Internal AI assistant for company data

An internal AI assistant answers questions from your own documents and systems — policies, runbooks, contracts, past tickets, product specs — instead of from general knowledge, and cites the source for every answer so the person asking can check it. The problem it solves is not knowledge, it is interruption: the senior person who fields the same twelve questions a week, and the new starter who waits half a day for an answer that exists in a document they did not know about.

It is also the automation with the widest gap between a convincing demo and a system people actually use.

What the manual version costs

Harder to measure than invoices, which is why it is usually under-costed.

Count for one week: how many times someone asks a colleague something that is written down somewhere. Then count the answer-shaped Slack messages from your two or three most-interrupted people. That second number is the real cost — not the minutes, but the fact that your most expensive people are a lookup service, and every interruption costs them the thread of what they were doing.

The other cost is silent: the decisions made without checking, because checking was too slow.

What building one actually involves

Retrieval is the whole game. The model matters far less than what you put in front of it. Getting the right handful of passages retrieved for a given question is where the engineering is, and it is where most internal assistants quietly fail — they return plausible answers assembled from the wrong documents.

Two things make this harder than it sounds. Accuracy degrades as you add more context rather than improving, so stuffing everything in is not a strategy — the same problem that makes long-running agents get worse over time. And your documents contradict each other: the 2024 policy and the 2026 policy both exist, and only one is right.

Permissions have to be enforced at retrieval. The assistant can only search what the person asking is already allowed to read. Designed in from day one; retrofitted at real cost.

Citations are not optional. Every answer names the document it came from and links to it. This is what makes an assistant trustworthy rather than merely fluent, and it is what lets someone catch it being wrong.

Where a human still has to stay in the loop

Differently from the other automations here: nobody approves each answer. The loop is on the content.

The assistant should be allowed to say “I don’t know”, and that path has to be built deliberately, because the default behaviour of a language model is to produce something. An assistant that answers everything is less useful than one that answers most things and admits the rest, because the first gives you no way to know which answers to trust.

Someone also has to own the corpus. Questions the assistant answered badly are the most valuable feedback your documentation will ever get — they point precisely at what is missing, stale or contradictory. Without an owner acting on that, the assistant degrades exactly as fast as your documentation does.

What it costs and how long it takes

Fixed price after a 30-minute scoping call. What moves the number: how many sources have to be connected, how messy the permissions model is, and what state your documentation is in — by a wide margin the biggest variable, and the one least under our control.

Six weeks at most to a first working version answering real questions from real documents. We would normally start with one well-documented domain rather than everything at once, because a narrow assistant that is right is more useful than a broad one that is sometimes right.

When not to build an internal copilot

This is the automation we most often talk people out of.

• Your documentation is out of date. Then you do not have a retrieval problem, you have a documentation problem, and an assistant will confidently serve stale answers faster than anyone could find them manually. Fix the corpus first.
• The answer is a search box. If people mostly need to find a document rather than synthesise across several, good search is cheaper, faster and more predictable.
• The knowledge genuinely lives in people’s heads. An assistant cannot retrieve what was never written down. Sometimes the honest recommendation is to spend the budget on writing things down and revisit in six months.
• Under about twenty people. Below that, asking a colleague is usually still faster than any system.

Havoric is an AI automation and web development agency based in Ahmedabad, India, working with clients worldwide. Internal copilots are part of our AI automation work and normally ship as a web application your team opens in a browser rather than another tool to install.

← All AI automation work