What is optical character recognition?

Optical character recognition (OCR) is technology that converts images of text, such as scans, photographs and image-based PDFs, into text a computer can read, search and process. It is often the first step before an AI agent can use a document, so any character it misreads becomes a fact the agent may act on.

Why does OCR matter once agents act?

When a person reads a scanned contract, a smudged digit is obvious. When OCR reads it, the smudge becomes a confident wrong number, and an agent reading the output has no way to see the smudge. A misread sort code, date or amount then flows into a draft letter, a schedule or a payment file.

Modern language models can read images directly, which blurs the line between OCR and the agent. That makes it easier to build, and harder to see where a misreading entered.

How does OCR work in practice?

OCR systems clean up the image (straightening, removing noise), find blocks and lines of text, recognise characters and words, and often return a confidence score for each. Better systems also recognise layout: tables, columns, headers and handwriting.

For agent work, the useful outputs are not just the text but:

• confidence per field, so low-confidence values can be checked;
• position, so a reviewer can see the original area of the page;
• page and document reference, so the fact can be traced back.

What does an example look like?

A software company's finance team uses an agent to prepare revenue recognition schedules. Many customer contracts are signed scans. OCR reads each contract; the agent extracts start date, term, price and performance obligations, and drafts the schedule at depth 2 for an accountant to review. On one contract, OCR reads a start date of 01/08 as 07/08. Because the extracted date carries a low confidence score and a link to the page, the accountant spots the error in seconds. Without the link, it would have shifted a month of revenue.

How is OCR different from intelligent document processing?

OCR produces text. Intelligent document processing goes further: it classifies the document, extracts specific fields with meaning ("this number is the invoice total"), validates them against rules and routes the result. OCR is one stage inside IDP. An agent sits beyond both: it decides and acts on the extracted data.

That layering matters for controls. OCR errors should be caught at extraction, with confidence scores and validation, before they become the agent's facts. If they reach the agent, they need to be visible to whoever reviews the agent's draft.

Where does OCR matter most?

In operations still heavy with paper and scans: trade confirmations and legal agreements in capital markets, identity and address documents in onboarding, claims forms in insurance. A UK capital markets operations team may receive scanned settlement instructions and signed agreements daily; the operating context is in AI agents in UK capital markets operations. Where a misread figure could move money, the extracted value should be checked against a sealed record, such as the firm's own static data, rather than trusted on its own.

Which terms sit close to it?

Intelligent document processing, handwriting recognition and straight-through processing. For how document reading fits into running agents, see our guide to deploying AI agents.

Where Pryme fits

In Pryme, text read from a customer's scan is declared or external evidence, depending on where the document came from, and keeps that provenance class through the agent's work. A value read from a scanned document cannot on its own authorise a payment or change payment details.