Why this is the best first AI project
Internal, low-risk, immediately useful, and it exposes your documentation problems before you build anything customer-facing. Teams who start here learn what their content is actually like — which is knowledge every later AI project depends on.
Every organisation has documentation nobody can find. This is the most reliable AI deployment we see, and the one with the fewest ways to go wrong.
Why traditional internal search disappoints
| Problem | Why it happens |
|---|---|
| Must know the right words | Keyword matching, not meaning |
| Returns documents, not answers | You still read to find the line |
| Cannot combine sources | Answer spans two documents |
| Ranking is poor | Little signal about relevance |
| Outdated results surface | Superseded documents still indexed |
| Content is scattered | Wiki, drive, email, tickets |
The gap is not that people cannot search. It is that they must already know what the document calls the thing they are looking for — which is precisely what a new employee does not know.
What good looks like
- Answers the question directly, then cites the source document and section.
- Combines information from multiple documents where the answer requires it.
- Says when it does not know rather than assembling something plausible.
- Respects permissions — surfaces only what that person may see.
- Lives where people work — Slack or Teams, not a separate portal.
- Shows recency, so users can judge whether the source is current.
The documentation audit that comes first
Do this before commissioning anything. Take your twenty most common internal questions and try to answer each from existing documentation alone. Whatever you cannot answer, the AI cannot either — and the exercise frequently reduces question volume on its own, sometimes enough that the AI project becomes unnecessary.
What the audit typically reveals:
- Contradictions — two documents stating different policies, both current.
- Stale content — superseded documents never removed.
- Undocumented knowledge — the answer exists only in someone's head.
- Fragmentation — the answer requires three documents nobody links together.
- Access confusion — nobody is sure who should see what.
Every one of those becomes an AI failure if left unaddressed. Fixing them is the actual project.
Sources, in order of value
| Source | Value | Difficulty |
|---|---|---|
| Policy and process documents | High | Low |
| Internal wiki | High | Low |
| Product and technical documentation | High | Low |
| Resolved support tickets | Very high | Moderate — needs curation |
| Shared drive documents | Mixed | Moderate — much is stale |
| Email archives | Low | High — noisy and sensitive |
| Chat history | Low | High — conversational fragments |
Resolved support tickets are the most undervalued source. They contain real questions in the words people actually use, paired with verified answers. Curated properly — the confirmed resolution, not the whole thread — they are frequently better grounding material than the formal documentation.
Permission-aware retrieval
The technical requirement that most affects architecture:
- Tag every chunk with the access groups permitted to see its source.
- Resolve the requesting user's groups at query time.
- Filter before ranking, so restricted content never enters the candidate set.
- Never rely on the model to withhold content that was placed in its context.
- Re-check on access changes, so departures and role changes take effect.
Measuring whether it works
- Questions asked per week — usage is the first signal.
- Answered versus refused — a high refusal rate points at documentation gaps.
- Repeat questions — the same query twice suggests the first answer failed.
- Source coverage — which documents are actually being cited?
- Questions with no good answer — this list is your documentation backlog, generated automatically.
That last point is a genuine secondary benefit. The logs tell you exactly what your organisation cannot answer about itself.
A deployment sequence
- Audit documentation against your twenty most common questions.
- Fix the contradictions and remove stale content. Measure question volume after.
- Index a narrow, high-quality subset — one department or topic area.
- Deploy in Slack or Teams where people already are.
- Read the logs weekly — what was asked, what failed.
- Expand coverage guided by the gaps the logs reveal.
Staff repeatedly asking questions your documentation should answer? Tell us what they ask most — the audit alone is frequently worth doing. See our AI agent service, RAG architecture, and choosing a deployment channel.