← All posts
· Tangle Web Services · 4 min read

The unglamorous work of document AI

Document AIAutomation
Isometric illustration of paper invoices and contracts flowing through a scanner and turning into structured data and a database

Nobody puts "extract fields from a PDF" on a conference slide. It's not exciting. But if I look back at the projects that paid for themselves fastest, document AI is near the top of the list almost every time.

The reason is simple. A huge amount of small business work is reading a document, finding a few pieces of information, and typing them somewhere else. Invoices into accounting. Application forms into a database. Contract terms into a tracking sheet. It's slow, it's mind-numbing, and people make mistakes when they're bored. That's a perfect target.

What document AI actually does

Strip away the marketing and document AI does three things, in order:

  1. Reads a document, including scans and photos, and turns the image into text it can work with.
  2. Finds the specific pieces you care about. The invoice total, the due date, the vendor name, the line items.
  3. Hands off that structured information to wherever it needs to go, like your accounting software or a spreadsheet.

The first step has been around for years under the name OCR. What's changed recently is the middle step. Older systems needed every invoice to be in the exact same layout. Move the total an inch and they'd break. Modern models can read a stack of invoices from forty different vendors, each laid out differently, and still find the total. That flexibility is the whole reason this got useful.

Where it earns its keep

A few patterns we've built more than once:

Accounts payable. Invoices arrive by email in every format imaginable. The system reads each one, pulls vendor, amount, date, and line items, and drops them into the accounting system as a draft for someone to approve. A task that took a bookkeeper a couple of minutes per invoice becomes a quick glance and a click.

Intake forms. Anyone who processes applications, whether for loans, rentals, services, or memberships, knows the pain of retyping the same fields from a PDF. Document AI does the typing. The human reviews.

Contracts and renewals. Pulling key dates and terms out of a folder of signed contracts so nothing renews or lapses without you knowing. This one quietly prevents the expensive surprise of an auto-renewal you forgot about.

The common thread: high volume, structured-enough information, and a real cost when it's done slowly or wrong.

The number that matters: accuracy on your documents

Vendors love to quote accuracy figures. "99% accurate!" Ignore those. The only accuracy that matters is on your documents, with your vendors, your handwriting, your scan quality.

So before committing, we run a test batch. Take a few hundred of your real documents, run them through, and check the results by hand. This tells us the truth: maybe it nails clean digital invoices but struggles with the faxed ones from that one supplier who's still in 2003. Now you know where to keep a human, and you've got a real number instead of a brochure promise.

That test batch is non-negotiable in how we work. Anyone who'll set up document AI for you without testing on your actual documents is guessing.

Keep a human where it counts

Here's a position I'll defend: for anything involving money or legal commitment, a person should approve before the action is final. Not review a report afterward. Approve before.

That doesn't slow things down the way you'd think. Approving an invoice the system already filled in correctly takes a few seconds. The human is a safety net, not a bottleneck. And the day the system misreads a 1 as a 7 on a five-figure invoice, that few-second glance is what saves you.

For lower-stakes work, like sorting documents into the right folder, you can let it run on its own and spot-check. Match the oversight to the cost of being wrong.

What it struggles with

I'd rather you hear this from me than discover it in production.

  • Bad scans. Garbage in, garbage out still applies. A crooked, low-resolution photo of a crumpled receipt will give worse results than a clean PDF. Sometimes the real fix is upstream: ask people to submit documents a better way.
  • Handwriting. It's improved a lot, but messy handwriting is still the hardest case. If your work is heavy on handwritten forms, test that specifically and expect to keep more humans in the loop.
  • Truly unusual layouts. A document type the system rarely sees will get worse treatment than your common ones. Volume helps accuracy.
  • Judgment calls. "Is this expense legitimate?" is not an extraction problem. The system can pull the data; deciding what it means is still a person's job.

What it tends to cost

This varies, but document AI is usually priced per page or per document processed, which makes it pleasantly easy to predict. You can estimate your monthly volume, multiply, and know roughly what you're signing up for. For most small businesses we've worked with, the processing cost is small next to the labor it frees up. The bigger cost is the upfront setup and testing, and even that's modest compared to a full custom build.

Is it right for you?

Quick gut check. If your team regularly retypes information from documents, and that work is measured in hours per week rather than minutes per month, document AI is worth a serious look. It's one of the few AI projects where the payback is concrete enough to put on a spreadsheet before you start.

If your document volume is low, or every document is a unique snowflake requiring real judgment, the math may not work, and I'd tell you so.

Want to find out which camp you're in? Send us a description of the documents you wrestle with and roughly how many per week. We can usually tell you within a conversation whether it's worth a test batch. Get in touch.