How to automate document intake: orders, invoices, forms and PDFs

The documents you're actually dealing with

"Documents" sounds like one thing. In practice it's a mix, and each type behaves differently.

• Emails with the details in the body. A customer writes "please send 40 cases of the usual to the Dayton site." No attachment, no form. The useful part is one sentence.
• PDFs made by software. Invoices and orders exported from another company's system. The text is already inside the file, so reading it is the easy part.
• Scans. Paper run through a copier. The file is a picture of text, often crooked, sometimes with a coffee ring.
• Faxes. Still common in some industries. Low resolution, and a stamp often lands on top of the one number you need.
• Phone photos. A driver photographs a signed delivery receipt on the hood of a truck. Shadows, angles, glare.
• Bundles. One email with eleven attachments, three of which matter.

Before you plan anything, pull a sample of real documents from the last month. A few hundred is ideal. Sort them by type and by sender. That pile tells you more about the project than any vendor demo will.

The six steps of intake

When someone says "automate the invoices," they usually picture the reading step. Reading is one step of six, and often not the hardest.

1. Capture. Get every document into one place: the shared inbox, the fax line, the upload form, the scanner by the front desk. Nothing downstream works while a fax sits in someone's tray.
2. Classify. Decide what each document is: an invoice, a credit memo, a new order, a change to an existing order, a referral, a duplicate. One PDF may hold several documents, so this step also splits files.
3. Extract. Pull out the fields you need: names, dates, amounts, line items. This is where text recognition and AI models do their work.
4. Validate. Check every value against what you already know. Does this customer exist? Does this order number match an open order? Do the line items add up to the total?
5. Route exceptions. Anything that fails a check, or that the system isn't sure about, goes to a person with the reason attached.
6. Write to the system of record. Create the order, bill or case in the software you already run, with the original document attached.

Many projects get step 3 working and stop. The result is a tool that reads documents well and still leaves someone retyping, because nothing checks the values or enters them. The work you want to remove lives in steps 4 to 6, and those steps are specific to your business, your records and your software. Our page on document processing shows what the finished system looks like.

How accuracy really works

Vendors quote accuracy figures. Treat them as a starting point. They were measured on someone else's documents, and yours are not average.

What matters is accuracy for each field, on your own documents. A system can read the invoice number perfectly and the ship-to address badly. One score for the whole document hides that.

Confidence is useful, and not enough

Most document AI services return a confidence score for each value they pull out. Microsoft's documentation for its document service explains that a field confidence of 0.95 means the value is likely correct 19 times out of 20, and suggests using the score to decide what to accept and what to flag for human review.

Nineteen out of 20 sounds good until you multiply. On an order with 20 fields, each at that level, you should expect about one wrong value per document. Confidence alone can't carry the system.

Your own records do the heavy lifting

The strongest check isn't a smarter model. It's your data. When the system reads an order number and then finds it in your list of open orders, the value is confirmed far more reliably than any score can confirm it. The same goes for customers that exist, prices that match the agreed list, and totals that equal the sum of their lines.

Here is a made-up example of one invoice moving through the checks.

The order number had low confidence and was accepted anyway, because it matched a real open order. The total had high confidence and was flagged, because the lines don't add up. A reviewer opens it and finds the vendor transposed two digits. The old manual process might have missed that.

Which checks to run, how strict each one should be, and what happens when one fails: those decisions are the real design work. They come out of sitting with the people who handle the documents today, and they're different in every organization.

The review queue is half the system

People will spend part of their day in the review screen, so it has to be fast and it has to say why a document is there. "Total doesn't match line items" is useful. "Low confidence" isn't. A good queue also sends a small random sample of accepted documents to a person. Amazon's human review tool for documents, for example, offers random sampling as a trigger alongside confidence ranges. It's how you find the mistakes the system was sure about.

The messy cases

Multi-page and bundled files. A 30-page PDF may hold an order, a contract and a certificate. Line items that continue onto page two are a common source of missed or doubled rows.

Handwriting. Treat handwritten fields as review-by-default until your own sample shows the system reads them reliably. Checkboxes and signatures are easier, because you usually need to know they're there, not what they say.

New layouts. A vendor changes its invoice template, and accuracy for that sender drops overnight. Microsoft's guidance warns that documents can look similar to people but appear dissimilar to an AI model. A system that tracks accuracy by sender notices in a day, not a month.

Poor scans and photos. When quality is too low to read, asking the sender for a better copy is cheaper than guessing.

Things that aren't documents. Replies, out-of-office notices, marketing, and "see attached" emails with nothing attached. They need sorting out early so they don't clog the queue.

None of these are reasons not to automate. They're reasons the system has to be built around your documents rather than bought off a shelf.

Examples across industries

• Freight paperwork. Bills of lading, delivery receipts and carrier invoices. The system matches each invoice to its shipment, compares charges with the agreed rate, and flags extra fees for a person to approve.
• Referrals. A specialty clinic gets referrals by fax and portal. The system pulls patient details and the reason for referral, checks for an existing record and required items, and creates a task for scheduling. Missing items trigger a request back to the referring office.
• Purchase orders. Manufacturers and distributors receive orders as PDFs and plain emails. The system maps the customer's part numbers to yours, checks price and stock, and creates a draft order for review.
• Insurance forms. Applications, claim forms and certificates arrive in many versions. The system sorts them, extracts the key details, fills the record, and sends incomplete forms back with a list of what's missing.
• Lease abstracts. A commercial real estate team needs rent, dates, renewal options and key clauses from long leases. AI drafts the abstract with a page reference for each item, and an analyst confirms it. This is judgment work, so every lease gets reviewed.

The six steps are the same everywhere. What changes is the checks. See industries for how this looks in specific kinds of organizations.

What to measure

Measure the current process for two weeks before you change anything. Without a baseline, every improvement is a guess.

• Touchless rate: the share of documents entered with no human edits.
• Correction rate by field: how often reviewers change each field. This tells you where to improve next.
• Time to entry: from arrival to record created. Often the biggest win is speed, not labor.
• Queue age: how long exceptions wait. A growing queue means the system is creating work.
• Downstream errors: wrong shipments, payment disputes, rework. These are the mistakes nobody caught.

A made-up example: a distributor receives 300 purchase orders a week, and each takes about six minutes to enter. That's 30 hours a week. If 70% go through untouched and the other 90 take two minutes each to review, review takes about three hours, plus time for spot checks. Run the same math with your own numbers before you decide what the project is worth.

Build or buy

The big cloud providers all sell the reading step. Amazon's service extracts invoice and receipt fields such as totals, due dates and order numbers. Google lists ready-made processors for invoices, expenses and pay slips, among others.

In practice, the systems that work are a mix: buy the reading, build the checking. The checks are where your business rules live, and no vendor knows those rules. Your team does, and the job is getting them out of people's heads and into software.

What most guides miss

Matching is harder than reading. "ACME Corp," "Acme Corporation" and "ACME Dayton" are the same customer. Customer part numbers differ from yours. A system that learns from every match a reviewer makes gets better over time, and reading errors matter less.

Every correction is a lesson. A month of logged corrections tells you exactly which senders, fields and rules need work. Most systems throw that information away.

The sender is a lever. Often a handful of senders produce most of your documents. Asking a big sender for a cleaner format, or a direct electronic feed, can do more than any model.

Keep the original forever. Attach the source document to the record it created. When someone questions a number six months later, the answer is one click away.

Don't chase 100% touchless. The last few percent are strange documents that need judgment. Aim for a fast, pleasant review queue instead.

How we do it

We start with a few hundred of your real documents and sit with the people who handle them today, because the checks that matter are the ones they already run in their heads. Then we build the intake system, the review screen and the connection into your software, and measure it against your baseline. After launch we stay on: senders change their forms, volumes grow, and the correction log keeps pointing at the next thing to improve, so we offer ongoing support to tune and extend the system as your business changes. Our AI systems page explains how we keep people in charge of what matters, and a short conversation is the fastest way to find out what your documents would need.