AI accounts payable automation: what the tools actually do (September 2026)

Where accounts payable time and money go today

Ardent Partners surveyed 212 accounts payable and finance leaders between March and May 2024 for its 2025 benchmark report, 59 percent of them in North America and 30 percent in Europe, the Middle East and Africa. The average team spent $9.40 to process one invoice, all costs included, and needed 9.2 days to process it.

The gap between teams is large. Ardent defines best-in-class as the 20 percent of companies with the lowest cost and shortest cycle time. They process an invoice for $2.78, against $12.88 for all others, and in 3.1 days, against 17.4. Their exception rate is 9.0 percent, against 22.0 percent.

For the first time in the 19 years of the study, invoice exceptions ranked as the top challenge, named by 53 percent of respondents. Exceptions are invoices that stop because a code is wrong, data is missing, an approval stalls or a purchase order does not fit. That is where AI has to prove itself, because clean invoices were already cheap to process.

Sources: [1]

What AI accounts payable automation does, step by step

Split the process into its steps and the picture gets clearer. Capture reads supplier, amounts, VAT and lines from a PDF, scan or e-invoice. Coding proposes the general ledger account, cost center and tax code. Matching compares the invoice with a purchase order and goods receipt. Approval routing finds the right person. Checks look for duplicates and fraud. Payment schedules the transfer.

Capture and coding are the most mature steps. Candis states that its AI data extraction captures up to 91 percent of information correctly and learns recurring patterns for pre-coding. Medius says its SmartFlow model fills in coding, tax and approver values for invoices without a purchase order with 95 percent precision after just two invoices from a supplier.

Agents now reach into the later steps. Tipalti lists separate agents for invoice capture, purchase order matching based on contextual descriptions, and finding bill approvers. BILL describes an Invoice Coding Agent that codes multi-line bills based on previous coding behavior and says it reduces manual time by 20 percent. Hypatos runs a set of agents that also resolve the supplier, check for duplicates and prepare the posting in the ERP.

Payment itself remains mostly rule based: approved invoices are paid on their due date or when a discount applies. The AI work sits upstream, in making sure the invoice that reaches the payment run is correct.

Sources: [2][3][4][5][6]

How to read vendor accuracy and touchless claims

Vendor numbers describe different things. Candis reports a share of correctly captured information. Medius reports precision for coding suggestions after two invoices. Rossum says its Aurora model, trained on 11 million transactional documents, reaches 92.5 percent average accuracy across customers. BILL reports saved manual time. None of these is the share of invoices that go from inbox to payment without a person.

Medius goes further and speaks of a 100 percent touchless capture rate. Capture is one step. Ardent's benchmark measures the whole process, where even best-in-class teams processed 49.2 percent of invoices without human touch. Both can be true at once, because an invoice captured perfectly can still stop at coding, matching or approval.

Ask every vendor the same three questions. What exactly was measured, on which invoices, and over what period? Is the figure per field, per invoice or per process? And what does it look like for customers with your mix of suppliers, languages and purchase order share?

Sources: [2][3][7][5][1]

Fraud checks: the step where AI adoption is thin

Fraud is the risk that makes accounts payable automation dangerous if done badly: a faster process also pays a fake invoice faster. In Ardent's survey, 29 percent of respondents named fraud as a top threat.

The AFP's 2026 Payments Fraud and Control Survey, based on 465 treasury professionals in US organizations, found that 76 percent experienced attempted or actual payments fraud in 2025 and 74 percent were affected by business email compromise. Only 17 percent used AI to combat payments fraud.

Some tools now add checks before approval. GetMyInvoices describes a Fraud Risk Assistant that analyzes incoming invoices with statistical models and flags anomalies before release, and states plainly that it cannot detect every fraudulent invoice. Treat any fraud feature as a second line behind supplier master data controls, not as a replacement for them.

Sources: [1][8][9]

The system behind the tool decides what AI can learn

Coding and matching suggestions are only as good as the data they learn from: past bookings, supplier master data, purchase orders and goods receipts. That data lives in your accounting system or ERP, so the connection matters more than the AI label on the invoice tool.

Tools differ by the systems they are built around. Hypatos names integrations with SAP, Workday, Coupa and xSuite and uses ERP and master data to assign the correct legal entity. Esker describes processing from receipt to ERP posting across several ERP environments from one cloud platform. Stampli says its AI was trained on more than $150 billion in annual spend across more than 70 ERPs.

Smaller companies usually start from their accounting software or tax advisor. Candis offers a direct connection to DATEV and native connections to SAP Business One, Microsoft Business Central and Sage 100, including three-way matching for goods invoices. Before a pilot, check whether coding suggestions flow back into your system as proposals your team confirms, or as bookings that nobody reviews.

Company size narrows the field quickly. Our directory marks 40 accounts payable tools as serving companies under 50 employees and 47 as serving enterprises with more than 1,000, with many tools covering several sizes.

Sources: [6][10][11][12][13]

The accounts payable market in our directory

Our directory lists 92 AI tools for accounts payable automation, including tools where it is a secondary category, such as spend management platforms. 29 of them are headquartered in the United States and 12 in Germany.

Pricing is often a sales conversation: 52 percent of the tools in this category publish no price model, which usually means enterprise pricing. 27 offer a free trial and 23 can host data in the EU. Browse the full list on our accounts payable category page.

The explorer below shows the same tools with their security reports, trial and hosting. If your invoices come from German or Belgian suppliers, check e-invoice support separately: in our check of 32 invoice tools, none of the US-based vendors named the German formats XRechnung or ZUGFeRD.

Sources: [13]

How to run a pilot that tells you something

Start with your own numbers. Count invoices per month, the share with a purchase order and how many stop as exceptions today. Without that baseline, no vendor figure means anything for you, and you cannot prove the result afterwards.

Run the pilot on real invoices, not the vendor's samples, and include the difficult ones: foreign suppliers, multi-line invoices, credit notes and invoices without a purchase order. Measure the touchless rate and the exception rate over at least one full month close, and track how many AI suggestions your team changes.

Then compare cost. The calculator below uses Ardent's all-in cost per invoice as a reference. It is a benchmark from a survey, not a quote for your company, but it shows how much the gap between average and best-in-class is worth at your volume. For a broader view of finance tools by task, start from our accounting category.

Sources: [1]