How to automate invoice and document processing with AI

AI document processing turns a document into structured data that another system can use. It covers four common types.

• Purchase invoices and bills. Supplier, invoice number, date, due date, net, VAT, total and line items, posted as a bill in the ledger.
• Receipts. Merchant, date, amount and VAT, matched to a card or bank transaction.
• Contracts. Parties, start and end dates, notice periods, renewal terms and fees, written into a tracker with reminders.
• Forms. Application forms, delivery notes, timesheets and onboarding packs, written into a CRM, a spreadsheet or a database.

Bank and supplier statements are a fifth type. Dext and AutoEntry both sell statement extraction as an add-on or at a higher credit rate.

The six steps

The model call is one step of six, and the checks around it decide whether the data is safe to post.

1. Capture. Documents arrive by a dedicated email address, an upload page, a shared folder or a phone photo. Xero and QuickBooks accept email, upload and photo.
2. Classify. The system decides what each document is (invoice, credit note, receipt, statement, contract) and splits multi-document PDFs.
3. Extract. A vision model or an OCR service reads the page and returns fields as structured data, such as JSON with a fixed schema.
4. Validate. Code checks the result. Line items should add up to the net, net plus VAT should equal the total, the VAT number should match the supplier on file, and the invoice number should not already exist.
5. Review. Anything that fails a check, or falls below a confidence threshold, goes to a person in a review queue.
6. Post and file. The system creates the bill or record through the ledger's API and attaches the original file.

Xero's own capture does a version of steps 4 and 5. Its page says it matches each record against bank data and flags matches and duplicates, and that pre-filled transactions need a review before posting.

Xero and QuickBooks both keep a person in the loop by default. QuickBooks says uploaded receipts and bills "are not yet included in your books" until someone reviews or matches them.

Review everything

Each extracted document waits for a person to approve it. This suits the first month of any new system, while you learn where it goes wrong.

Review the exceptions

Documents that pass every validation check post straight through, and the rest go to a queue. A failed total, a new supplier, a VAT mismatch or a possible duplicate all send the document to a person.

Review by risk

Set rules on value and type. A firm might auto-post receipts under a set amount, review every invoice from a new supplier, and require a second approver before any payment goes out.

Spot-check the straight-through items

Pick a random sample of auto-posted documents each week and check them against the originals. Track the error rate per field. If it rises, tighten the rules.

A custom build makes sense when the documents are not standard invoices, when the data goes into a system the off-the-shelf tools don't connect to, or when your checks are specific to your business.

Build prices

Softomate, a London agency, lists invoice extraction and workflow setup from £4,500 plus £299 a month, excluding VAT. That covers one intake route and one invoice type. Full ERP posting and purchase-order matching cost extra. No independent survey covers document processing builds, so this range comes from a firm that sells the work.

API cost per page, worked out

Anthropic's vision docs say an image costs one token per 28 × 28 pixel block. A scanned A4 page at 150 dpi is 1,240 × 1,754 pixels, which is 45 × 63 = 2,835 tokens. Add instructions and a field schema, and one page is about 3,500 input tokens. Assume 500 output tokens for the extracted fields.

Anthropic's batch processing halves these prices for work that can wait. A two-page invoice roughly doubles the page tokens.

A worked example at 500 invoices a month

A firm receives 500 supplier invoices a month and needs line items. On AutoEntry each costs 2 credits, so it needs 1,000 credits. Platinum has 500, so the firm needs Diamond at £331.50 a month, which is £3,978 a year.

A custom build on Sonnet 5.5 uses 500 × $0.012 = $6 a month in API costs. The build itself, at Softomate's "from" prices, costs £4,500 + (12 × £299) = £8,088 in year one. For standard invoices the ready-made tool wins in year one. The build wins on cost only when it replaces staff time or tools that the off-the-shelf products cannot.

Yarigai builds document reading and filing, receipts into Xero or QuickBooks, invoice chasers and finance dashboards, inside the client's own accounts. The client owns the code, the data and the accounts from day one, and pays AI and hosting bills at cost. The monthly running cost is on the quote. A person approves anything that sends money, emails a customer or deletes data, so no payment leaves without a human sign-off.

A project starts with a free 30-minute scope call. The blueprint takes about a week, is credited to the build, and sets out the plan, a clickable prototype, the running cost, a fixed price and the payback maths. Yarigai quotes a fixed price after the blueprint and does not publish prices. The guarantee reads "If the build isn't profitable for you, we won't build it." If AutoEntry or Xero's own capture does the job for less, the blueprint maths will show it. See AI finance automation for what Yarigai builds and custom invoicing or Xero for the invoicing side.