Accounts payable automation with AI: invoice to bill

What AI invoice processing in accounts payable means

Accounts payable automation with AI turns incoming supplier invoices into draft bills: it reads the PDF or e-invoice, checks the arithmetic and VAT, suggests the account and cost centre, matches the invoice to the purchase order and goods receipt, and flags duplicates and changed bank details. It matters because a person then approves exceptions instead of retyping every invoice.

The value is not speed of typing. For a small business the real gains are consistency and control: every invoice gets the same checks, including the ones a busy person skips at five o'clock on a Friday, and every exception arrives with the reason it was raised. The risk is equally specific. An AI that reads a fraudulent invoice fluently and codes it neatly has made fraud faster, so the design has to put payment controls outside the model's reach.

For the accounting behind a bill, see our guides on accounts payable versus accounts receivable and on invoices, receipts and bills. This guide covers the automated route from inbox to approved bill.

The accounts payable pipeline, step by step

Every AP process, manual or automated, runs the same six steps. Automation changes who does each step and how exceptions surface, not the steps themselves. The mode column is the one to agree with your auditor and your approvers before you switch anything on.

• Capture: receive the invoice in one dedicated mailbox or portal and store the original unchanged. Mode Auto.
• Extract: read supplier, invoice number, dates, lines, net, VAT and total, or take them straight from a structured e-invoice. Mode Draft.
• Validate: arithmetic, VAT rate and VAT number, mandatory invoice content, supplier and bank details against the supplier master. Mode Read, raising flags.
• Code: propose the expense or inventory account, cost centre and VAT code, with the reason. Mode Draft.
• Match: compare with the purchase order and the goods receipt within agreed tolerances. Mode Draft for matched invoices, exception queue for the rest.
• Approve and pay: a named person approves the bill; payment runs are released separately under dual control. Never Auto.

Capture: PDFs, scans and structured e-invoices

Invoices arrive in three forms, and they need different handling. A structured e-invoice, such as an XML file following the European standard EN 16931, already contains the data as labelled fields; there is nothing to extract, only to validate and map. A PDF generated by the supplier's system contains text that software can read accurately. A scan or phone photo is an image, and every figure on it is an interpretation.

AI earns its place on the last two. As of September 2026, Claude reads PDFs by converting each page to an image and extracting its text alongside, so it can read tables, charts and other visual content on the page as well as the text layer. Anthropic's citations feature can point to the passage an answer relied on, but its documentation notes that scanned PDFs with no extractable text are not citable, so the reviewer cannot click from a figure to the source on a scan. Treat scanned invoices as higher risk and route them to review more often.

Germany shows where Europe is heading, and why the phrase kreditorenbuchhaltung automatisieren now starts with e-invoices. According to the Federal Ministry of Finance, every domestic business has had to be able to receive e-invoices since 1 January 2025, and an ordinary e-mail inbox is enough to do so; issuing becomes mandatory in stages from 1 January 2027, for businesses with prior-year turnover above EUR 800,000, and for everyone from 1 January 2028. A plain PDF is not an e-invoice under the German rules. An AP process that only knows how to read PDFs will therefore spend the next few years handling a shrinking share of German invoices. Our guide on EU e-invoicing mandates covers the other countries.

The practical rule: when a structured invoice is available, use its data and keep AI for exceptions and coding; when only a PDF or scan exists, let AI extract, and make the validation step prove the extraction.

Validation and VAT checks

Validation is deterministic and should be done by software rules, not by asking the model whether the invoice looks right. The model's extraction is the input; the checks are code. In the EU the right to deduct input VAT depends on holding an invoice that meets the Directive's content rules (Council Directive 2006/112/EC, Article 178 and Article 226), so a missing VAT number or supply date is not a formality.

When a check fails, the invoice goes to a person with the failed check named. A reviewer who sees the flag net 1,250.00 plus VAT 228.50 gives 1,478.50 but the printed total is 1,487.50 can open the PDF, see that the VAT is really 237.50 and fix it in seconds; a reviewer who sees only a low-confidence score has to start from scratch.

• Arithmetic: each line's quantity times price equals the line net; lines add to the net total; net plus VAT equals the gross total printed on the invoice.
• VAT rate: the rate on each line is a valid rate for the supplier's country and the goods or services, and VAT equals rate times net within a cent of rounding.
• VAT number: present, in the right format for the country, and matching the supplier master; for cross-border EU supplies, validated through the European Commission's VIES service.
• Mandatory content: invoice number, invoice date, supply date where different, supplier and customer names and addresses, description and quantity, and the reverse-charge wording where the customer accounts for the VAT.
• Dates: the invoice date and supply date fall in an open period; an invoice dated before the goods receipt is flagged.
• Currency: the invoice currency matches the purchase order, and any foreign-currency amount is converted at the rate your policy prescribes.

Coding the bill: accounts, cost centres and capital items

Coding is where AI is most useful and most in need of your policies. A model can learn from last year's postings that a supplier's invoices go to software subscriptions and the IT cost centre, and it can read a description well enough to tell a laptop from a laptop repair. What it cannot know unless you tell it is your capitalisation threshold, your prepayment policy or which projects are billable to customers.

Put those policies in the instructions and make the model state its reason for each line: supplier history, description, purchase order account or policy rule. Coding suggestions without a reason are hard to review; suggestions with a reason can be reviewed in bulk by scanning for the reasons that look wrong.

Watch three recurring errors. Annual subscriptions coded entirely to the month of the invoice instead of being split into a prepayment. Capital items below or above the threshold coded the wrong way. And freight or duty on stock purchases coded to expenses, when IAS 2 includes transport and other directly attributable costs in the cost of purchase of inventory.

Two-way and three-way matching

A two-way match compares the invoice with the purchase order: right supplier, items, quantities and prices. A three-way match adds the goods received note, so the business pays only for what it ordered and actually received. Services without a delivery are usually two-way matched and approved by the budget holder instead.

Tolerances decide how many invoices a person has to look at. A typical policy allows small price differences per line, expressed as the lower of a percentage and a fixed amount, and no quantity difference above what was received. Set tolerances deliberately: too tight and every rounding difference queues for review, too loose and small overcharges become a habit.

Rules do most of the matching. AI adds value on the invoices rules cannot pair: descriptions that differ from the order (a box of 12 on the order, 12 units on the invoice), one invoice covering several orders, partial deliveries, and freight or surcharges that are not on the order at all. For each of these the model should propose the pairing and explain it, and the proposal stays a draft until a person accepts it.

Worked example: a three-way match reviewed by a person

A German wholesaler orders on 20 August: 200 metres of cable at EUR 6.00 (1,200) and 50 connectors at EUR 2.40 (120), total net 1,320. The goods received note records 200 metres of cable and 40 connectors; 10 connectors are back-ordered. The supplier's invoice arrives as a PDF: 200 metres of cable at EUR 6.30 (1,260), 40 connectors at EUR 2.40 (96) and freight of 35, net 1,391, VAT of 264.29 at 19 percent, total 1,655.29. The arithmetic holds: 1,260 plus 96 plus 35 is 1,391, and 1,391 times 19 percent is 264.29.

The AI's proposal and the reviewer's decisions follow. The model's reading of the documents was accurate; its one wrong suggestion was a judgement about which price applies, which is exactly the kind of call that belongs to a person.

• Connectors: 40 invoiced, 40 received, price as ordered. AI proposes a match. Accepted.
• Freight 35: not on the order. AI cites the supplier contract, which charges freight on orders below EUR 1,500; this order was 1,320. Accepted.
• Cable: invoiced at 6.30 against 6.00 ordered, a variance of 200 x 0.30 = 60, or 5 percent, outside the 2 percent tolerance. AI flags it correctly.
• AI also suggests accepting the new price because an e-mail in the thread announces a price list from 1 September. Rejected: the order was placed on 20 August at 6.00, and a price-list e-mail is not an agreed change to an order. The buyer confirms 6.00 and requests a credit note.
• Checks with no exceptions: VAT number and IBAN match the supplier master; no invoice with the same number or the same amount from this supplier in the last 12 months.
• Bill posted: Dr Inventory 1,391 (freight included in cost of purchase) / Dr Input VAT 264.29 / Cr Trade payables 1,655.29.
• Credit note posted on arrival: Dr Trade payables 71.40 / Cr Inventory 60.00 / Cr Input VAT 11.40.
• Net amount payable: 1,655.29 minus 71.40 = 1,583.89, which equals the agreed net of 1,331 plus 19 percent VAT. Payment is held until the credit note is posted.

Duplicate invoice detection

Duplicates are rarely exact copies. The same invoice arrives as a PDF by e-mail and on paper by post, a supplier resends it as a reminder with a new date, or a clerk keys INV-0045 where the system already holds 45. Exact-match checks on supplier and invoice number catch only the easy cases.

A layered check works better: normalise invoice numbers by removing prefixes, spaces and zeros at the start; compare supplier, amount and invoice date within a window; compare the same amount and date across different supplier records, which catches a supplier set up twice; and compare bank accounts, since two supplier records paying into one IBAN deserve a question. AI helps with the fuzzy layer, spotting that a reminder describes the same delivery as an invoice already posted.

A suspected duplicate is held, never deleted, and a person decides. If a duplicate is paid, the recovery is a debit to the supplier's account and a request for a refund or credit, and the case goes on a list for the next review of the matching rules.

Payment fraud and supplier bank-detail changes

The costliest AP fraud does not involve a fake invoice at all. A fraudster impersonates a real supplier, often from a look-alike e-mail domain, and asks for future payments to go to a new bank account. The next genuine invoice is then paid to the fraudster. Auditors consider this kind of risk under ISA 240, and it is the control area where AI must be kept furthest from the decision.

Verification of payee helps in the euro area. The European Central Bank states that under the Instant Payments Regulation, payment service providers in euro-area member states had to offer a verification of payee service from 9 October 2025, and providers in non-euro member states from 9 July 2027. The service compares the account identifier with the name of the intended payee and returns match, close match, no match or other before the payment is made. It is a useful extra check, not a replacement for your own control over supplier master data.

• No bank-detail change from an invoice or e-mail alone: confirm by calling the supplier on a number already on file, never one given in the request.
• Dual control: one person enters a bank-detail change, a different person approves it, and the change is logged with old and new values.
• Cooling-off: hold the first payment to changed bank details for review, whatever the amount.
• AI flags, never acts: the model flags an IBAN on an invoice that differs from the master, a new supplier with urgent payment terms, or a sender domain that differs from previous invoices. It has no tool that can change supplier master data or release a payment.
• Treat verification of payee results of close match or no match as a stop, not a warning to click through.

Security and data protection for AI in AP

Supplier invoices are untrusted input from outside the business, and an AI that reads them can be targeted. A PDF can carry hidden text such as an instruction to change the bank account or to approve the invoice. Anthropic's guidance on indirect prompt injection is to treat third-party content as data, tell the model where it came from, state in the instructions that such content must never override them, and apply least privilege so that a successful injection can do little harm. In AP that means the extraction model can read invoices and write drafts, and nothing else.

Invoices also contain personal data: sole traders' names and addresses, employee expense details and sometimes bank details of individuals. Under the GDPR you need a lawful basis for the processing, a processor agreement with any AI provider, and a view on international transfers. Anthropic states that by default it does not use inputs or outputs from its commercial products, including the API, to train its models. Send only the documents the process needs, keep originals in your own archive, and log each AI action: the document read, the model version, the proposal and the person who approved or rejected it. This is general information, not legal advice.

Accounts payable automation for small business: choosing software

Accounts payable automation for small business does not need a separate product for every step. The questions below separate software that controls payables from software that only speeds up data entry. When people ask for the best accounts payable automation software, the honest answer is the one that passes these tests on your own invoices, so run a pilot on last month's real invoices before you decide.

• Does it validate structured e-invoices (EN 16931) as well as read PDFs, and keep the original file unchanged?
• Does every extracted figure link back to the document, and are VAT and totals recomputed by rules rather than trusted from the model?
• Are two-way and three-way matching tolerances configurable per supplier or category?
• Can a bill above a threshold be approved only by a named approver, and is that recorded?
• Are supplier bank-detail changes under dual control, logged, and out of reach of the AI?
• Does the tool post into your ledger with the right VAT and cost centre, and refuse closed periods?
• Where is data processed and stored, and is there a processor agreement?

Accounts payable in Skyline Nexus ERP

In Skyline Nexus ERP a purchase or purchase return posts its journal to accounts payable, inventory or expense and the VAT Input Account as soon as it is saved, once an administrator has switched on automatic posting for purchases. Product categories can carry their own purchase account, purchase returns work as debit notes, and the supplier side has a Supplier Ledger, an Accounts Payable Report and an AP Aging Report. Treasury adds Payment Processing with Scheduled Payments, Bulk Payments, Hold and Release and Invoice Matching screens, plus payment approvals.

The in-app assistant answers how-do-I questions from the product's help library and, within the user's permissions, gives live figures such as supplier dues or purchases this month; Claude Desktop can reach the same figures through the Skyline Nexus ERP MCP server with a personal token an administrator issues. AI invoice capture and Claude drafting bills for approval are being rolled out on the Skyline Nexus ERP roadmap, and national e-invoicing connectors for European countries are being rolled out market by market. Tell us your country and we will confirm your go-live date.