AI document processing for a UK SME: invoices, applications and referrals into your systems, with accuracy and cost.

A document arrives: a supplier invoice by email, a job application from a board, a referral from a GP, a signed form, a delivery note photographed on a phone. Today a person opens it, reads it, and types the parts that matter into the system, then files it. Document processing does the reading and the typing: it extracts the fields the process needs, matches them to existing records (the purchase order, the vacancy, the patient, the job), writes them to the system of record, and files the document against the record. Where something is missing, ambiguous or unusual, it stops and asks a person, with the document and its best reading attached. The person’s day changes from typing everything to checking the exceptions.

Before the build, take fifty real documents, including the scanned, the skewed and the handwritten, and record the right answer for each field. That set is the test. Agree the threshold the process needs: for an invoice total, near perfect with every miss caught by a match against the purchase order; for a free-text field, lower, with a person reviewing. Run the pipeline against the fifty, count the misses per field, and route the fields that fall short to the exceptions path. After go-live, sample weekly and keep the count. The threshold is what makes “finished” a fact rather than an opinion, and it is what we agree in writing before quoting.

Thresholds are agreed per process before the build and measured on fifty real documents.

A typical invoice or application is a few thousand tokens of input and a few hundred of output. On a small model at September 2026 prices that is a fraction of a penny per document; a firm processing 2,000 documents a month spends pounds, not hundreds, on the model. The costs that matter are elsewhere: the integration with the system of record, the exceptions interface a person will use every day, and any channel around it. Our running-cost guide gives worked examples, and our model guide explains why the smallest model that passes the test is the right one for this work.

A first document pipeline, one document type, one system of record, an exceptions path and a measured threshold, is an implementation at £4,000 to £15,000 fixed at UK rates. Several document types or several systems is a custom system at £8,000 to £25,000. Published UK ranges sit around those bands with day-rate suppliers higher. Off-the-shelf capture products exist for common cases such as invoices into mainstream accounting packages and are worth checking first; custom is right when the documents are yours (applications, referrals, forms) or the system of record has no product connector.

Applications, referrals and forms contain personal data and often special-category data. The pipeline runs on the API under a data processing agreement, the model is not trained on the content, documents stay in your systems, and where health or similar data is involved a DPIA covers the use and the region of processing is chosen deliberately. Supplier invoices are lighter, but payment details and personal names still carry the ordinary duties. Our GDPR guides set out the provider side; the DPIA template covers yours.

Four hundred supplier invoices a month, keyed and matched by hand at nine minutes each across the people who touch them, at a loaded £28 an hour: 60 hours and £1,680 a month, with one in twenty reworked and three days from arrival to posting. After the pipeline: 14 hours a month on exceptions and review, £392, one in fifty reworked, same-day posting, and a model cost of a few pounds. Monthly saving about £1,250, annualised about £15,000, against a build at £4,000 to £15,000. The figures are illustrative; the structure, hours times rate minus build and running cost on one measured process, is how the case is made and how it is checked at day 90.