Accounts payable is the clearest example of work a person should not be doing by hand, and the clearest example of automation done badly.
The pitch you usually hear is "rip out your process, buy our platform, migrate everything, retrain your team, and in two quarters your invoices will be automated." For most small and mid-size businesses that is a bad trade. The project is bigger than the problem. So nothing happens, and a capable person keeps retyping invoices off PDFs into your accounting system, one at a time, forever.
There is a better path, and it does not start with new software. It starts with the work.
What AP actually is, step by step
Strip away the jargon and invoice processing is a short, predictable sequence:
- An invoice arrives, usually as a PDF or an email attachment.
- Someone reads it: vendor, invoice number, date, amounts, line items, tax.
- They decide which account and cost centre it belongs to.
- They match it against a purchase order or a contract, if there is one.
- They enter it into the accounting system or ERP.
- They route it for approval.
- Someone approves, and it gets scheduled for payment.
Look at that list honestly. Steps 2 and 5 are pure data movement, reading and re-keying, and a machine does them better than a tired human at four in the afternoon. Steps 3 and 4 are mostly rules with a few exceptions. Steps 6 and 7 are judgment and authority, and they should stay with people. The opportunity is not to automate "accounts payable." It is to automate the reading, the coding, and the entry, and to hand a clean draft to the human who already does the approving.
The approach that does not require a platform migration
Here is the shape of an AP automation that fits a real business instead of replacing it.
Read the document where it already arrives. Invoices come into an inbox or a shared folder. The automation watches that source. No new portal for your vendors to learn, no change to how invoices reach you.
Extract the fields with AI, then check them. Modern AI reads an invoice well. That is the easy part, and it is not where the value is. The value is in everything after: validating that the numbers add up, that the vendor is known, that the invoice is not a duplicate, and flagging anything that looks off.
Code it using your own rules. Map the vendor and line items to the right accounts using your chart of accounts and your historical patterns, not a generic template. The automation should learn how you code, because that is the part that actually saves your team time.
Draft the entry into your real system, with the source attached. This is the line that separates useful from useless. The result is not a summary in a chat window. It is a draft record in your actual accounting system or ERP, coded, with the original PDF attached so anyone can verify it later. Nothing is posted. A person approves.
Grade the confidence and surface only the exceptions. The automation flags the entries it is unsure about so your team reviews those, not all of them. For one North American back-office services firm, a confidence grader built into the flow ran their monthly close at about 95% match to the prior manual process, which meant the team only had to look closely at the small slice that did not line up. That is one firm's result, not a typical or guaranteed outcome, but it is the pattern: humans on the exceptions, the machine on the volume.
Notice what is not in that list. No multi-quarter implementation. No vendor portal rollout. No ripping out the system you already run. You are connecting the tools you have through their interfaces, removing the re-keying, and keeping a human on the approval. The systems stay. The typing goes away.
What about trust, security, and "the AI making things up"
Two fair objections, both answerable.
On accuracy: a well-built AP automation never posts on its own, grades its own confidence, and attaches the source document to every draft so every number is traceable back to where it came from. If an entry is wrong, a person catches it at approval, the same checkpoint you already have, now with less typing in front of it.
On security: this runs inside your own controls. Your identity system, your access rules, your data staying where it already lives. If an automation needs to ship your financial data to an outside platform you have not vetted, your IT team should stop it. Real automation is built to pass that review, not to dodge it.
Where to start
You do not need to automate all of accounts payable on day one. Start with the highest-volume, most repetitive slice: the invoices that arrive the same way, from the same kinds of vendors, coded the same way every month. Automate the reading, coding, and drafting for those. Keep approval with your people. Measure the hours you get back. Then expand.
The reason to do it this way is simple. A small, well-scoped automation that ships in weeks and earns trust beats a big platform project that arrives in two quarters and asks your team to change everything at once. Operations runs on momentum. So should your automation.
If you want help finding which slice to start with, that is the first thing the AI Automation Audit does: it maps your actual workflows and tells you what to automate first, with the hours and dollars attached, before you commit to building anything.