Building a Company Knowledge Assistant That People Actually Use

Part of our guide hub: AI solutions and automation

The idea is compelling: point an AI at everything the company knows, and let anyone ask it a question. Most organisations that build one abandon it within a couple of months, and the reasons are predictable enough to design around.

Why they get abandoned

The documents were already out of date

An assistant is only as good as what it reads. Most companies have a shared drive containing four versions of the leave policy, a pricing document from 2023 marked final, and process notes that describe how things used to work.

Pointed at that, the assistant answers confidently and wrongly. Two wrong answers and people stop trusting it, which is a permanent condition rather than a temporary one.

The uncomfortable implication: a knowledge assistant project is mostly a knowledge cleanup project. Businesses that treat it as a software purchase get a fluent liar.

It answered questions nobody was asking

Teams build the assistant around the documents they have rather than the questions people actually ask. The result answers policy questions nobody was struggling with, and cannot answer the operational questions that occupy real time.

It required a new habit

An assistant that lives on its own page gets opened four times and then forgotten. One that answers inside Slack, Teams or WhatsApp, where people already are, survives.

The permissions problem

This is the one that ends projects, and it is almost never planned for.

If the assistant reads everything, it will eventually tell a junior employee what someone’s salary is, or repeat the contents of a board document to a contractor. It does not understand that some sources are restricted; it only sees text.

The correct design applies permissions at retrieval: the assistant searches only the documents that the person asking is allowed to see. That has to be built in from the start, because retrofitting it means rebuilding the retrieval layer.

The practical shortcut for a first version: include only documents that everyone in the company may read. It sounds limiting and it covers most of the useful questions anyway.

What to build first

Start with one domain where the questions are frequent, the answers are stable, and being wrong is recoverable.

The usual best candidates:

• HR and policy questions. Leave entitlement, expense rules, working hours, onboarding. High volume, genuinely repetitive, and currently answered by one person being interrupted.
• Product and pricing for a sales team. What a package includes, what the current price is, what we do and do not support. High value because the alternative is a salesperson guessing on a call.
• Technical runbooks for support staff. How to resolve a known issue, what the escalation path is.

What to avoid in version one: anything involving numbers that change daily, anything legal or contractual, and anything where a confident wrong answer causes real harm.

The design that works

1. Curate the source set deliberately. Not the whole drive. A defined, current, owned set of documents. Assign an owner to each.
2. Answer only from those sources, and say clearly when the answer is not in them. The most important behaviour in the entire system.
3. Cite the source document with every answer, linked. This lets people verify, which is what builds trust, and it turns the assistant into a way of finding the document rather than a replacement for it.
4. Put it where people already work. Slack, Teams, or WhatsApp for field staff.
5. Log every question asked. This is the most valuable output of the whole project: it tells you what the organisation does not know, which is a gap in documentation, training or process.
6. Provide a way to flag a wrong answer, routed to the owner of that document.

Point three deserves emphasis. An assistant that cites sources is useful even when it is partly wrong, because the person can check. One that answers without citation is either trusted blindly or not at all, and both are bad outcomes.

What it costs

• Build: 500,000 to 2,000,000 PKR for a scoped assistant over a curated document set with permissions and a chat integration.
• Running: usage based model costs plus hosting. For internal use at a few hundred queries a day this is usually modest. See what AI features actually cost per month.
• The real cost: the document cleanup and the ongoing ownership. Budget someone’s time for keeping sources current, permanently. Without it the assistant decays back to being wrong.

How to know whether it is working

• Queries per week, trending. Falling usage after week three is the abandonment signal.
• Proportion of questions answered from sources versus “not found”. A high not found rate is a documentation gap, which is useful information.
• Flagged wrong answers, and how quickly the underlying document was fixed.
• Interruptions to the person who used to answer these questions. The actual outcome you are buying.

When not to build one

If your company is under about thirty people, the person who knows the answer is usually two desks away and an assistant solves a problem you do not have. If your documentation does not exist, write it first; the writing is the value and the assistant is a convenience on top.

The general principle is the same one that decides which AI tools survive ninety days: it has to sit inside work people already do, and it has to be right often enough to be trusted.

Frequently asked questions

What is a company knowledge assistant?

An AI system that answers employee questions from your own documents, citing the source, rather than from general knowledge.

How do we stop it leaking confidential information?

Apply the asking person’s permissions at retrieval so it only searches what they may already read. For a first version, restrict sources to documents everyone can access.

How long does it take to build?

Six to twelve weeks for a scoped first version. Document cleanup usually takes longer than the software.

Why do these projects fail?

Out of date source documents, no permission model, and living somewhere people do not already work. All three are decisions rather than technical limits.

Ezitech builds internal AI assistants connected to company data with proper permission handling. See our AI solutions or tell us what your team keeps asking.