Hand your document processing over to AI, and give your teams back the time spent copying them

Three families of tools can process documents:

• Text recognition (OCR) software, such as ABBYY FineReader, Adobe Acrobat or the open-source Tesseract. It turns a scan or photo into text, but doesn’t understand the content. It is enough to archive documents and make them searchable;
• Data extraction platforms, such as Azure AI Document Intelligence (Microsoft), Google Cloud Document AI, Amazon Textract or the French solution reciTAL. They recognise the type of document and pick out specific information, such as the amount on an invoice or an order number. They are ideal for large volumes of standard documents;
• Large AI models able to analyse text and images, such as GPT (OpenAI), Claude (Anthropic), Gemini (Google) or Mistral AI’s models. They understand a document as a whole, including free text, a contract clause or an unusual layout. They do, however, require checks, because a plausible answer can be wrong.

Most projects combine these tools: an AI model to understand varied documents, an extraction platform for standard flows, and control rules to check the result. The right choice depends on how varied your documents are, their volume and the level of reliability expected.

Fewer than you might think. Recent AI models can already read most common documents, with no lengthy prior training. A few dozen representative examples are generally enough to tune the checks and measure reliability. Remember to include the hardest cases: crooked scans, handwritten notes, unusual layouts. That is where a solution is really judged!

Yes. Assistants such as NotebookLM, ChatGPT, Claude or Mistral AI’s Le Chat let you summarise or query a PDF, often for free. They are suitable for occasional use, on non-confidential documents. To process company documents at scale, they quickly show their limits: no connection to your software, no systematic checking of results, and data usage terms often incompatible with contracts or personal information.

Vendors claim up to 99.5% accuracy for reciTAL and up to 99% at Automation Anywhere. These results are obtained on their customers’ documents. Yours may give different results, because reliability varies with:

• the quality of the scan or photo,
• the orientation and resolution of the document,
• the legibility of the writing, especially handwriting,
• the variety of layouts from one supplier to another.

Measurement focused on the useful information. An overall error rate can be misleading. In the example given by Microsoft, an AI that gets 2 words wrong out of 30 shows only a 6% error rate. But if those 30 words make up 10 customer names, 2 names out of 10 are wrong, a 20% error rate. At Kreante, we therefore measure reliability on the information you actually use: amount, reference, due date.

A confidence threshold so nothing slips through. Every piece of information picked out by AI comes with a confidence score. You just need to set a threshold:

The threshold is set according to the cost of an error: stricter for bank details than for filing a letter.

No. Your ERP, your CRM and your document management tool stay in place: AI processing connects to them and deposits the information it picks out. For your teams, nothing changes in their habits, except that they no longer have to copy everything.

At Kreante, every project is priced on a custom quote, based on a scope defined together in writing, after a free assessment. The budget depends on five factors:

After go-live come running costs: use of AI models or extraction services, often billed per page, hosting and maintenance.

• Personal data protection: as soon as a document contains names, contact details or information about employees or customers, the GDPR applies. It requires you to pick out only the necessary information, limit how long it is kept and govern the providers who access it, especially when data is processed outside the European Union.
• The European AI regulation: most document processing, such as reading orders or contracts, is not high-risk. However, automatic sorting of job applications and CVs is explicitly classified as such^[2]. For these systems, the obligations have been postponed to 2 December 2027^[3].

At Kreante, these points are identified at the scoping stage, without replacing a lawyer’s opinion when your situation calls for it.

The market brings together vendors of ready-to-use platforms, the services of the major cloud players and integrators who design custom solutions.