AI automation · Documents Intelligent document processing with AI: from inbox to approved data, with a person at the end

The problem

In most businesses the same thing happens a hundred times a week. A document arrives, someone opens it, finds the supplier, the date, the amount or the termination clause, types it into another system and files the original. The typical sources of error are re-keying, documents with unfamiliar layouts, clauses that are overlooked on page 14 and documents assigned to the wrong case. The work is repetitive, but it needs attention, so it is slow and tiring.

The workflow in seven steps

1. Intake: emails, portal downloads, scanner output and shared folders flow into one entry point, with the sender and the time recorded.
2. Recognition: the system identifies the type of document, such as invoice, contract, certificate or letter, its language and its quality. Scans go through text recognition first.
3. Extraction: the AI reads the fields and passages you need and returns them as structured data, each with the page and the position it was found.
4. Validation: rules check the result against your master data: the supplier exists, the totals add up, dates are plausible, mandatory clauses are present. Values the AI is unsure about are marked.
5. Review: a person sees the document next to the extracted values, and only the exceptions. Corrections are kept and improve later runs.
6. Release: approved data goes to the ERP, the document system or the CRM, the original is archived, and deadlines found in the document become tasks.
7. Log: who approved what, what the AI changed and which version of the rules applied, as the basis for audits.

Three levels of automation

Rules and templates

For a few stable layouts, such as your own forms, fixed rules are enough. They are predictable and cheap to run, and they break when the layout changes.

AI‑assisted extraction

For many layouts, such as supplier invoices or contracts from different counterparties, a language model reads the document and fills the fields. A person reviews exceptions. This level covers most of what people call intelligent document processing, and for contracts it covers review against a checklist of clauses and deviations.

Agents

When a document starts a process, the system also takes the next steps within limits you set: ask the sender for the missing page, create the task, book the deadline, chase a reminder. Each step with consequences goes through an approval. More on this under AI agents for business.

Systems and interfaces

The workflow connects to what you already use: mailboxes through standard protocols or the Microsoft and Google APIs, shared drives and SharePoint, document systems such as DocuWare or M-Files, and ERP systems such as SAP, Microsoft Dynamics 365 or Odoo. For invoices in structured formats, the format itself carries the data and the check is simpler, see automating incoming invoices. The names are for orientation only and do not imply any partnership.

Law and risk

• Data protection: documents contain personal data. You need a data processing agreement with every provider (Art. 28 GDPR), processing in the EU or UK and a clear rule on how long originals and extracted data are kept. The ICO guidance on AI and data protection and the EDPB Opinion 28/2024 explain what regulators expect.
• Automated decisions: Art. 22 GDPR limits decisions about a person that are based solely on automated processing. The workflow extracts and checks, people decide.
• Out of scope: reading applications to rank candidates and assessing creditworthiness are high-risk under Annex III of the AI Act. We do not build them, see risk categories.
• Retention and evidence: accounting, tax and contract documents have retention periods set by national law. The archive keeps the original, and the extracted data supports it but never replaces it.
• Contracts: extraction and comparison against a checklist support the review. Legal assessment stays with a lawyer, see AI for law firms.
• Staff: Art. 4 of the AI Act asks for sufficient AI literacy. This overview is general information, not legal advice.

How to measure success

Measure before and after on a sample of your own documents. Useful figures are the share of documents that pass without any correction, the correction rate per field, the time from receipt to booking or routing, the share of documents that land in the exception queue and the number of deadlines missed. Results depend on the quality of the documents and the rules, so we agree the measures with you at the start and do not promise figures in advance.

Where this applies

• Accounts payable: invoices, order confirmations and delivery notes matched, see AP automation.
• Law firms and legal teams: contracts and bundles checked against a clause list, see AI for law firms.
• Insurance: policies against the order and claim documents sorted, see AI for insurance.
• Logistics: CMR notes and delivery receipts matched to orders, see logistics AI.
• Manufacturing: certificates and test records checked against specifications, see AI in manufacturing.