How to Launch an Internal AI Assistant People Actually Use

Internal assistants fail on content and permissions far more often than on models. This playbook covers curating what it knows, enforcing access correctly, and launching in a way that builds trust rather than spending it, drawing on FISTA Solutions' AI enablement work.

When is this worth doing?

When people are repeatedly asking the same questions in help channels, when onboarding takes longer than it should because information is scattered, or when a knowledge base exists and nobody can find anything in it.

It is not worth doing where the underlying content is bad and nobody will own fixing it. The assistant will surface the problem rather than solve it.

What does the sequence look like?

Step	Purpose
1. Find the real questions	From help channels, not from a survey
2. Curate the corpus	Owned, current, non-contradictory
3. Enforce permissions at retrieval	Against the asking user
4. Require citations	Verifiable answers
5. Launch narrow	To a team, on their questions
6. Assign corpus ownership	Or currency decays

Step 1 — Find the questions people actually ask

Mine the internal help channels, the IT and HR ticket queues, and the onboarding questions new starters ask.

That list is both the scope and the evaluation set. It is considerably more useful than a survey, because people report the questions they remember rather than the ones they ask most.

Count the frequencies. A handful of question types usually dominate, and covering those well is a better launch than covering everything badly.

Step 2 — Curate the corpus

Include only documents that are owned, current, and non-contradictory, covering the questions from step one.

This is where the work is. Indexing everything produces an assistant that surfaces drafts, superseded policies, and personal notes with the same confidence as authoritative content.

Resolve contradictions before launch. Two documents saying different things about the same policy produce answers that vary by phrasing, which destroys trust faster than an obvious failure. See how to overhaul a knowledge base for ai.

Step 3 — Enforce permissions at retrieval

The system should only be able to surface content the asking user could already access, enforced when retrieving rather than filtered afterwards.

This is the most serious failure mode available. An assistant that indexes HR documents and answers a question from one to someone who should not see it is a disclosure incident, and the affected person will tell everyone.

Test it deliberately with users at different permission levels before launch, and again after any change to the corpus or the access model. See what is metadata filtering in rag.

Step 4 — Require citations on every answer

Each answer should show which documents it came from, with links.

Citations do three things: they let users verify in seconds, they let the team diagnose failures precisely, and they calibrate trust. An answer with a source people recognise is trusted appropriately; one without is either over-trusted or dismissed.

Where the system cannot cite a source, it should say it does not know rather than answering. Abstention is the behaviour that sustains trust.

Step 5 — Launch narrow to a friendly team

One team, on their questions, with someone available to hear complaints.

Broad launches to the whole organisation on day one spend all the available goodwill on a version that has not been tuned against real use. People who get a bad answer in week one do not return in month three.

The narrow launch also produces the failures you need to fix, at a volume somebody can actually review.

Step 6 — Assign ownership of the corpus

A named person responsible for currency, with review dates on documents and a process for adding and retiring content.

Without this, currency decays and the assistant starts producing confidently wrong answers from superseded policies. That decline is gradual and nobody attributes it to the corpus, which means it gets blamed on the model.

Enforce review dates in the system rather than in a policy: content past its date is excluded from retrieval until an owner confirms it.

What should it refuse to answer?

Anything where the corpus has no relevant content, anything requiring a judgement rather than a lookup, and anything in areas where a wrong answer has consequences — employment decisions, legal questions, or safety matters.

Make those refusals helpful: say it cannot answer and point to the person or team who can. An assistant that declines usefully is more valuable than one that attempts everything.

How do you sustain use?

By keeping the corpus current, adding content in response to unanswered questions, and telling people when it improves.

Track the questions it could not answer. That list is the highest-value content backlog available, and working through it makes the assistant visibly better in the areas people care about.

Measure whether questions get answered rather than whether the tool gets used. See how to measure ai adoption.

Who needs to be involved?

A corpus owner, an engineer, and document owners from each function whose content is included.

The document owners are the constraint. An assistant built without their participation launches on content nobody is maintaining.

How long does it take?

Four to eight weeks, dominated by corpus curation rather than by building. The technical work is the smaller half.

What are the common failure modes?

Indexing everything. Permissions filtered after retrieval. Broad launch on day one. Answers without citations. No corpus owner. And measuring usage rather than answers.

How do you know it worked?

Questions answered correctly with sources people recognise, permission boundaries respected, unanswered questions feeding a content backlog, and use sustaining months after launch.

What does it cost?

Mostly people's time rather than tooling. The expensive version is the one that stalls halfway and leaves the organisation with neither the old state nor the new one, which is why a narrow first pass beats a comprehensive plan nobody finishes.

Budget the work as an operated change rather than a project with an end date, because most of these need a maintenance tail. See AI total cost of ownership.

What should you do first?

Collect fifty real questions from your internal help channels. That list scopes the corpus, becomes the evaluation set, and tells you whether the content to answer them even exists.

How FISTA Solutions helps

FISTA Solutions runs this work alongside client teams rather than around them: corpus curated to the questions people actually ask, permissions enforced at retrieval against the asking user and tested before launch, evidence produced as the work proceeds, and handover that leaves your people able to continue without us. Delivery runs through AI agents, AI enablement, and forward deployed engineers. The record is 150+ projects for 50+ companies across 12+ countries, with 47% average efficiency gains where measured.

To run this with support, message FISTA on WhatsApp, or read how to improve RAG answer quality.