The straight-through rate decides what document automation saves.

Straight-through means nobody touched the data.

A supplier invoice is straight-through when it is received, read, checked and posted with no person looking at its data. Ardent Partners, which publishes an annual set of accounts payable benchmarks, calls this touchless processing: processing an invoice in a straight-through manner, without human intervention.

Three definitions decide whether your number means anything:

• The denominator is every document received in the period, including the unreadable ones, the duplicates and the ones you rejected. Leaving the awkward ones out is the easiest way to flatter the rate.
• A touch is any person opening, checking, editing or keying the data. Someone who opens a document and changes nothing has still spent the time.
• Business approval is separate. A manager approving spend is policy, not an exception, so report two rates: data straight-through, where nobody touched the data, and end-to-end touchless, where nobody touched anything.

Per-field accuracy compounds across every field on the page.

A document goes straight through only if every field it needs clears the bar. Accuracy is quoted per field, the cost is paid per document, and the two numbers sit far apart.

Here is the arithmetic, as an illustration. Suppose each field is read correctly, and with enough confidence to be accepted, with probability p, and suppose the fields are independent of each other. A document with n fields then goes through untouched with probability p to the power of n.

At 98% per field and 45 fields, about 40% of documents go straight through and 60% go to a person. An invoice header, with supplier, TRN, number, dates, currency and totals, is a dozen fields before a single line item, and every line adds its own description, quantity, unit price and tax. Line-heavy invoices are where rates fall.

Two corrections to the illustration matter in practice. First, the per-field number that counts is the auto-accept rate at your chosen threshold: a field read correctly but with low confidence still sends the document to a person. Second, errors are not independent. They cluster by supplier, layout and scan quality, which is good news, because fixing the worst few suppliers moves the rate more than the average suggests.

Measure it on your own invoices, bad scans included.

1. Take a sample that looks like a real month: a few hundred documents, including the bad scans, the photographed ones and the Arabic and mixed-script ones.
2. Write the field list for each document type: what posting needs, header and lines.
3. Key the correct values by hand, once. This is the ground truth everything else is scored against.
4. Run the pipeline at the confidence threshold you would use in production, not a relaxed one.
5. For each document record two facts: would it have posted untouched, and if so, was everything it posted right.
6. Report the results per supplier, per document type and per week, not only as one average.

Four numbers come out of it:

What does a manual invoice cost to process?

The benchmarks below come from Ardent Partners' surveys of accounts payable teams. Cost and exception rate are from the 2026 edition, the latest. Touchless rate and processing time are from 2025, the latest edition that publishes them openly.

Read them with their method in mind. The 2025 metrics come from 212 accounts payable and finance leaders surveyed between March and May 2024, and 53% of them worked for companies with more than USD 1 billion in revenue. The cost is all-inclusive: staff and operating costs across receipt, processing and approval, with salaries, benefits, technology and overhead. A business with a few thousand invoices a month should treat these figures as a range to test against.

The rate sets the cost per document.

Every exception is a person's time, so the running cost of a document programme comes down to three terms:

• The processing cost of every document: capture, model inference and hosting.
• The exception rate multiplied by the cost of one review.
• The false-post rate multiplied by the cost of finding and fixing one wrong posting.

An illustration, with invented round numbers. Five thousand invoices a month at a 40% straight-through rate leaves 3,000 for review. If a review takes four minutes, that is 200 hours a month. At 70% it is 1,500 reviews and 100 hours. Capture, inference and hosting cost much the same in both cases. The rate is what moved.

That is why we build a cheap, deterministic path for the predictable majority, the suppliers you see every week in the formats they always use, and reserve a model for the variance.

Sort the exceptions by supplier before tuning anything.

Sort them by document type as well. If a handful of suppliers produce most of the exceptions, the cheapest fix usually sits upstream: a PDF generated by their system instead of a scan, a structured feed, or a rule for their layout.

The UAE e-invoicing mandate will change this for part of the pile. As suppliers inside its scope go live, and once you receive through your own Accredited Service Provider, their invoices arrive as structured PINT AE data rather than as a PDF to be read. Businesses with revenue of AED 50 million or more go live on 1 January 2027. For those invoices extraction disappears, and the straight-through rate is decided by validation and matching against your purchase orders. Foreign suppliers, delivery notes, receipts and customs paperwork will still need capture. The Ministry of Finance has changed the timetable once already, so confirm the dates that apply to you with your tax adviser.

Five questions that turn a vendor demo into numbers.

Accuracy claims are easy to make and hard to compare. Put these five to any supplier:

• Is your accuracy figure per field or per document, and at what confidence threshold?
• Was it measured on our documents, including the bad scans, or on a clean sample of yours?
• Does the straight-through rate count rejected, duplicate and unreadable documents in the denominator?
• What is the false-post rate at that threshold, from an audit of untouched postings?
• What does one document cost at our volume, exceptions and reviews included?

A supplier who answers all five in writing, on a sample of your own documents, has told you more than any accuracy percentage on a slide.

An audited false-post rate keeps the number honest.

Any straight-through rate can be pushed up by lowering the confidence threshold. The rate rises, and so do wrong postings. Three habits stop that:

• Audit a random sample of untouched postings every week, and report the false-post rate beside the straight-through rate.
• Raise a threshold only after the audit numbers support it.
• Fail closed on money. A total that does not add up, an unreadable TRN or a duplicate check that cannot run sends the document to the queue instead of the ledger.

When we scope a document programme, a few hundred of your real documents go through first, the bad ones included, and the rate and the cost per document are measured before we commit to either.