The AP manager runs the SAP payment proposal on Thursday morning. Most invoices on it went through without anyone touching them. The week went on the rest: invoices on payment block, a vendor disputing a short payment, a non-PO invoice nobody would code, a bank detail change that needed a call-back. AI in accounts payable is useful for the rest, and that is also where it needs the firmest limits.
What AI in accounts payable does well now
Classification. Coding non-PO invoices to GL account and cost centre from the vendor's history, and tagging exceptions by likely cause rather than by SAP blocking reason alone.
Matching suggestions. Proposing the purchase order line and goods receipt for an invoice that arrived without a usable reference, and scoring near-duplicate invoices before the run.
Exception routing. Sending each exception to the person who can fix it, with the documents attached, and tracking its age. A late goods receipt goes to the warehouse; a price variance goes to the buyer.
Anomaly flags. Marking items in the payment proposal that look out of pattern: a first payment after a bank detail change, an amount well above the vendor's usual range, a vendor reactivated after a long gap.
Each is a suggestion with evidence that a person accepts or rejects, which gives an auditor a trail to follow.
How common it is, and the baseline to measure against
Surveys suggest AP is a common place to start. Gartner put AP process automation among the leading AI use cases, at 37%, in a May and June 2025 poll of 183 CFOs and senior finance leaders, while 91% reported low or moderate impact initially. Hackett's 2026 Finance Key Issues Study (sample not disclosed) found 33% of organizations were already scaling AI solutions in AP, ahead of other finance processes.
Know the baseline before measuring anything. APQC Open Standards Benchmarking (cross-industry median, all companies) shows a total AP cost of $6.50 per invoice processed, from 6,546 companies, and 15.0 days between invoice receipt and payment transmission, from 2,226. Sponsor-funded vendor research from Ardent Partners (State of ePayables 2025; the page gives no sample) reports averages of $9.84 per invoice, 8.2 days to process and an 18.4% exception rate. The two sources measure differently, so do not compare them directly.
What still needs people
| AP task | What AI can do now | What stays with a person |
|---|---|---|
| Non-PO invoice coding | Suggest account and cost centre from history | Approve where policy requires |
| Blocked invoices | Classify, suggest a cause, draft the vendor query | Release the block, accept a price, short-pay |
| Duplicate checks | Score near-matches before the run | Decide whether to hold |
| Vendor bank changes | Flag a change followed by a payment | Make the call-back, approve the change |
| Payment proposal | Flag anomalies | Approve the run |
| Vendor disputes | Summarise the history, draft a reply | Negotiate and agree |
| Policy exceptions | Spot that an exception applies | Grant or refuse it |
Approvals are delegated authority, and a spending limit belongs to a named person. Vendor disputes involve a relationship and often a negotiation. Policy exceptions, such as paying early, exist because someone judged that the rule should bend this time. Each needs a name against it.
A read-only layer on top of SAP, or changes to SAP
There are two broad routes. One is to change SAP: configuration, add-ons, and the AI features SAP and other vendors are building into their own products. The other is a layer that reads SAP data (invoices, orders, receipts, blocks, vendor change documents) and returns suggestions to people who then act in SAP.
| Read-only layer over SAP | Changing SAP | |
|---|---|---|
| Getting to a first test | One company code, no transport needed | Follows the release and testing cycle |
| Risk of a wrong posting | None from the layer itself | Needs the usual change controls |
| Where the record sits | Partly outside SAP; needs its own log | Inside SAP |
| Effect of an ERP migration | Data connection must be rebuilt | Configuration redone in the new system |
Neither is right in every case. A read-only layer suits a proof of concept, because it can be tested on one company code without touching configuration. It adds a system that needs access controls, logging and an owner. Changing SAP keeps controls in one place and is the better long-term home for anything that should post.
Whichever route you take, decide the access model first: what the tool can read, who sees its output, and how each suggestion and decision is logged. Earlier issues covered how AI changes the control environment and access, logging and sign-off for finance data.
What AI in accounts payable will not do
It will not fix the process underneath. If goods receipts are late, tolerances wrong and the vendor master full of duplicates, AI will report those problems faster and more clearly. Someone still has to fix them.
Nor will it satisfy an auditor alone. If a suggestion leads to a payment, the auditor will ask who accepted it, on what evidence and under what authority.
Frequently asked questions
What can AI do in accounts payable today?
Today AI can reliably classify invoices and exceptions, suggest matches, route exceptions to owners and flag anomalies in the payment proposal, all as suggestions a person reviews.
Can AI approve invoices in SAP?
It should not approve them. Approval is delegated authority that belongs to a named person, so AI should prepare the evidence and a person should approve.
Should an AI tool write back to SAP?
Not in a first project. A read-only tool is quicker to test and cannot make a wrong posting; write access can follow once controls, logging and ownership are in place.
How many companies use AI in accounts payable?
Published surveys suggest AP is one of the commonest uses. Gartner's 2025 survey put AP process automation at 37% among top AI use cases, and Hackett's 2026 study, sample not disclosed, reported 33% of organizations scaling AI in AP.
What does it cost to process an invoice?
APQC's cross-industry median total cost to process AP is $6.50 per invoice, across 6,546 companies. Ardent Partners' vendor research reports an average of $9.84, measured differently.
Where CFOLogic fits
Our AP work begins as a single-process proof of concept that reads from SAP or another ERP and never posts to it. A person signs off anything that goes outside the team. You can see the design in a purchase-to-pay demo on fictional data; in production it reads from SAP and drafts nudges that people choose whether to send.