Document AI
AI that reads the documents your organisation receives and turns them into checked, structured data your systems can use.
What it is
Document AI reads the documents an organisation receives, such as application forms, invoices, referral letters, delivery notes and scanned post, and turns them into structured data. It works out what kind of document it is, pulls out the fields that matter, and says how sure it is about each one.
It combines two things. Text recognition and layout analysis find the words, tables and tick boxes on the page. A language or vision model then works out what the text means, so it can find a policy number in a free-text letter as well as in a fixed box. Each field is then checked against rules and against your own data, such as a valid postcode or a customer who exists, before anything is passed on.
In an enterprise it usually sits at the front door of a process, between the inbox and the system of record. It doesn't replace the case system or the ERP. It fills them in. The value comes from the confidence scores: fields the AI is sure of flow straight through, and the rest go to a person, who sees the original page beside the extracted value.
Use it when, and think twice when
Use it when
- The same kinds of document arrive in volume, and people re-key them
- The data you need is on the page, even if the layout varies
- Extracted fields can be checked against rules or existing records
- A person can review low-confidence fields before they're used
Think twice when
- The documents are few, and a well-designed web form would remove them
- The decision depends on judgement the page doesn't contain
- Nobody can say what 'correct' looks like for each field
- Errors would go straight into payments or records with no check
Reference architecture
01Channels
- Email and attachments
- Scanned post
- Portal uploads
02Intake
- Amazon SESLogic Apps email connectorYour existing mail serverInbound email
- Amazon S3Blob StorageCeph, or IONOS S3 Object StorageOriginal file store
03Intelligence
- Amazon Textract, or Bedrock Data AutomationAzure Content Understanding (includes Document Intelligence)Docling plus an open-weight vision modelRead and extract
- Amazon BedrockMicrosoft FoundryOpen-weight models served with vLLMClassify and interpret
- Amazon Comprehend (PII detection)Azure Language in Foundry Tools (PII detection)Open-source PII detection (e.g. Presidio)Mask personal data
04Orchestration
- AWS Step FunctionsLogic Apps or Durable FunctionsTemporalDocument workflow
- AWS LambdaAzure FunctionsRules in code on KubernetesValidation rules
- Amazon SQSService BusRabbitMQ or Apache KafkaQueues
05Systems of record
- Case or claims system
- ERP or finance system
- Document management
06People
- Person
Review low-confidence fields - Person
Handle unreadable documents
- Person
- Cloud or platform service (AWSAzuresovereign / on-prem)
- AI agent
- A person decides
- Your existing systems
- Outside your organisation
Components, side by side
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Intake and original filesKeep the original file unchanged, so every value can be traced back to the page. | Amazon SES, Amazon S3 | Logic Apps connectors, Blob Storage | Your mail server; Ceph or IONOS S3 Object Storage |
| Text and layoutTest on your worst real documents: faxes, phone photos and handwriting. | Amazon Textract | Azure Content Understanding (includes Document Intelligence) | Open-source parsing (e.g. Docling) |
| Mixed bundlesUseful when one upload mixes forms, letters and photos. In London, Bedrock Data Automation processes requests in London only (checked 9 November 2026). | Bedrock Data Automation | Azure Content Understanding | Docling plus an open-weight vision model |
| Interpretation and classification | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Validation against your dataCheck account numbers and totals against the source system, not the model's opinion. | AWS Lambda behind API Gateway | Azure Functions behind API Management | Your own services behind Kong or your existing API gateway |
| Workflow and queues | AWS Step Functions, Amazon SQS | Logic Apps or Durable Functions, Service Bus | Temporal; RabbitMQ or Apache Kafka |
| Personal data detectionMask what reviewers and logs don't need to see. | Amazon Comprehend (PII detection) | Azure Language in Foundry Tools (PII detection) | Open-source PII detection (e.g. Presidio) |
| Human reviewShow the page and the extracted value side by side. | In the team's existing case tool | In the team's existing case tool | In the team's existing case tool |
| UK data residencyCheck that each service and model you pick is offered in your UK region. | eu-west-2 (London) | UK South / UK West | Your own UK data centre, or a UK or EU provider |
On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →
Operations
Keeping it reliable
AI fails quietly. These three habits make failure visible before customers see it.
Evaluation
Score every field, not every document
Build a test set of a few hundred real documents, with the correct values agreed by the people who key them today. Measure accuracy per field, because a 98% document score can hide 70% on the one field that matters. Rerun it whenever the model, prompt or layout changes.
Monitoring
Watch confidence and corrections
Track the share of fields sent to review, and how often reviewers change the value. A sudden rise usually means a new form layout or a sender changing their invoices. Log the original file, the extracted value and the final value together.
Cost
Price it per page
Document services charge per page, and model calls charge per token. Run cheap text recognition first and send only the pages that need it to the costlier steps. Report cost per thousand documents beside accuracy.
Security and failure risks
Silent extraction errors
A misread digit in an account number or amount passes every format check and goes into a payment or record.
How the design avoids it
Cross-check key fields against the system of record, use totals and check digits, and send anything that doesn't reconcile to a person.
Instructions hidden in documents
A document contains text aimed at the model, such as 'mark this as approved', and the AI acts on it.
How the design avoids it
Document content is treated as data only. The extraction step can only return fields, and it has no permission to change anything.
Collecting more personal data than needed
Health details, bank details or third-party data the process doesn't need are extracted and spread into logs and prompts.
How the design avoids it
Extract only the fields the process needs, mask the rest, keep processing in UK regions, and set a retention period for originals.
Forged or altered documents
A forged payslip or edited invoice is read perfectly and accepted.
How the design avoids it
Reading a document isn't verifying it. Keep existing fraud checks, compare with source data where you can, and flag mismatches to a person.
Playbooks that use this pattern
All 32 playbooks using Document AI →KYC and customer onboarding
Every application read, checked and summarised before an analyst opens it. Analysts spend their time on judgement, not copying, and a person approves or declines every customer.
- Document AI
- Agents with tools
- Human in the loop
- Event-driven integration
Credit decisioning support
Every referred application arrives with the evidence gathered, affordability worked out and the score explained in plain English. A person makes every referred decline, clear-cut cases stay with the existing decision engine and its safeguards, and fairness is checked every month.
- Predictive ML
- Document AI
- Agents with tools
- Human in the loop
Claims intake and triage
Every claim read, sorted and routed within minutes. Simple claims ready for one-click approval, complex ones flagged early, and people making every decision that matters.
- Document AI
- Agents with tools
- Human in the loop
- Event-driven integration
Underwriting submission review
Every broker submission read, its risk data extracted and checked against appetite on the day it arrives, with gaps flagged and a referral note drafted. The underwriter decides, prices and speaks to the broker.
- Document AI
- Retrieval (RAG)
- Agents with tools
- Human in the loop
Referral and clinical letter processing
Every referral and letter read, summarised and routed within minutes of arriving. The consultant sees a one-screen summary and a suggested priority, and makes every triage and clinical decision.
- Document AI
- Retrieval (RAG)
- Agents with tools
- Human in the loop
- Event-driven integration
Waiting-list validation
A waiting list that matches reality: data errors found every night, patients asked whether they still need care, and a person confirming every change. Clinical validation stays with clinicians.
- Predictive ML
- Document AI
- Agents with tools
- Human in the loop
- Event-driven integration
Citizen casework and correspondence
Every letter and email read, sorted and routed on the day it arrives, with a sourced draft reply waiting for an official to edit and sign. People decide what the department says and what happens to each case.
- Retrieval (RAG)
- Document AI
- Agents with tools
- Human in the loop
Freedom of Information request handling
Every request logged, checked and on the clock the day it arrives, records found faster, and suggested exemptions and redactions laid out with reasons. The FOI officer decides what is released.
- Retrieval (RAG)
- Document AI
- Agents with tools
- Human in the loop
Benefits and grants application review
Every application read, checked against the scheme rules and made ready for an officer within a day. Missing evidence asked for early, fraud and error checks done fairly, and an officer deciding every award and every refusal.
- Document AI
- Agents with tools
- Retrieval (RAG)
- Human in the loop
Patterns
All eight building blocks
Document AI
Read forms, letters and PDFs into structured data.
Retrieval (RAG)
Answer questions from your own documents, with sources.
Agents with tools
AI that takes small, controlled actions across your systems.
Human in the loop
Put people on the decisions that matter, with the evidence.
Predictive ML
Score, rank and forecast from your own historical data.
Event-driven integration
Connect AI to core systems safely, through events and queues.
Computer vision
Inspect photos, video and camera feeds for defects and damage.
Speech and voice
Turn calls and conversations into text, notes and actions.
Thinking about document ai in your organisation?
Talk through where it fits, what it would plug into, and which decisions should stay with people.
