Stop paying a stranger
by the page to read your own files.

The technique underneath every credible document assistant is retrieval-augmented generation, first described by Lewis and colleagues in 2020. The model does not memorize your files. It searches an index of them, pulls back the passages that match, and writes an answer grounded in those passages. That is the good news, because it means answers can be made to cite their source. It is also the part vendors gloss over, because building that index means making a second copy of your entire archive in a form that lives wherever the vendor keeps it.

The usual reassurance is that the index stores mathematical vectors rather than readable text. That reassurance does not survive contact with the research. In Text Embeddings Reveal (Almost) As Much As Text, Morris and colleagues showed that embeddings can be inverted to recover a substantial portion of the original wording. A vector database of your privileged correspondence is not a scrambled version of your privileged correspondence. It is a recoverable one. OWASP lists sensitive information disclosure as a top-tier risk in language model applications for exactly this reason.

For regulated professions this stops being an abstraction. In July 2024 the American Bar Association issued Formal Opinion 512, its first ethics guidance on generative AI. The opinion reads the duty of confidentiality under Model Rule 1.6 to mean that a lawyer who puts client information into a self-learning tool may need informed client consent first, and that the duty of supervision extends to the vendors and models a firm chooses. The same logic runs through federal guidance for other sectors: NIST SP 800-171 Rev. 3 is built around keeping controlled information inside a defined boundary, and the NIST Generative AI Profile treats data leakage and provenance as governance problems you have to answer, not features you can buy.

Then there is the meter. The two largest document AI services publish their rates openly, and both bill by the page: Azure AI Document Intelligence and Google Document AI. Per-page pricing is honest pricing, and it is also the problem. An archive is not a one-time cost under that model. Every reprocess, every reindex, every new question asked of an old file is a fresh charge against documents you already own.

None of this argues that the technology is wrong. It argues that the boundary is the design decision. Run the same retrieval architecture on hardware inside your building and the confidentiality analysis, the audit trail, and the cost curve all change at once. The NIST AI Risk Management Framework is the checklist we work against, and a system you own is simply easier to answer it with.