The problem with retyping
In many offices, someone reads a document and transfers the details into another system. Invoice number, amount, delivery address, contract end date. The work is monotonous, error-prone and ties up people who really have other things to do.
This is exactly where AI is particularly strong today. It recognises the relevant details in a document and stores them as clean data in the right system. Afterwards, a person only checks the cases where the system is uncertain.
Which documents are suitable
In principle, almost any document that contains information in a recognisable form can be extracted. The most common are:
- Supplier invoices and credit notes
- Delivery notes and order confirmations
- Purchase orders received by email or as PDFs
- Contracts with terms, deadlines and termination clauses
- Forms, applications and inspection reports
- Job application documents and CVs
The more often a document type occurs and the more details are transferred from it, the greater the benefit.
Text recognition versus understanding: the difference
Classic text recognition
Text recognition, often called OCR, turns an image into text. But it does not know what that text means. Classic solutions therefore work with templates: for each layout, you define where the invoice number is located. If a supplier changes its layout, the template no longer fits.
AI with language understanding
Modern language models understand the content of a document. They recognise an invoice number even if it sits in an unusual place or is labelled differently. They can summarise line items, spot connections and answer questions about the document, for example whether a contract contains an automatic renewal clause.
In practice, the two methods are often combined: text recognition supplies the text, and the AI understands and structures it.
Handwriting, scans and poor photos
Not every document arrives as a clean PDF. Delivery notes with handwritten notes, receipts photographed at an angle or old faxes are part of everyday life in many businesses. Current models cope with such documents far better than earlier methods, but not without errors.
That is why it is important for the system to recognise when it is not sure. An illegible amount should not be guessed but passed on for review. It also helps to improve the quality of what comes in, for example by using a simple scanning app instead of handheld photos.
Checking against your rules
Extraction alone is not enough. The real value comes when the recognised details are checked against your rules and your data:
- Does the sum of the line items match the total amount?
- Is there a purchase order and a delivery note for this invoice?
- Is the supplier known, and do the bank details match the master data?
- Has this invoice number been recorded before?
- Does a contract end within the next few months?
Such checks are rule-based and reliable. They catch many errors before they reach your systems.
Confidence levels and queries
A good system works with confidence levels. If it is certain about all details and all checks pass, the case goes straight through. If it is unsure about a detail or a check fails, the case goes to a person with an explanation. That person sees the document, the recognised values and a note on what does not match.
At the start, it is advisable to set strict thresholds and have more cases reviewed. As experience grows, the share of cases that run through without intervention can be increased step by step.
Handover to your systems and data protection
The recognised data has to arrive where it is needed: in accounting, the ERP, the CRM or the document archive. The cleanest way to do this is via interfaces. Where none exist, data can also be handed over as a file.
Documents often contain personal data. So clarify where the processing takes place, whether a data processing agreement (Art. 28 GDPR) is in place and whether your content is used to train third-party models. With the Scalient Group, processing takes place in Frankfurt, and your content is not used to train third-party models.
Conclusion
Today, AI reads invoices, delivery notes, contracts and forms reliably enough to largely replace retyping in many offices. What matters is checking against your rules, clear thresholds for queries and a clean handover to your systems. Start with the document type that occurs most often, and expand from there.