Learn about AI · Document AI
How does AI read invoices and PDFs?
From inbox to draft, step by step
- Receive
The invoice arrives as a PDF, an image or an e-mail in a fixed mailbox.
- Read
OCR turns the document into text with the position of every word.
- Understand
The language model finds fields and line items, whatever the layout.
- Check
Amounts, article numbers and supplier are held against your orders and master data.
- Deliver
If everything matches, it is posted as a draft. Otherwise it goes to a person with the deviation marked.
The third step is the one that has changed. Previously you had to make a template for each supplier: the invoice number is here, the total is there. When the supplier changed layout, it broke. A language model reads the invoice the way a person reads it: it knows what an invoice number tends to look like and finds it whether it is top right or bottom left, and whether it is called "Invoice no.", "Faktura nr." or "Reference".
The fourth step is what makes it safe. The model can misread. An 8 can become a 3 on a bad scan. But if the amount is then held against the purchase order and the article numbers against your item master, the error is caught before it is posted. Extraction without a check is a risk. Extraction with a check is a bookkeeper who never gets tired.
An example: supplier invoices at a manufacturer
Take a company making ventilation components, with 45 employees and about 400 supplier invoices a month from 70 suppliers. They come in 70 layouts. Someone in finance keys them in, looks up the purchase order, compares, and passes the doubtful ones to purchasing. It takes one working day a week, and it is not the part of the job she was hired for.
With document AI, each invoice is read as it arrives. The fields are extracted, the lines matched to the purchase order, and one of three things happens. If everything is within the tolerance the finance manager has set, the invoice is posted as a draft with order number and account filled in. If amount or quantity deviates, it lands with the employee with the deviation highlighted and a suggestion of what is wrong. If the model cannot find an order, it asks purchasing. The employee approves the drafts in one sitting and spends the rest of the day on what she was actually hired for.
- Collect 50 invoices from the 20 largest suppliers
That is the test set. Each is read by the tool and by a person, and the two results are compared field by field before anything goes live.
- Decide what may be posted without a person
"Amount within 1 % of the order and a known supplier" is a typical starting rule. It is tightened or loosened after a month of numbers.
- Build the connection to the finance system
The draft has to land in e-conomic, Business Central or whatever you use, with the fields the bookkeeper would otherwise have filled in. Without that connection the result is a tidy table someone still has to type.
- Measure the errors, not just the time
How many invoices were corrected by a person, and why. That number shows whether the tool needs adjusting or whether the scans are the problem.
What document AI is good at, and what it is not
Good at
- Many layouts from many senders: invoices, order confirmations, delivery notes, contracts
- Line items, not just totals
- Several languages and currencies in the same pile
- Holding what it read against your orders, prices and master data
Not good at
- Bad scans and skewed phone photos; the quality of the answer follows the quality of the image
- Handwriting beyond the simplest
- Deciding whether an invoice is reasonable; that is a decision, not a reading
- Documents where the same number means two things depending on context, with no rule to lean on
What it takes on your side
- Examples: 30–50 real documents of each type you want read, with answers from a person.
- Somewhere for the result to land: finance system, ERP, a spreadsheet to begin with. And a connection to it.
- A rule for what may go through without approval, and who sees the rest.
- Reasonable scans. A multifunction printer at 300 dpi is fine; a photo taken at an angle is not.
- A person who looks at the deviations during the first month and adjusts the rule.
Common misunderstandings
| What we hear | How it actually works |
|---|---|
| It's just OCR, we tried that | OCR turns images into text. What is new is the step after it, where the language model understands the text and finds the fields without a template, and the step after that, where the fields are checked against your data. |
| It has to be 100 % accurate or we can't use it | It won't be 100 %, and neither is a person. The difference is the check: what doesn't match gets caught, and what matches nobody has to look at. |
| We'll need a template for every supplier | Not any more. The model reads the layout the way a bookkeeper does. A new supplier needs no set-up. |
| Our invoices will be sent to a service abroad | That depends on the set-up. Both OCR and the language model can run with a European provider or on your own server. |
What we would advise against
We would advise against letting the tool post without approval from day one, however well it does on the test set. Let it post drafts, count the corrections for a month, then loosen the rule. And we would advise against starting with the hardest documents because they are the most annoying. Start with the 20 suppliers who send the most, and take the rest once the numbers hold.
Frequently asked questions
Which documents can be read?
Invoices, order confirmations, delivery notes, contracts, forms; in principle anything with text. The more consistent the document type is in meaning, the better, even when the layout varies.
Does it work with our finance system?
If the system has an API, yes. e-conomic, Business Central, Dinero and most others do. We build the connection so the drafts land with the right fields filled in.
What happens to an invoice the tool cannot read?
It lands with a person, marked as unreadable, together with whatever the tool did manage to get out of it. It does not disappear, and it is not guessed.
How long does it take to get started?
With a test set and access to the finance system, the first document type can typically run as drafts within a few weeks. Each new document type takes less time than the previous one.
Read next
Want help getting from explanation to practice?
Have a no-obligation chat with us about where it makes sense to start. We reply within 24 hours.