Extracting documents with AI: what works reliably today
5 min read
Short answer
Today AI reliably reads invoices, delivery notes, forms, contracts and emails, even with changing layouts. It recognises the relevant fields, classifies documents and transfers data into your software. Limits remain with poor scans, hard-to-read handwriting and anywhere an error would be costly, so a person should check the result.
What does this mean for your company? Book an intro call
In many companies, staff read documents every day just to pick out a few details: an invoice number, a date, an amount, a tick in the right box. AI can take over exactly this work very well today. Your team then focuses on the cases where experience is genuinely needed.
What is the difference between text recognition and AI?
Classic text recognition, often called OCR, turns an image into letters. But it does not know what those letters mean. Older solutions therefore needed a template for every layout: the invoice total is at the bottom right. If a supplier changes its layout, nothing matches any more.
AI understands the content. It recognises which number is the total, even when it sits in an unusual place or is labelled differently. That is why extraction also works for senders the system has never seen.
Which documents are suitable?
- Incoming invoices and credit notes, see also Automating incoming invoices
- Delivery notes and order confirmations
- Completed forms and application documents
- Certificates and evidence, for example checked for validity and completeness
- Contracts, to capture terms, notice periods and amounts
- Emails containing orders or enquiries
A typical example from my work: application documents that used to be checked for completeness page by page. The AI reads the documents, compares them against the requirements and flags what is missing. The specialists only review the flagged items.
How reliable is it?
Very reliable with clearly legible documents, but never error-free. A good solution is therefore built so that errors stand out:
- Plausibility checks: do net, tax and gross amounts add up? Does the order number actually exist?
- Flagging uncertainty: fields the AI is unsure about are presented for review.
- A person in the loop: before data has consequences, such as a payment, a person looks at it.
Test with one hundred real documents before deciding. That shows faster than any presentation how well it works for you.
Where are the limits?
Poor scans, stamps over the text, very long contracts with many cross-references, and tasks that require judgement rather than recognition. Whether a clause is unfavourable for you should still be decided by a person.
Where is the data processed?
Documents often contain personal data. Look for processing within the EU, a data processing agreement (Art. 28 GDPR) and a guarantee that your documents are not used for training. What to look for when choosing a provider is covered in Using AI in line with data protection.
Julius Launhardt automates reading and checking documents for small and mid-sized companies, from recognition to handover into specialist software. More on the process automation page.
Frequently asked questions
How accurate is extraction with AI?
With clearly legible documents, most fields are recognised correctly. Only a test with your real documents shows how accurate it is for you. One hundred typical examples are enough for a reliable assessment.
Do I have to set up a template for every supplier?
No, that is the big difference from older solutions. Modern AI recognises fields without a fixed template, even from senders it has never seen before.
Can AI read handwriting?
It now recognises clearly legible handwriting in forms fairly reliably. With hasty notes or stamps over the text, the error rate rises. A person should check such fields.
Where are my documents processed?
That depends on the provider. For documents containing personal data, I recommend processing within the EU, a data processing agreement and a contractual exclusion from model training.