Document intelligence is software that reads business documents such as invoices, purchase orders and specifications, and returns their content as structured data: fields, tables and line items, each with a confidence score. AI document intelligence does this with trained models instead of fixed templates, so a new supplier layout needs no new setup.
At Lleverage we think document intelligence is the smaller half of the job. Reading an order correctly takes seconds. The hours go into what comes after: finding the customer in the ERP, matching their part numbers to yours, checking the price against the agreement and asking the account manager about the one line that looks wrong. That is the work our AI agents for order management and accounts payable take on, with document intelligence as their first step.
We wrote this guide for operations and finance leads at companies that make, move and sell physical products. If you would rather see it on your own documents, book a demo and bring the awkward ones.
What is document intelligence?
Document intelligence is the ability of software to understand a document well enough to pull out the information a business needs from it. The input is a file: a scanned delivery note, a PDF invoice, an Excel order with the customer's own column names, an email body. The output is data with labels on it. Invoice number, supplier, order date, 14 order lines with quantity, unit and price.
A document intelligence system knows that the number next to "Total incl. VAT" is the invoice total, even when one supplier prints it at the bottom right and another puts it in a table on page 3. It knows that a block of rows with article numbers and quantities is a set of order lines, and it keeps each line together.
Most systems do 4 things:
- Read the text on the page, printed or handwritten, from a native PDF or a scan.
- Work out the layout: tables, headers, sections, check boxes, which value belongs to which label.
- Classify the document as an invoice, order, credit note, certificate or something else.
- Extract named fields into a fixed shape, usually JSON, with a confidence score per field.
A score per field lets a system decide what a person needs to look at, so your team checks only the uncertain fields.
The term has 2 uses in practice. People use "document intelligence" as a general name for this category of software. Microsoft also uses it as a product name: Azure Document Intelligence.
How does AI document intelligence work?
AI document intelligence runs in roughly 5 steps, whatever a vendor calls them. Knowing the steps helps you see what a product does and what it leaves to you.
Step 1: capture
The document arrives from a shared mailbox, a scanner, a supplier portal or a folder. A native PDF already holds its text. A scan or a phone photo is an image, so the system first runs optical character recognition to turn the pixels into characters.
Step 2: layout analysis
The model works out the structure of the page. Which lines form a paragraph, where a table starts and ends, which cells belong to which column, which label sits next to which value. On a purchase order with 40 lines across 3 pages, this is the step that keeps line 23 from losing its quantity at the page break.
Step 3: classification
If one inbox receives invoices, orders, delivery notes and the occasional complaint, the system has to decide what each file is before it knows which fields to look for. A classifier sorts them. A misclassified document is expensive, because every later step looks for the wrong fields.
Step 4: extraction
The system fills a schema: the list of fields you want, with a type for each. A date comes back as a date, an amount as a currency value, a quantity as a number. Older systems needed a template per layout. Current models find fields by meaning and context, which is why they cope when a supplier redesigns their invoice.
Step 5: output with confidence
The result goes out as structured data, each field with a confidence score and, in good systems, a pointer to where on the page it came from. Fields below a threshold go to a person to check.
Most document intelligence products stop here and hand the clean data to the next system, or to your team.
How is document intelligence different from OCR and IDP?
Document intelligence and OCR
OCR is one part of document intelligence. It reads characters off an image and returns text. It has no idea what the text means. Give it an invoice and you get every word on the page in reading order, with no way to tell the invoice number from the bank account number.
Document intelligence adds the understanding: layout, document type, field meaning and confidence.
| Task | OCR | Document intelligence |
|---|---|---|
| Turn a scanned page into text | Yes | Yes, OCR is built in |
| Know which number is the invoice total | No | Yes |
| Keep order lines together across a page break | No | Yes |
| Tell an invoice from a credit note | No | Yes |
| Cope with a supplier's new layout | Not applicable, it returns text only | Yes, with model-based extraction |
| Return a confidence score per field | Per character or word at most | Per field |
| Output | Plain text | Named fields and tables, usually JSON |
For a searchable archive or a full-text search across old contracts, OCR text is enough. For anything that has to land in an ERP field, OCR alone still leaves a person to find the right value and type it into the right box.
We go deeper on the invoice side in our post on template-free invoice data capture compared with OCR.
Document intelligence and intelligent document processing
The 2 terms overlap.
Document intelligence usually means the capability: the models that read and understand a document. Intelligent document processing, or IDP, usually means the whole flow built around it, from intake in a mailbox through extraction and validation against reference data to a review screen and an export to the next system. Microsoft describes Azure Document Intelligence as a service for building intelligent document processing solutions, which puts the 2 in that order.
| Document intelligence | Intelligent document processing | |
|---|---|---|
| What the term points at | The reading and understanding models | The full document flow around those models |
| Typical buyer | IT and developers building their own flow | Operations or finance buying a finished tool |
| Typical output | Structured data through an API | Validated data handed to another system |
| Validation against your ERP | Left to you | Often included, depth varies |
| Who reviews exceptions | Whatever you build | A review queue in the tool |
Neither term covers what happens in your ERP after the data is ready. Our guide to intelligent document processing covers the IDP side in more detail, and the post on document data extraction into the ERP covers mapping fields to ERP records.
What is Azure AI Document Intelligence?
Azure Document Intelligence is Microsoft's cloud service for reading documents. Its full name today is Azure Document Intelligence in Foundry Tools. It was called Azure Form Recognizer until Microsoft renamed it in July 2023.
Microsoft describes it as a cloud-based service that uses machine-learning models to extract key-value pairs, text and tables from documents and return them as structured JSON. The current API version is v4.0. Version 2.1 retires on 15 September 2027 and version 3.0 on 30 March 2029, so teams on the older APIs have a migration to plan.
The service has 3 kinds of model:
- Document analysis models. Read pulls out printed and handwritten text. Layout adds tables and document structure.
- Prebuilt models. Ready-made schemas for common documents: invoices, receipts, bank statements, contracts, cheques, pay stubs, identity documents, health insurance cards and a set of US tax and US mortgage forms.
- Custom models. You label your own documents and train an extraction model on them. Microsoft offers neural models for mixed layouts, template models for fixed layouts, composed models that combine several, and a custom classifier that decides the document type before extraction runs. A custom neural model can be trained with as few as 5 labelled samples.
The v4.0 prebuilt list has no purchase order or delivery note model. For those, which are exactly the documents a manufacturer or wholesaler receives most, Microsoft points to custom models or to its newer service.
That newer service is Azure Content Understanding. Microsoft now places Document Intelligence inside Content Understanding: Document Intelligence for structured documents with familiar templates, and Content Understanding's large language models for unstructured documents, varying layouts and fields that have to be inferred. For invoices from many vendors, purchase orders with inconsistent layouts and delivery notes on different templates, Microsoft recommends a custom setup in Content Understanding, where you describe the fields in plain language without labelling. Existing Document Intelligence users keep their APIs, endpoints and billing.
On data handling, Document Intelligence keeps documents and results in Azure Storage in the region of the request and deletes them 24 hours after the request. For on-premises or air-gapped setups, Microsoft lists Document Intelligence containers as the only option today.
Azure Document Intelligence is a developer service. It gives an engineering team the reading layer, through an API and software development kits. Connecting it to your mailbox, your ERP and your exception handling is a project your team or a partner builds and maintains.
Which documents does document intelligence handle in a back office?
In a company that makes, moves or sells physical products, the documents fall into a short list. Each one ends in a different piece of work.
| Document | What document intelligence pulls out | The work that follows |
|---|---|---|
| Customer purchase order | Customer, delivery address, order lines, quantities, requested dates | Match the customer and articles, check price and stock, create the sales order |
| Supplier invoice | Supplier, invoice number, lines, VAT, totals | Match to the purchase order and goods receipt, book within tolerance |
| Supplier order confirmation | Confirmed quantities, prices and dates | Compare with your purchase order, flag changes to the buyer |
| Delivery note | Delivered quantities, batch numbers | Check against the order, book the goods receipt |
| Certificate of analysis or conformity | Batch, test values, limits | Compare with the specification, release or block the batch |
| Technical specification or manual | Requirements, process steps, references to other documents | Answer a question or draft a route, with the source cited |
The first 5 are structured or semi-structured. Every supplier has a layout, but the information is the same. The last row is different: a 60-page specification is read for answers, and the reader has to follow references from one document to the next.
SPL Treatments, an aerospace surface engineering company in Sheffield, is a clear case of that last row. Before a part it has never treated can enter the shop, someone has to read the customer's purchase order and pull the aerospace specifications it references. Those are checked against a library of more than 800 standards. Lleverage is indexing that library, about two-thirds of it so far, so the agent can retrieve the right section. The workflow reads the order, walks the references and drafts the route card, with every recommendation cited to its section. A first-time part at SPL used to take about an hour of expert reading. At SPL it now takes 2 to 2.5 minutes, and the person reviewing reads citations instead of source documents. The full account is in the SPL Treatments customer story.
"90% of things are correct within a result. It feels a bit negative to give it a thumbs down." Grant Riley, Operations and Technical Lead, SPL Treatments
Where does extraction stop and process work start?
Extraction stops when the data is correct. Process work starts when someone has to decide what the data means for your company. In our experience that second part takes most of the time a back office spends on a document.
Take a purchase order by email. Document intelligence returns a customer name, a delivery address and 12 lines with the customer's article numbers. A person handling that order still has to:
- Find the customer in the ERP, which may hold 3 accounts with similar names.
- Translate the customer's article numbers into yours.
- Check the price against the agreed price list and the quantities against stock.
- Apply the rules that live only in your company: this customer always gets a split delivery, that one orders in boxes of 50 but means single units.
- Create the sales order in the ERP with the right fields filled.
- Send the confirmation, and ask the account manager about the line that does not match anything.
Each of those steps needs your master data, your rules and access to your systems. A document intelligence service does none of them by design, and an IDP tool does the first few at best.
An AI agent receives the extracted data and does the process work, inside the systems you already run. The Lleverage agents read the order in any format, check the customer, articles, price and stock against the ERP, and create the sales order through the ERP's own interfaces, so the ERP applies its usual checks. Where there is no API, the agent works the screen with the access a person has.
Topa Bathroom Products gets orders from roughly 700 customers as plain email text, PDFs, CSVs and Excel sheets. At Topa, over 90% of incoming orders are now processed directly into Business Central by the agent, without manual input, and customers get their order confirmation within 30 seconds.
Microtechniek shows the same pattern on an older system. Each customer purchase order there needed more than 20 fields extracted and resolved by hand. Its Ridder iQ ERP runs on-premise, and the API module of the cloud version had to be installed on site before anything could write into it. At Microtechniek in August 2026, between 45 and 71 purchase orders a week went into the ERP without anyone typing them. The admin steps in only where the workflow is unsure.
Validation is process work too. Solifi, which builds the platform banks and lenders use for secured finance, developed an agentic document intelligence capability with Lleverage. It checks contracts, invoices, bank details and identity documents against application data and business rules as they arrive. Each result comes back with the evidence and the reason for any exception.
What does a person still decide?
Every agent we run scores how sure it is of each piece of work. Below a confidence line a person decides. Above it the work goes through on its own, and only the customer moves the line. When a person has to look, the request shows the field next to the source PDF. Hover a value and the matching text lights up on the page, in searchable PDFs up to their first 100 pages. When information is missing, the agent asks the right colleague in Slack or Microsoft Teams and carries on with the answer.
Corrections feed back. A Skill, the reusable description of how a task is done, collects ratings on the work it was used for, and the agent drafts an improvement that a person approves before it goes live. Over time, what your team corrects becomes knowledge the company brain keeps.
How do you measure whether document intelligence works?
Measure whole documents, because your process depends on whether a whole order or invoice went through correctly. Field-level accuracy, the figure vendors usually quote, runs higher.
Take an invoice with 20 fields, each read correctly 98% of the time. If the errors are independent, all 20 are right on about 67% of invoices, so roughly 1 in 3 still needs a person.
Microsoft's own guidance for choosing a document tool lists straight-through processing as the first thing to evaluate: the share of documents you can process without human review. We count one step further along. A document counts only when the whole process finished with no one touching it. The sales order is created, the invoice is matched and booked within tolerance, the confirmation is sent.
| Measure | What it tells you | What it hides |
|---|---|---|
| Field-level accuracy | How well the model reads | How many documents still need a person |
| Document-level accuracy | How many documents were read fully correctly | Whether the data matched your ERP and rules |
| Straight-through rate after extraction | How many documents passed extraction without review | The process work still left after extraction |
| Straight-through rate end to end | How many documents finished the process with no one touching them | Little, which is why it is the hardest number to get |
To test a product, collect a month's sample of your own documents, weighted towards the ones your team complains about, and ask for the end-to-end rate on that set. Then ask to see the screen a person gets when the system is unsure, since that screen decides how long each exception takes.
How should you start with document intelligence?
Most companies are early. According to Eurostat's 2025 survey of enterprises with 10 or more employees, 19.95% of EU enterprises used at least one AI technology, and text mining, the analysis of written language, was the most used at 11.75%. Among large enterprises Eurostat measured 55.03% using AI, against 17% of small ones.
Pick a first document flow where many layouts arrive every week and your team can say which checks each document gets. Orders that already come in by EDI are a poor first choice, because they arrive structured. A few hundred emailed orders in 40 layouts, each checked the same 6 ways, make a good one.
Then choose one of 3 routes:
- Build on a document intelligence service. Azure Document Intelligence or a similar service gives your developers the reading layer. This fits a company with an engineering team, one high-volume document type and a workflow system ready to receive the data. You own the connectors, the validation, the exception handling and the maintenance.
- Buy an IDP tool. You get extraction, a review screen and an export. This fits when the next step after extraction is simple, or already handled well by another system.
- Use AI agents that include document intelligence. The reading is one step of an agent that also matches, applies your rules, works in your ERP and routes exceptions with context. Lleverage builds this route, because the teams we work with lose their hours in the checks and the ERP work after the reading.
Whichever route you choose, plan for the work after extraction from day 1. Write down the rules your team applies, the master data each document is checked against and who decides when something does not fit, and use that list as the scope of the project.
At Lleverage, our forward deployed engineers build the first process with your team, and it typically runs supervised within a few weeks. The agent connects to the ERP you run today, SAP, Business Central, Exact, AFAS or one of the others on our integrations page, without a migration.
Frequently asked questions
What is document intelligence in simple terms?
Document intelligence is software that reads a business document the way a trained person would and returns the important information as structured data. It recognises the text, understands the layout, decides what kind of document it is and pulls out named fields such as supplier, invoice number, order lines and totals, each with a confidence score.
What is the difference between document intelligence and OCR?
OCR turns an image of text into characters and stops there. Document intelligence uses OCR as its first step, then adds understanding: which value is the total, which rows are order lines, what type of document it is. OCR returns plain text. Document intelligence returns labelled fields and tables a system can use directly.
Is Azure AI Document Intelligence the same as Form Recognizer?
Yes. Microsoft renamed Azure Form Recognizer to Azure Document Intelligence in July 2023, with no change to its capabilities or pricing. The full name today is Azure Document Intelligence in Foundry Tools. Microsoft now positions it inside Azure Content Understanding, which adds large language models for unstructured documents and varying layouts.
Does document intelligence enter data into the ERP?
A document intelligence service returns structured data and stops there. Getting that data into the ERP, matching it to customers and articles and applying your rules is a separate piece of work. Some IDP tools export to an ERP. An AI agent goes further and creates the sales order or books the invoice through the ERP itself.
How accurate is AI document intelligence?
Field-level accuracy on clean documents is usually high, but it overstates the result. With 20 fields per document at 98% each, only about 67% of documents come out fully correct. Judge a product by its straight-through rate on a sample of your own documents, measured all the way through the process to the ERP.
Which documents is document intelligence best suited for?
It suits documents that arrive in many layouts but carry the same information: customer purchase orders, supplier invoices, order confirmations, delivery notes and certificates. Long technical documents such as specifications and manuals also benefit, when the system can retrieve the right section and cite it, so a person checks a reference instead of the whole document.
Start with the document flow that costs you the most hours
Pick the flow where your team spends the most time between receiving a document and finishing the work it triggers: usually customer orders or supplier invoices. Count the documents per week, the layouts and the checks each one needs. That gives you the scope and the business case.
Then book a demo with 20 of those documents, including the worst ones. We will show you how the agent reads them and what it does after the reading, so you can judge the result on your own documents.