one day early. Set a tolerance, say 2 percent or $25, whichever is lower. Inside it, the agent clears the line and records why. Outside it, the agent writes a one-paragraph summary and sends it to the buyer.

Suggest coding. Agents are good at “last six invoices from this vendor went to account 6140.” The agent proposes. A reviewer approves under a limit you set.

Find the missing piece. A missing PO number is often in the email thread, the PDF cover page, or the vendor’s earlier invoices. The agent looks, proposes a match, and shows the evidence. A person confirms.

Chase people. Half of AP delay is waiting on someone else. An agent can send the receiver a polite reminder with the invoice attached, and escalate on a schedule. This is unglamorous and very effective.

A labeled example

Say a 40-person AP team gets an invoice for 120 units against a PO for 100, with a receipt for 100. The agent sees a 20-unit overbill. It checks whether a second receipt is pending. None is. It drafts a short note to the vendor asking for a corrected invoice or a credit memo, and holds the invoice. A person reviews the draft and sends it. Nothing was paid, and nothing was approved by the agent.

That is the shape you want: the agent does the lookup and the paperwork. A person owns the money.

Controls that stay human

I treat an AP agent like a new AP clerk. It gets the same segregation of duties as a clerk, no more.

The Washington State Auditor’s guide is blunt about what AP staff should not do. Clerks should not add or edit vendors in the vendor master file. They should not perform or approve the check run. They should not reconcile the bank account. The supervisor should be independent of processing. If you give an agent those powers, you have collapsed the control you were trying to keep.

So the following stay with people:

1. Vendor creation and bank account changes. This is the fraud door. The FBI’s 2024 Internet Crime Report logged 21,442 business email compromise complaints and about $2.77 billion in losses. The 2025 AFP Payments Fraud and Control Survey found BEC was the top fraud vector, reported by 63% of respondents, and that vendor impostor fraud rose to 45% from 34%. A change request is exactly what an attacker sends. Verify it by calling a known number on file, done by someone who does not process invoices.
2. Payment release. The agent can prepare a batch. A person releases it.
3. Tolerance and policy settings. Finance sets them. The agent cannot edit its own limits.
4. Anything above the tolerance. Over-limit variances go to the PO owner.
5. Credit memos and write-offs. Money moves in the vendor’s favor or against it. A person signs.

This also matches broader guidance for agents. OWASP’s excessive agency entry says to require a human to approve high-impact actions before they run. Singapore’s agentic AI framework says to define checkpoints where human approval is required. An AP agent is a clean example of both.

Also remember invoices are untrusted input. An invoice PDF can carry text that tells a model what to do. Treat the content as data and never as instructions, and keep the agent’s tools narrow. Tool design matters a lot here: give the agent “propose a match,” not “post a payment.”

How to handle duplicate invoices

Duplicates deserve their own section because they are common and costly. The Washington State Auditor cites industry experts who put duplicate payments at 0.8 to 2 percent of total payments. That is an expert estimate, not a measured rate for your company. It still tells you the problem is real.

The same source explains why ERP checks miss them: most software only catches a repeated invoice number within the same vendor number. Duplicate vendor records defeat that. So does a changed invoice number.

An agent helps in three ways:

• Normalize first. Strip spaces, dashes, and leading zeros from invoice numbers. Map vendor name variants to one record. The auditor’s guide recommends standard naming conventions for the same reason.
• Compare fuzzy. Look at vendor, amount, date, PO, and line items together. Flag near matches, not just exact ones.
• Hold and explain. The agent places a hold and writes why: “Same vendor and amount as invoice 5521 paid on May 3. Different number, same PO.”

Then fix the cause, not only the symptom. The auditor’s list is practical: one intake channel for invoices, one payment method per vendor, limited rush payments, a clean vendor master, and a PO that closes when fully invoiced. A second invoice against a closed PO is a red flag by itself.

Auditability for the controller

A controller will ask one question: “Show me why this was paid.”

Every agent action on an invoice should leave a record you could hand an auditor:

• The invoice as received, with a timestamp
• What the agent extracted, and what it matched against
• The rule or tolerance it applied
• Its recommendation, and the human decision that followed
• The person who approved, with time and role

If an agent cleared a variance, the record should show the tolerance and the numbers. If a person overrode the agent, the record should say so. I cover what to capture in what to log for an AI agent audit trail. Log from day one of the pilot. Adding it later is much harder.

Measuring exception rate and rework

Pick the metrics before the pilot starts. Here are the ones I would track, with plain definitions:

Metric	Definition	Why it matters
Exception rate	Invoices needing a human action, divided by invoices received	Your baseline. Define it in writing
Agent resolution accuracy	Agent-proposed resolutions a reviewer accepts unchanged	Tells you if the agent is right
Override rate	Proposals a reviewer changes or rejects	Rising means drift or a bad rule
Rework rate	Invoices reopened, reversed, or re-coded after clearing	The number a straight-through rate hides
Exception cycle time	Hours from flagged to resolved	Where the speed gain shows up
Duplicates caught before payment	Holds that proved correct	Direct dollar value
False holds	Holds that turned out to be legitimate	The cost of being too strict

Compare against your own baseline, not an industry average. Your cost per invoice depends on your volume, your systems, and how you count staff time. For the wider ROI frame, see AI agents and back-office ROI.

Pilot scope: one week

Can AI automate accounts payable? In part, yes. But the pilot should prove the exception work, not the whole process. The scope I’d use matches the approach in our one-week pilot method and I’d run it in shadow mode.

Scope

• One legal entity or one vendor group
• 200 to 500 invoices from the last quarter, already resolved by your team
• Read-only access to the ERP, the PO and receipt data, and the invoice inbox
• No write access. The agent posts nothing.

Week plan

1. Day 1. Tag exception types from history. Write tolerances and the “agent can clear” column. Agree on metric definitions.
2. Days 2 and 3. Connect read-only data. Run the agent on the history and have it propose a resolution for each exception.
3. Day 4. Compare proposals to what your team actually did. Score accuracy, overrides, and false holds.
4. Day 5. Review misses with the controller. Decide what moves to live “draft only” mode and what stays manual.

Go or no-go

• Proposals match your team’s resolution at a rate you set in advance
• No proposal would have released a payment it should not have
• Every proposal has a complete case file

If it passes, go live in draft mode. The agent prepares, a person decides. Widen from there. For another finance use case with similar controls, see AI agents in financial services.

Bottom Line

Judge an AP agent by what it does with the invoices that stop, not by the touchless rate.

Let it match within tolerance, suggest coding, find the missing PO, chase people, and flag duplicates. Keep vendor changes, payment release, and policy settings with people.

Next step: tag one month of held invoices by exception type, then run a shadow-mode pilot on them.

FAQ

Can AI automate accounts payable? Partly. AI handles reading invoices, matching within tolerance, suggesting coding, and chasing approvals well. Payment release, vendor bank changes, and policy settings should stay with people. Start with the exception queue, where the effort is.

What is a realistic exception rate? It depends on how you define it. Ardent Partners’ February 2025 survey reported a 14% exception rate for 2024, and other published figures run higher. Measure your own baseline using a written definition, then track the change after the pilot.

Which controls must stay human? Vendor creation, bank account changes, payment release, tolerance settings, over-limit variances, and credit memos. Treat the agent like a new AP clerk with the same segregation of duties.

How do I handle duplicate invoices? Normalize invoice numbers and vendor names, compare vendor, amount, date, and PO together, and hold likely copies with a written reason. Also clean the vendor master and use one intake channel. A person decides whether to release a held invoice.

What does a one-week pilot cover? One entity or vendor group, 200 to 500 past invoices, read-only access, and shadow mode. Compare the agent’s proposed resolutions to what your team did, and score accuracy, overrides, and false holds before going live in draft mode.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect