AI document processing: how an intake agent reads your mail and enters the data

AI document processing uses AI models to read documents the way a person would: scans, PDFs, photos, faxes and email attachments. It pulls out the details you need, checks them against your own systems, enters them, and sends anything unclear to a person. Unlike older OCR, it handles new layouts, messy scans and handwriting without a template for each one.

How does a document intake agent work?

Six steps, from the moment a document arrives to the moment the data is in your system:

1. ArrivesAn email attachment, a scan from the mailroom, a fax, an upload.
2. SortsWorks out what it is: an invoice, a purchase order, a legal notice, or junk.
3. ReadsPulls out the fields you care about, even from crooked, faint or handwritten pages.
4. ChecksMatches against your data: does the customer exist, does the PO match, do the totals add up?
5. AsksAnything it isn't sure about goes to a person, with the page and its reading side by side.
6. EntersCreates the record in your system and files the original, with a log of what it did.

The checking step is what separates a useful intake agent from a text reader. Reading a number off a page is easy now. Knowing the number is wrong, because the PO doesn't exist or the price doesn't match your contract, is where the work is saved.

Which documents is it good for?

• Invoices and bills, matched to purchase orders and receipts.
• Purchase orders from customers, in whatever format each one sends.
• Legal letters and notices, sorted by type, with deadlines and parties pulled out and routed to the right person.
• Insurance forms and certificates, checked for coverage dates and limits.
• Applications and intake forms, including scanned and handwritten ones.
• Shipping paperwork: bills of lading, packing slips and delivery receipts.
• Remittances and payment advices, matched to open invoices.

The best first candidate is a document your team types in by hand every day, in several layouts, where a mistake costs something.

What is the difference between OCR, intelligent document processing and an AI agent?

Intelligent document processing (IDP) is the broader category. Newer IDP products use the same AI models and are closing the gap. The question that matters is whether you need fields extracted, or the whole job done.

How accurate is AI document processing?

It depends on your documents, so measure it on them. Take 50 to 100 past documents where you already know the right answers, including the ugly ones, run them through, and score each field. That gives you a real accuracy number per field before anything goes live.

In production, accuracy comes from checks, not from trusting the reading. A total that doesn't add up, a PO number that doesn't exist or a customer it can't match sends the document to a person. The agent never guesses to fill a gap.

What does a person still do?

Review the exceptions. The agent shows the original page next to what it read, highlights what it's unsure of, and the reviewer approves or corrects in one step. For the first weeks, run it as a practice run where a person approves everything, then let the clean cases go through on their own once the numbers hold up. Corrections become new rules, so the review pile shrinks.

Is document intake an agent or a workflow?

Most of it is a workflow with an AI step: the reading needs AI, and what happens after is fixed. It becomes an agent when the follow-up needs judgment, such as matching a PO to the right customer and price, chasing a sender for a missing page, or deciding who on the legal team a notice goes to and by when. Start with the workflow and add judgment where the exceptions are.

Should you buy IDP software or build your own?

Buy when you handle a high volume of one standard document, like supplier invoices, and the product connects to your accounting or ERP system as it ships.

Build when the documents are mixed and every sender's layout is different, the data goes into older or custom software, the checks are your own business rules, or the follow-up work is the real cost. Building also means the rules and the data stay yours.

What do you need to get started?

• 30 to 100 real documents, including the worst scans you get.
• The list of fields you need and where each one goes in your system.
• The checks your team does today, even the ones nobody wrote down.
• One person who reviews exceptions and owns the results.

What about sensitive documents?

Legal, insurance and customer documents often hold personal information. The agent should reach only the systems its job needs, keep originals where your retention rules say, log every action and send nothing to AI providers that train on your data. Unusual or sensitive cases go to a person. Our guide to keeping people in control of AI agents covers the approval rules.

Common questions

Can AI read handwriting?

Often, yes, including handwritten notes on printed forms. Handwritten fields are the ones to check most closely in your test set, and the ones most likely to go to a person.

Can it handle faxes, photos and crooked scans?

Yes. Pages are straightened and cleaned up before reading. Very faint or cut-off pages get flagged instead of guessed.

What about paper mail?

Paper has to be scanned first, by a scanner in the mailroom or a mail scanning service. From there it works like any other document.

Does it replace our ERP or accounting system?

No. It feeds the system you already use, through its import, its API or, for older software, the same screens your team uses.

How long does a first intake agent take?

Our First Agent Sprint builds a first working version on your real documents in two weeks, running as a practice run your team approves.