- Capture, matching and coding suggestions are the parts AI handles well. They are also where most AP hours go.
- Exceptions, new vendors, bank detail changes and anything over an approval limit stay with a person.
- Measure accuracy per field and per vendor, not as one overall number.
- Check what your accounting system and AP tool already do before you build anything.
- Custom work earns its place in the gaps: your matching rules, your exception queue, your approvals.
The 28th of the month
It's the 28th. The AP clerk has a shared inbox with a few hundred unread messages. Some are invoices. Some are statements. A few are the same invoice sent twice, once as a PDF and once as a photo from a driver's phone.
Each one gets opened, read and keyed into the accounting system. Then it gets matched to a purchase order, if there is one. Then someone has to confirm the goods arrived. Then a manager approves it, usually after two reminders.
None of this is hard. It's just slow, repetitive and easy to get slightly wrong. That combination is exactly where AI earns its keep, with a person still in charge of the money.
The five steps, and who should own each
Accounts payable breaks into five steps. AI helps with some far more than others.
- Capture. Reading the vendor, invoice number, dates, totals, tax and line items from whatever arrived. Modern AI reads varied layouts well, including scans and photos. This is the clearest win.
- Matching. Comparing the invoice to the purchase order and the receiving record. A two-way match checks invoice against PO. A three-way match adds proof the goods or services arrived. Software does the comparison. Your team sets the tolerances.
- Coding. Assigning the general ledger account, department, project or job. AI can suggest coding from past invoices for the same vendor. A person confirms anything unusual.
- Exceptions. Price differences, short shipments, missing POs, duplicate invoices. This is judgment work. AI can sort and explain the problem. A person resolves it.
- Approval and payment. Always a person, with limits set by your finance team. Automation can route the request and chase the approver. It should never approve on its own.
The pattern is simple. Let software do the reading, comparing and routing. Keep people on the decisions that move money.
Capture: better than it used to be
Older invoice scanning relied on templates. Someone drew boxes around where the total sat on each vendor's layout. When a vendor redesigned its invoice, the template broke.
Current tools read invoices without a template for each vendor. Microsoft, Google and Amazon all offer prebuilt invoice models that return fields and line items, each with a confidence score. Our guide to intelligent document processing explains how that reading works in plain terms.
The confidence score matters more than the headline accuracy. It lets you send a clean, high-confidence invoice straight to matching. A blurry photo with a low score on the total goes to a person first.
Matching: where the real time goes
Capture gets the attention. Matching is often where the hours actually disappear.
The invoice says 40 units at one price. The PO says 40 at a slightly different price. Receiving logged 38. Is that a problem? It depends on your tolerances, the vendor and who ordered it.
Most accounting systems support matching rules. Dynamics 365, NetSuite and Sage Intacct all handle some form of PO matching. Check what yours does before you build anything.
Where custom work helps is the logic your system doesn't express. Partial shipments that arrive over three weeks. Freight billed separately. A vendor who always invoices the whole order on day one. These rules live in your AP lead's head today. Writing them down, then putting them in software, is most of the value.
Coding: suggestions, not decisions
A vendor you've paid 200 times is easy to code. The software can see how the last invoices were coded and suggest the same.
A new vendor, or a familiar vendor billing for something new, is different. The suggestion might be plausible and wrong. Show the reviewer why the system chose that account, so they can disagree quickly.
Over time, track how often reviewers change the suggested coding. If they almost never change it for a vendor, you can loosen review there. If they often do, the suggestion isn't helping yet.
What stays with a person
Some decisions should stay human no matter how good the software gets.
- Changes to vendor bank details. The FBI's guidance on business email compromise is to confirm any change to payment instructions through a separate channel, such as a call to a number you already have. No AI should process a "please update our bank account" email on its own.
- New vendors. Someone should confirm the vendor is real and approved before the first payment.
- Anything outside tolerance. Price or quantity differences beyond your limits, or invoices with no matching PO.
- Approvals above a threshold. Your finance team sets the limits. Software enforces them.
- Disputes. Talking to a vendor about a short shipment is a relationship, not a data task.
Our guide to designing human review for AI work covers how to build review screens people actually use, so they don't just click approve.
Statements, credits and the messy middle
Not everything in the AP inbox is an invoice. Vendors send statements, credit memos, past-due reminders, price change notices and plain questions. A good system sorts these before it reads anything.
Statements are worth handling well. Comparing a vendor's statement to your own open items catches invoices that never arrived, and payments the vendor hasn't applied. It's tedious work for a person and a natural job for software, with a person reviewing the differences.
Credit memos need care. They should reduce what you owe on a specific invoice, not sit unused in the system. Software can suggest which invoice a credit belongs to. Someone on the team should confirm it.
Then there are the questions. "Can you confirm you received our invoice?" "When will we be paid?" AI can draft a reply from the invoice status in your system. A person sends it. That alone can clear a surprising amount of inbox.
Measuring accuracy honestly
A vendor will quote you one accuracy number. It won't tell you much.
Measure per field. Reading the invoice number correctly matters less than reading the total and the bank details correctly. Measure per vendor too. A system can be excellent on your top twenty vendors and poor on the long tail of small ones.
Run a test on your own invoices before you commit. Take a few hundred from last quarter, including the ugly ones. Compare what the system read against what your team keyed. That tells you more than any demo.
Then keep measuring after launch. Track how often reviewers correct each field, how many invoices flow through without a touch, and how long exceptions sit in the queue.
Your AP tool, or something custom
Start with what you already pay for. Many accounting systems now include invoice capture. Dedicated AP tools add approval workflows, vendor portals and payment runs. Procurement platforms like Coupa handle purchasing, invoicing and approvals in one place.
An off-the-shelf tool fits when your process looks like most companies' processes. Standard POs. Clear approval chains. One accounting system.
Custom work fits the gaps around those tools:
- Matching rules your AP tool can't express.
- Receiving data that lives somewhere your AP tool can't see, like a warehouse system or a field app.
- Several entities or accounting systems after an acquisition.
- An exception queue that shows the clerk the invoice, the PO, the receipt and the suggested fix on one screen.
Often the right answer is both. Keep the AP tool for what it does well. Build the connection or the queue it leaves out. Our invoice processing page lists more of what that can look like, and we connect to QuickBooks, NetSuite, Sage Intacct and Coupa.
Rolling it out without a big bang
Don't switch every vendor over on the same day. Start with a group of high-volume vendors whose invoices are clean and whose POs are reliable.
Run the new process alongside the old one for a few weeks. The clerk still checks everything, but now against what the system read and matched. Disagreements show you where the rules need work.
Then widen the group, and loosen review where the numbers support it. A vendor whose invoices have gone through cleanly for two months can skip the routine check. A vendor who keeps changing formats stays under a closer eye.
Tell your vendors what's changing, too. A single address for invoices, PDFs rather than photos where possible and a PO number on every invoice will lift your results more than any model upgrade.
What most guides miss
Your PO data is the bottleneck. Matching only works if the POs are accurate and receiving gets logged. If the warehouse logs receipts a week late, every invoice looks like an exception. Fix that first, or at least know it going in.
The exception queue is the product. Clean invoices take care of themselves. Your team's day will be spent in the exceptions. Design that screen with the people who'll use it.
Duplicates hide in plain sight. The same invoice arrives by email, by post and through a vendor portal, each with a slightly different number format. Checking vendor, amount and date together catches more than checking the invoice number alone.
Month-end changes the rules. Volume spikes in the last days of the month. Plan review capacity for the 28th, not for an average Tuesday.
How we do it
We start with a free intro call and a look at a sample of your real invoices, including the messy ones. Then we work alongside your AP team to build the piece your current tools leave out, usually matching rules or the exception queue. Every payment decision stays with your people, and we stay on after launch. Start with an intro call.
