AI in document management: what it adds to your DMS, and what it does not replace

Every document management vendor now adds “AI” to its offer. Behind the word lie very different things: a search box a little more forgiving of typos, a one-click summary of a document, or the ability to question an entire collection and get an answer you can check. For a records manager or an IT department, the question is therefore not “do we need AI in our DMS?”, but “which kind, for what, and on what terms”.

Two different questions

A document management system is very good at one precise question: where is the document? It was designed to store, file, version, control access and archive. AI asks a different question: what do we know? To answer it, the unit of work changes, from the document to the passage. A DMS can hand you the contract; it can rarely tell you which of your 3,000 contracts contain a clause that matters to you today. We developed this distinction in Document management and AI: two different questions.

The gap shows in time. According to Atlassian’s State of Teams 2025 report, based on 12,000 knowledge workers and 200 executives, a quarter of working time goes into simply looking for answers. It is a self-reported survey, run by a vendor of collaboration tools: an order of magnitude, not a measurement.

What AI really adds

Five contributions deserve the name, and a demonstration should focus on them.

Searching by meaning, and by exact terms. Search by meaning finds a passage even when it does not use your words. Search by exact terms finds what meaning misses: an article number, a proper name, a standard’s reference. Good systems combine the two, then rerank the results by real relevance.

Answering with citations. Rather than a list of documents to reopen, a written answer in which every statement points to the passage it rests on. It is the condition of any check: a fluent answer inspires trust long before it has been verified. The Hallucination-Free? study, published in 2025 in the Journal of Empirical Legal Studies, showed that professional tools also produced “misgrounded” answers, citing a real source that does not say what it is made to say. Only opening the cited passage reveals the gap.

Reading what the DMS stores without reading. Scanned documents, emails and their attachments, compressed archives: many collections keep them without their content being searchable. Character recognition and email reading bring them into search, just like a native document.

Classifying and extracting. Recognising what kind of document it is (contract, invoice, council decision) and extracting its key information (parties, amounts, deadlines), each value tied to the place where it was read. These are metadata nobody had time to enter, and they can be checked like a citation.

Alerting. Part of what your documents contain should reach you without you asking: a deadline, a tacit renewal, a notice window. It is the rarest function in today’s offers. Ask for it explicitly, and check that every signal arrives with its source.

What the DMS keeps

The DMS keeps its role as the system of record, with its access rights, its versioning, its compliance and the evidential value of its documents. AI comes on top of it as a knowledge layer that reads, indexes and connects what already exists. This approach avoids a migration project, respects past investments and makes the IT department an ally rather than an obstacle.

It also assumes usable documents. In February 2025, Gartner reported that 63% of organisations do not have, or do not know whether they have, data management practices suited to AI, and forecast that many projects not backed by AI-ready data would be abandoned. The forecast remains to be confirmed; it is a reminder that a well-kept collection is an asset, not a detail.

Six questions to ask before choosing

1. Access rights. Does the solution carry over your DMS’s rights, and at what granularity: per document, per folder, per workspace?
2. Updates. What happens when a document changes or a new version arrives: automatic reindexing, or a new upload?
3. Citation. Does every answer point to the exact passage, in the version of the document that was read?
4. Formats. Scans, emails and attachments, spreadsheets: what is really read, and what is only stored?
5. Hosting. Where do the documents, the index and the models live, and under which law?
6. Exit. What becomes of the index and the extracted information if you change tools?

No answer is disqualifying in itself. What matters is that the answers are clear before you sign, and that they match how you work.

Start small

Rather than a demonstration on documents chosen by the vendor, take a well-defined collection: one department’s contracts, a year of council decisions, the notes of a project. Gather a dozen questions your teams really ask. Then measure two simple things: is the answer right, and does its source check out in one click? Within a few weeks, you will know what AI changes for your collection, and on what terms.