Document intelligence for invoice processing, measured in cost per document.

The straight-through rate sets the cost.

An invoice goes straight through only when every field it needs clears the bar, and a field accuracy figure hides how rarely that happens. If an invoice needs 20 fields and each is read correctly 98% of the time, independently, only about two invoices in three come out entirely right (0.98 to the power of 20 is roughly 0.67). Vendor sheets quote the 98%, and the 67% is what lands on your finance team.

Ardent Partners' 2026 report puts the average exception rate at 19.9%, against 11.8% for its top performers, which leaves plenty of room to move. The touchless rate and processing time below come from the 2025 edition, the latest to publish them openly.

Your own two numbers come first: the straight-through rate and the cost per document, measured on your documents. The build is designed to move both.

Your volume, priced at published benchmarks.

The calculator applies the published cost per invoice to your monthly volume. It is there to show whether a conversation is worth having at your size.

Rules for the predictable, models for the rest.

Your regular suppliers, in the formats they always send, go through deterministic rules that cost almost nothing to run and cannot invent a value. The model takes the variance: new suppliers, bad scans, Arabic and mixed-script documents, a photograph of a screen. Where that line falls decides the running cost at your volume.

A pipeline from the inbox to the ledger.

• Capture from wherever documents arrive: email, a scanner, a phone photograph, a supplier portal, a shared drive.
• Deterministic extraction for known layouts and structured feeds.
• Model extraction for the long tail, with a confidence on every field.
• Validation against your own rules: totals that add up, dates that are possible, a tax registration number in its valid format, a purchase order that exists with room left on it, and a duplicate check.
• An exception screen people will use: the crop on the left, the field on the right, one keystroke to accept.
• Straight-through reporting per supplier, per document type and per week, which shows the supplier behind most exceptions so the fix can start there.

Everything financial fails closed.

When a total does not add up, a tax number is unreadable or the duplicate check cannot run, the document waits in the queue.

Posting an invoice and paying it are separate acts. Payment runs stay with the people who approve them today, and no model can release one.

Where e-invoicing is mandated, some invoices arrive as data.

In the UAE, a Ministry of Finance mandate moves business invoices onto PINT AE, exchanged through Accredited Service Providers. Businesses with revenue of AED 50 million or more go live on 1 January 2027, and the rest on 1 July 2027. Once you and a supplier are both live, that supplier's invoices reach you through your ASP as structured data, with nothing left to read.

Plenty stays on paper: invoices from suppliers abroad, delivery notes, statements, receipts and shipping documents. Those still need reading after the mandate. Making your own ERP issue PINT AE is a separate job, covered under UAE e-invoicing readiness.

It starts with your worst few hundred documents.

• Measure first. A few hundred of your real documents, bad scans included, give a straight-through rate and a cost per document before any promise is made.
• Split the work. Rules, a model or a person, decided field by field and supplier by supplier.
• Build the pipeline. Capture, extraction, validation, posting into your ERP, and an audit trail that can reconstruct any booked document.
• Build the queue. The exception screen, the keyboard path through it, and who may approve what.
• Tune on corrections. The most-corrected fields get attention first, and a threshold is raised only when the correction log supports it.

The price rests on the volumes we measured.

The build is priced once, against the volumes and document types we measured, and paid in milestones. Hosting plus inference is estimated from your real volumes before you commit, and rules keep that running cost low for the suppliers you see every week.

On full payment the source code, the extraction rules and the validation tests are yours. Your documents train no model outside your company, and warranty and hosting follow pricing and engagement.

Bring your monthly volume and a handful of your worst documents to a free 30-minute call, or send the numbers through the contact form.