Invoice digitisation with OCR and AI, from a case in production

Invoice digitisation with AI means a scanned document turns itself into a clean row of data, with nobody typing amounts or dates. The technology for reading paper has existed for decades. What changed is what comes after, because the bottleneck was never reading the invoice. It was being able to trust what was read.

What changes compared to plain OCR

OCR, optical character recognition, turns an image into text. That solves the easy half. A supplier invoice can arrive in a hundred layouts, with the total wherever each issuer chose to put it and line items that never match across companies. That is where the language model comes in, interpreting the text the way an experienced clerk would. It finds the amount even when it moves, understands that two differently worded concepts are the same charge and lands every field in its column. OCR reads. The model understands what was read.

How it looks in production

For Stanton, a property manager, the tenants’ electricity, gas and water invoices used to be processed by hand. Each document arrived in its utility company’s own layout and someone turned it into usable data, invoice by invoice. Today those invoices arrive through a Telegram chat that acts as the inbox, OCR reads them, the model structures them and each one lands as a normalized row in the spreadsheet the team already used. Two AI agents in production. And the client keeps extending the automation to more back-office processes, which is the signal that the first piece worked.

That last detail matters more than it looks. The administrative automation that survives does not arrive as one grand project that changes everything. It arrives process by process, starting with the one that hurts most.

Validation is the actual product

Extracting data is the showy part. The part that decides whether the system deserves trust is validation, the checks that run before any figure is accepted. Totals that must square with the line items. Dates that must be possible. An empty field that gets flagged as a doubt instead of passing in silence.

One of our own stories explains why we insist so much. In one of our systems, a double timezone conversion shifted every entry time by two hours, without a single visible error. It was caught by comparing against the source document, not because anything failed. Since then the house rule is that the silent error is the enemy, because a badly extracted figure that looks well extracted propagates into every decision made with it. Doubtful cases go to a human review queue with an owner. The rest pass on their own.

Suppliers, delivery notes and whatever comes next

The same pieces serve the rest of the paperwork. Supplier invoices with their hundred templates, delivery notes that must be matched against orders, scanned forms that feed a case file. Each document type has its nuance, but the pattern repeats, reading, interpretation, validation and a clean row in your systems. Which is why the useful conversation is not “digitise everything”. It is choosing the first process, measuring it and growing from there.

What to ask before you hire

Three questions separate a pretty demo from a system that holds. What happens when a layout the system has never seen arrives, does it break silently or fall into the doubts queue. Where does the data land, in your systems and your accounts or in a third party’s platform. And which checks run before a figure is accepted, because extraction without validation is just fast typing with extra steps.

The yardstick is still the one from the opening. Can you trust what was read without looking at the paper? The day the answer is yes, that process has left your list.

If your paperwork looks like this, see how we approach it in AI workflow automation or go straight to the accounting firms vertical. And if you want to know why we always start by ordering the data, that story is in data before prompts.