Short answer
Knowledge management AI can help people find, organise and reuse work knowledge, but a useful purchase starts with a specific knowledge problem. Evaluate whether the system can show the source, scope, owner and review date of an answer, and whether it fits how knowledge is actually captured and corrected. A polished answer without those controls is difficult to trust or maintain.
Start with the knowledge problem
Do not start a buying process with “we need an AI knowledge base.” Write one work question that people currently struggle to answer. For example: “How does a store supervisor document an unresolved return at shift handover?” or “Which approved procedure applies when an operational policy changes?”
The answer may already exist in a controlled document. In that case, search and source retrieval are the main needs. If the useful practice is still held by experienced people, the task is different: capture a bounded example, validate it and decide whether it should be shared. The UK National Archives' knowledge principles make the same practical point: not all knowledge can or should be captured, and captured knowledge needs ownership and refreshing.
A buyer worksheet
Use this worksheet in a vendor demonstration or pilot.
| Question | Evidence to request | Decision it supports |
|---|---|---|
| What answer must users find? | Three real, permitted questions | Whether retrieval is the actual problem |
| What is the approved source? | A clickable source, its owner and date | Whether an answer can be checked |
| Who can see it? | Roles and a test user journey | Whether the answer respects intended access |
| What happens when it is stale? | An owner, review date and correction workflow | Whether knowledge can stay useful |
| What cannot be inferred? | A demonstration of uncertainty or missing source handling | Whether the tool overstates its evidence |
| How will the pilot be judged? | A small task and human review of results | Whether the workflow helps real work |
Ask to see the system answer a question whose source has changed and one that has no approved answer. A credible response may point back to the source, identify uncertainty or send the question to an owner. It should not invent a policy.
For a repeatable demonstration, use the organizational-memory evaluation pack. It provides a sample collection, conflicting-version tests and a result record you can reuse across options.
Fictional example: choose retrieval before capture
Fictional example: a retail operations team wants a faster answer to “which exception code applies to a damaged delivery?” The current policy is written, but staff cannot locate it during a shift. The buyer tests three synthetic questions against the approved policy, checks whether the answer links to the relevant section and asks the process owner to correct one deliberately outdated paragraph.
This pilot does not attempt to map employee capability or create a new people record. It tests a narrow retrieval job. If the team later discovers that the written policy does not explain the practical handover between warehouse and store, it can run a separate capture exercise with people who do that work.
Compare the operating model, not just the interface
Different categories can be appropriate:
- Search and retrieval tools suit approved documents, policies and records that people already need to find.
- Knowledge-base tools suit articles that have editors, publication rules and a regular review cycle.
- Meeting or project records can preserve decisions when the organisation has a clear record owner.
- Knowledge-capture workflows suit a specific practice that needs context, validation and a reusable form.
The category label does not establish freshness, permission or quality. NIST's AI Risk Management Framework is a voluntary reference for considering AI risks through design, use and evaluation. For a buyer, the practical application is simple: inspect the source path, test the output and assign a human owner before relying on it.
Where Lontra fits
Lontra is relevant when a team needs to gather contextual accounts of work before a human decides what practice is useful to retain or revisit. It supports focused employee conversations and manager briefs; managers receive briefs rather than raw conversations. It is not presented as a replacement for an organisation's document repository, HRIS or general enterprise-search stack. Explore the product.
Frequently asked questions
What should a knowledge management AI pilot prove?
It should prove that people can find a permitted, traceable answer for a defined work question and that an accountable owner can correct or retire it.
Is knowledge management AI only enterprise search?
No. Search is one use case. Other products support capture, curation, reuse or learning, but each needs a clear source, owner and review process.