Automate Accounts Payable with AI: An SME Playbook

Manual accounts payable is expensive mostly because of waiting. Invoices sit in a shared mailbox until somebody opens them. Approvals get chased over chat. Duplicate payments surface weeks later, early-payment discounts pass unnoticed, and month-end absorbs all of it because the numbers eventually balance.

Work out what one invoice costs you to process before you do anything else. Take the fully loaded cost of the people who touch supplier invoices, add the licence cost of the tools they use for the job, and divide by the number of invoices posted in the same period. The result will be rough. It is still good enough to decide with, and without it the automation decision gets argued on instinct rather than arithmetic.

A second cost never reaches the ledger. Unsanctioned tool use is worth checking for before you design anything, because reading a document is exactly the kind of task a person will hand to whatever general-purpose AI tool they already have open.

That figure covers the population as a whole rather than finance departments specifically, so it proves nothing about your AP team on its own. It does set the base rate. Ask whether anyone has already pasted an invoice into a chatbot and retyped the answer into the ERP, and treat the answer as a design input. A sanctioned pipeline with a confidence threshold, a review queue and an audit trail is the governable version of that habit.

Automating AP starts with writing the current cycle down honestly, including the parts nobody documents. Written out, the cycle in a small or mid-sized company runs to nine steps.

1. Arrival. The invoice lands as a PDF attachment, a scan, an e-invoice, or a link in a supplier portal that someone has to log into.
2. Capture. Header and line data are read: supplier, invoice number, date, due date, net, VAT, gross, currency, bank account, PO reference.
3. Supplier identification. The document is tied to a supplier record, which is harder than it sounds when the trading name on the invoice differs from the legal name in the ledger.
4. Validation. VAT arithmetic, currency, duplicate check against invoices already posted, registration number, and whether the bank account matches the one on file.
5. Matching. Two-way against the purchase order, three-way if there is a goods receipt.
6. Coding. General ledger account, cost centre, project, VAT treatment.
7. Approval. Routed by amount, cost centre and category to a named approver.
8. Posting and payment. Written into the ERP as a purchase invoice, then included in a payment run.
9. Archiving. The document stored where an auditor can find it, linked to the ledger entry.

Write down who owns each step today and how long each one waits. Cycle time accumulates in the waiting far more than in the doing. Time capture and approval separately, because capture runs in minutes per invoice and approval in days. Automation that speeds up capture while leaving approval untouched will barely move the number a CFO looks at.

AI belongs in exactly one part of accounts payable: turning an unstructured document into structured fields, a job we take apart field by field in our guide to AI supplier invoice processing. Everything downstream of that should be deterministic code, because finance work has to be reproducible. The same invoice fed in twice must produce the same posting twice, and a language model is the wrong tool for any step where that guarantee matters.

We hold ourselves to this rule in our own products. The free AI-readiness check on our homepage runs eight deterministic checks with no LLM scoring anywhere in the pipeline, specifically so that running it twice on the same site gives the same answer. The same reasoning applies to a purchase ledger, with more at stake.

The practical consequence is that an AP automation project is mostly integration and rules engineering with a model bolted onto the front. If a vendor describes the whole thing as AI, that is a sign the deterministic half has not been thought through.

It will read some invoices wrong. The design has to assume that from the first day. The answer we give when finance buyers ask is always the same: confident cases post automatically, uncertain ones go to a review queue with the document and the system's best guess side by side, and a person confirms or corrects in seconds rather than retyping from scratch. That queue is the part most projects underbuild, and designing a review queue that stays fast and shrinks over time is a discipline that applies well beyond accounts payable.

Every extracted field therefore carries a confidence value, and the threshold is a setting the client controls rather than something buried in the code. Start it high and let it come down on evidence. A pipeline that sends most invoices to review in its first month and earns its way down to a small exception queue has worked. A pipeline that posts everything from day one and produces a reconciliation problem in month two can cost far more to unwind than the cautious version costs to sit through.

Corrections have to be worth making. When a reviewer fixes a field, the fix should be stored as a supplier-level rule, so that the next invoice from the same supplier arrives already knowing where the invoice number sits and which cost centre it usually belongs to. Without that feedback path, the review queue never shrinks and the team correctly concludes the automation is not helping.

The requirement worth writing into the specification is narrower than automating AP. It is this: cut manual data entry, and keep exception handling and manual review wherever a field cannot be identified reliably. That sentence also works as a test of a supplier. If it makes them uncomfortable, they are selling a demo rather than a production system.

Three metrics decide whether AP automation paid for itself, and all three need a baseline measured before the build starts. Retrofitting a baseline afterwards is guesswork, and everyone in the room knows it.

Use the median for cycle time, not the average. One invoice stuck behind a director on holiday for six weeks will drag an average far enough to make a real improvement invisible.

We apply the same discipline to our own operations and publish the parts that do not flatter us. On aiasemu.ee, the lawn-care business we run to test these methods, PostHog recorded 57 address searches, 39 refined estimates, 9 bookings started and 2 bookings completed in one week. Our internal conclusion was that the site generates leads well and closes bookings badly. That was only visible because every step in the funnel was counted separately. An AP pipeline deserves the same treatment: instrument every step from arrival to payment.

Roll out in four phases, and do not let anything post to the live ledger until the automation has agreed with a human on real invoices for a few weeks.

Phase 0: measure the baseline

Export a year of purchase invoices. Count volume, compute median cycle time, work out cost per invoice, and rank your suppliers by invoice count. Read down that ranking until the tail turns into one-off invoices. Everything above the cut is what the first build should handle.

Phase 1: shadow mode, a one to three week build

The automation reads incoming invoices and produces a proposed posting, but a person still enters it. You compare the proposal against what the person did. This is where extraction accuracy per supplier becomes visible, and where the confidence threshold gets set with evidence instead of a guess.

Phase 2: auto-post below the threshold

Confident invoices from known suppliers post automatically. Everything else goes to the review queue. Touchless rate becomes a number you can watch weekly.

Phase 3: extend outward

Once posting is stable, the adjacent workflows have the same shape and reuse the same integration. The set buyers ask for next is fairly consistent: read bank payment notification emails and reconcile them against open invoices, sync sales orders from the ERP to Excel and OneDrive, generate delivery planning from sales-order delivery dates, automate monthly file creation and archiving. Each of those is its own one to three week workflow rather than a second project from scratch.

The failure mode to avoid is starting with the hardest supplier. Begin with the ones that send the most invoices in the most consistent format, and let the awkward ones stay manual for as long as they need to. Our wider view of this is in our guide to AI back-office automation for SMEs.

We quote AP automation per project, because the price is driven by the number of systems that have to be integrated rather than by invoice volume, which is the same logic behind the ranges in what an AI business automation project costs. Scope is the planning figure we can state up front: one well-defined workflow, such as supplier invoice PDFs extracted and posted as purchase invoices in the ERP, is typically a one to three week build. Our adjacent work carries published prices on aiside.ee, 290€ for an AI-readiness audit and from 3 999€ for a new site, and automation quotes follow the same principle of naming a number rather than hiding behind a contact form.

Two commercial terms are worth settling before any of that starts. The first is ownership. Source code, infrastructure and rights stay with the client, and the build is not tied to a single AI vendor, so a model can be swapped when a cheaper or better one appears. The second is your ERP, which you almost certainly keep. Cloud systems such as Exact Online, Merit and e-conomic expose APIs, so the automation is built around what you already run.

One last point, this time about the buying side. When somebody asks an assistant for automation suppliers rather than opening a search engine, the shortlist gets assembled from whatever those systems can read about a company. We cannot measure how often that happens and we will not pretend otherwise. The mechanics are worth understanding either way, and they are the subject of why ChatGPT does not recommend your business.