Quick answer
AI in accounts payable can read invoice fields, suggest GL codes from past coding, and flag invoices that look unusual. It doesn’t know whether goods arrived, whether a vendor is legitimate or whether past coding was right, and it can be wrong with high confidence. CPA firms can check a client’s AP software with a 50-invoice test that scores field accuracy, coding agreement by confidence level, and flag quality.
Much of the writing on AI in accounts payable covers what it can do. Your clients will ask whether they can trust it. Vendors quote accuracy figures that are hard to compare, because each one comes from different invoices.
A better question is how the tool performs on this client’s invoices. A 50-invoice test answers it in an afternoon. Pull a sample, score the AI’s output against a person’s, and check whether its confidence scores mean anything.
Free worksheet
Get the AP AI test worksheet
Score a client’s AP software on 50 invoices: capture accuracy by field, GL coding by confidence band, and flags on planted test invoices, with the range beside every rate. Use it on any client.
What can AI do in accounts payable today?
AI does three useful things in AP today. Each one produces a prediction that needs checking.
Invoice capture
The software reads an invoice from a PDF, scan or email and fills in the fields: vendor, invoice number, date, totals, tax and line items. Many current tools work without a template for each vendor and learn from corrections. Capture works best on clean invoices from repeat vendors, where the layout is familiar. ProcureDesk’s AP automation page describes capture from email or a supplier portal across PDFs, scanned images and electronic invoices.
GL coding suggestions
The software proposes an account, department or class for each invoice by comparing it with how similar invoices were coded before. A good tool shows a confidence score with each suggestion, so a person can accept high-confidence coding and review the rest. The suggestion is only as good as the history behind it.
Flagging odd invoices
The software compares a new invoice with past ones and flags what looks off: a possible duplicate, a price above the usual, an unusual amount, or a pattern that doesn’t fit the vendor. In AFP’s 2026 survey, only 17% of organizations used AI to fight payments fraud. Those that did reported better efficiency in fraud reporting (49%), deepfake detection (45%) and real-time identification (43%). AFP names cost, the perceived immaturity of the technology and reliance on existing controls as the reasons organizations hold back.
Where does AI fall short in AP?
AI fails in predictable places. The table lists them by task.
| Task | Where it falls short | What the CPA checks |
|---|---|---|
| Invoice capture | Poor scans, handwriting, unusual layouts, long multi-page invoices, credit memos, and totals read as the wrong number | Compare captured totals, tax and invoice numbers with the source image on a sample |
| GL coding | New vendors with no history, one-off purchases, items where context decides the account (such as repair versus capital), and mis-coding it learned from the past | Compare suggested accounts with a person’s coding, grouped by confidence score |
| Flagging odd invoices | False alarms that bury real problems, and fakes that look normal, such as a real vendor’s name with changed bank details | Count flags that were real and problems that were missed, using planted test invoices |
Three limits cut across all of them.
It learns from your history, errors included. If a client has coded repairs to the wrong account for two years, a model trained on that history will keep doing it, consistently. Review a sample of coding against the client’s chart of accounts policy and not only against past entries.
Confidence is the model’s own estimate. A score of 95% doesn’t mean the suggestion is right 95% of the time. The test in the next section checks whether higher scores really go with higher accuracy.
It sees only what is on the page and in the data. AI can’t know whether the goods arrived, whether a vendor is who it claims to be or whether a bank-detail change is genuine. Those checks need receipts, vendor approval and a call to a number already on file.
How do you test a client’s AP AI on 50 invoices?
The test takes an afternoon and needs the client’s AP software in review or test mode. The AP AI test worksheet scores the results. Don’t post or pay the planted invoices in step 5.
Step 1: Build the sample
Pick 50 paid invoices: about 30 from regular vendors, 10 from new or rarely used vendors, and 10 unusual ones. Unusual means poor scans, long multi-line invoices, credit memos or foreign currency. The mix matters because the AI will look best on regular invoices and worst on the rest.
Output: a list of 50 invoice IDs, each tagged Regular, New vendor or Unusual.
Step 2: Run them and save the AI’s output
Run the sample through the software. For each invoice, save the captured fields, the suggested GL code, the confidence score and any flag.
Output: one row per invoice with the AI’s results.
Step 3: Score capture
Compare the AI’s vendor, invoice number, date, total, tax and line items with the source image, and mark each one right or wrong. Report accuracy for each field, and also the share of invoices where every field was right. That second figure shows how many invoices could pass through untouched. Field accuracy can look high while many invoices still carry one error.
Output: accuracy by field and an all-fields-right rate.
Step 4: Score coding by confidence band
Have someone who knows the client’s chart of accounts code the same invoices without seeing the AI’s suggestion. Enter the AI’s code, the correct code and the AI’s confidence score. The worksheet groups the results into High, Medium and Low bands. Look for two things: whether accuracy rises with the score, and how coding does on new vendors, which have no history to learn from.
Output: coding accuracy by band and by invoice group.
Step 5: Plant problems and count the flags
Add about six test invoices. Good ones to plant: a copy of an earlier invoice with a different invoice number, an invoice above its PO amount, an invoice from a vendor with changed bank details, and an invoice with no PO. Record whether the AI flagged each one, and record any flags on clean invoices. Expect some misses. The aim is to learn which ones.
Output: false alarms, missed problems and the share of real problems caught.
Step 6: Read the result with its margin
A small sample gives a rough read. If the AI codes 48 of 50 invoices correctly, that is 96%, but the true rate could plausibly sit anywhere from about 87% to 99%. The worksheet shows this range beside every rate. Treat a band as proven only when it holds enough invoices. Rerun the test after the client adds many new vendors or changes its chart of accounts.
Output: a one-page result for the client, with the range beside each rate.
What should a client automate, sample or keep human?
Use the test results to set one of three levels for each kind of invoice.
| Level | Use it for | Review rule |
|---|---|---|
| Suggest only | New vendors, the first period after go-live, and any confidence band below the client’s target | A person accepts or fixes every suggestion |
| Auto-accept with sampling | Repeat vendors in a band that met the client’s target in the test | A person reviews a random sample each month, and the band is retested if errors rise |
| Touchless | Invoices that match an approved PO and receipt within tolerance, from vendors in good standing | A person reviews exceptions and a monthly sample |
Some items stay with a person whatever the score:
- The first invoice from a new vendor
- Any bank-detail change
- Invoices above the approval limit
- Capital versus expense decisions
- Credit memos and adjustments
- Invoices with no PO that are not on the exempt list
Write the rule down, set a monthly sample size the client can keep up with, and name who reviews it. When accuracy drops or the vendor mix changes, move that band back one level. The software doesn’t carry accountability for the books, so a named person at the client or your firm still does.
Where does ProcureDesk fit in?
ProcureDesk is one of the tools a client might be using or considering. The same test applies to it and to any alternative.
ProcureDesk’s AI features map onto the three tasks. According to its AI capabilities page, GL coding predictions come with a confidence score for every suggestion and learn from each correction. Duplicate detection catches differences in amount, date and invoice format. Approvers see comparative pricing, budget impact and spending alerts next to each request. Capture takes invoices by email or supplier portal and handles PDFs, scanned images and electronic invoices. Its AP software page describes validation against POs and vendor master data.
Confidence scores matter only if they line up with accuracy, which step 4 of the test checks. ProcureDesk’s pages also publish accuracy figures for capture and coding. Treat those figures, and any other vendor’s, as claims to test on the client’s own invoices.
ProcureDesk integrates with QuickBooks Online, QuickBooks Enterprise, Xero, Sage Intacct, NetSuite, Microsoft Business Central and Bill.com, and implementation takes 2 to 3 weeks. Firms can recommend it to clients or offer it through the CPA partner program.
See it working
See predictive GL coding and duplicate detection in a demo
What do you tell a client who asks if AI will replace the AP team?
Tell them the work changes before the headcount does. Keying and basic matching shrink. Review, exceptions, vendor setup and judgment calls grow, and someone has to own each one. Three questions help the client see it:
- How many hours a week go to typing invoice data, and how many to resolving problems?
- Who reviews the AI’s output today, and how would anyone know if it got worse?
- Which invoices would the client never want paid without a person looking?
The answers set the review rules above and give your firm recurring work: the first test, a monthly sample review, and a retest when vendors or accounts change. That’s advisory work built on evidence, and it avoids promising savings the test hasn’t shown.
We have heard this on sales calls. In May 2026, a finance team evaluating ProcureDesk told us their leadership wanted to avoid hiring dedicated AP staff by using automation and AI to handle invoice processing and approvals.
Free worksheet
Download the worksheet and run the test on your next client
Scores capture, coding by confidence band and flags, with the range beside every rate.
Frequently asked questions
What can AI do in accounts payable today?
It can read invoice fields, suggest GL codes from past coding, flag invoices that look unusual and help match invoices to POs. Each output is a prediction that needs checking.
How accurate is AI invoice capture?
It depends on the invoices. Accuracy tends to be higher on clean invoices from repeat vendors and lower on poor scans, long multi-line invoices and new vendors. Vendors’ figures come from different sets of invoices, so test on the client’s own.
Can AI code invoices to the GL on its own?
It can suggest codes, and for high-confidence cases it can auto-accept them, but only after a test shows those suggestions are right often enough for that client. It learns from past coding, including past errors.
What is a confidence score?
It’s the model’s own estimate of how likely its suggestion is to be right. The score is useful only when higher scores go with higher accuracy, which the test checks.
Can AI detect invoice fraud?
It can flag duplicates, unusual amounts and odd patterns. It can miss fakes that look normal, such as a real vendor’s name with changed bank details, so call-backs and PO controls still apply. AFP’s 2026 survey found that 17% of organizations use AI against payments fraud.
How many invoices do you need to test?
Fifty gives a rough read and 100 gives a better one. The range around each rate is wide on small samples: 48 right out of 50 is consistent with a true rate of about 87% to 99%.
How often should the test be repeated?
Repeat it after a large batch of new vendors, a chart of accounts change, a new invoice type or a major software update. Between tests, review a monthly sample.
Is AI replacing accounts payable?
No. Someone still owns the books. AI moves work from entry toward review and exceptions.
Does ProcureDesk use AI for GL coding?
Yes. Its AI capabilities page describes GL coding predictions with confidence scores that learn from corrections.
How long does implementation take?
Implementation takes 2 to 3 weeks.
Resources
- How CPA Firms Get Clients to Approve Spend Before the Invoice Arrives: the first article in this series
- The 30-Minute AP Exposure Review CPA Firms Can Run on Any Client: the second article in this series
- AI capabilities: GL coding predictions, approver insights and duplicate detection
- AP automation: invoice capture, matching and approval
- AP software overview: capture, validation and ERP sync
- Invoice matching: matching invoices to POs and receipts
- Three-way match: PO, receipt and invoice matching explained
- PO-Before-Invoice Rule framework: the full definition of No PO, no pay
- The Missing Layer in Your CAS Tech Stack: why a PO before the invoice matters for CAS firms
- Beyond AP Automation: A CAS Playbook: turning AP cleanup into advisory work
- CPA partner program: offering ProcureDesk to your clients
About the author
Sachin Sharma
Sachin Sharma is the CEO of ProcureDesk and has spent over 23 years in procurement and supply chain technology. He previously led procurement operations at a Fortune 500 company before founding ProcureDesk. Connect with him on LinkedIn.
