Summarising confidential documents with AI: how to do it safely

Where the document goes

Whoever uploads a PDF to an AI chat service sends its full content to that service: names, amounts, agreements, medical or criminal data. Depending on the terms, the text is stored, viewed by the vendor’s staff during abuse checks or used to improve models.

For a public press release that is no problem. For a contract, a patient file or a due diligence report it is. Those documents are the longest, and therefore the most likely to end up in a chatbot. See Shadow AI.

What a good summary needs

1. Processing in your own environment. The model runs on a node that belongs to your organisation alone.
2. Source references. Every statement in the summary points to a page or passage, so you can check quickly. See AI on your own documents.
3. A fixed instruction. A standard prompt per document type (contract, judgment, report) gives consistent summaries.
4. Human review. Language models sometimes leave something out or add something. See AI hallucinations.
5. Follow permissions. Whoever may not open a document may not see its summary either. See Access control for AI.

Long documents and stacks of files

A model can only process a limited amount of text at once, its context window. For a file of hundreds of pages it works better to summarise per chapter or per document first and then combine those summaries. An automation can do that for a whole folder at once, and prepare the result for whoever handles the file.

Summarising with werqly

With werqly, the language model runs on a dedicated node in your server room or in a dedicated rack. Documents are processed there and do not leave your environment. In werqly AI Control you decide per role which sources are available, and the audit log records who had which document summarised.