The channel the team already writes in
The assistant is in the same window the team writes to each other in. A question gets asked there in the middle of a conversation the team is having anyway.
Maciej Nuzia · AI assistant
I build an AI assistant that answers a team’s everyday questions
If the same questions keep going back to a few experienced people, an AI assistant can prepare some of the answers from company documents and procedures. It can work in a tool the team already uses, with a scope limited to questions from one team. I describe the search mechanism on the RAG for business page.
I start a rollout with one group inside the company, and with the questions that group actually asks.
Interruptions
A repeated question interrupts two people
Somebody looks up and asks about the warranty terms on one particular client. Whoever happens to sit nearby and happens to know answers them. Two minutes later both are back at their own tasks and both have to read their way in again. Others simply put down whatever seems right and move on.
- the same question reaches one person over and over, from a different colleague each time
- whoever works remotely writes and waits for someone to notice
- whoever is nearest answers, so the same thing gets answered two ways
- nobody in the company knows how much time goes into this
Who it is for
For teams where knowledge sits with a few people
I start by asking who fields the most questions and whether that sits anywhere in their job description. It is usually the same person, the one who has been around longest, and they do it alongside their own work. That person is the reference point here. If the assistant takes anything off anyone, it shows on them first.
- Teams spread across several sites
- Companies running shifts
- Departments with high turnover
- Internal IT helpdesk
- HR fielding employee questions
- Production and service with procedures of their own
Through the day
Where an assistant helps during the working day
Whoever shows a new hire around
A new hire has somebody who answers everything for the first few weeks. Some of those questions can go to a machine at any hour, and nobody has to check whether that person has a minute to spare. What stays with them is everything nobody ever wrote down.
Someone who has been here for years
This person asks about the exception: the client on a different addendum, the rule that changed last month. They ask rarely, and only when they genuinely do not know. If the assistant only handles the easy question, they will try twice and stop. The rest of the team watches what they do.
Questions asked quietly
There are things nobody puts to a manager. How to report broken equipment, what to do about a clash in the team, whose approval a training course needs. A machine does not judge the person asking, and that is the whole difference here.
Evenings and the night shift
Somebody finishing a proposal on a Sunday will not message their manager at that hour, so the question waits for Monday or gets guessed at. The night shift has had this problem forever.
The person everybody writes to
Every company has one. Their inbox is the best list of topics anyone in the building could gather, and I ask to see it early on. Some of those messages repeat closely enough that the answer can arrive without them.
When the questions are meant to arrive from outside, from customers, the project looks different and I describe it under AI customer support automation. Sometimes the assistant is one screen inside an application written from scratch, and that project belongs under AI web applications.
Among people
Launching it is not enough. The team has to use it
The first part is engineering and it closes on the day the answers come out right. The second part happens outside the code and comes down to what people actually do with it day to day.
01
Who is going to ask, and about what
I sit down with whoever runs the team and pick the group that asks most. I write down what those people carry around today and who they carry it to. That is enough to know whether there is anything here to take over, and to say so if there is not.
02
The questions in their own words
I collect questions the way they actually get asked, abbreviations and in-house names included. "Has Smith still got SLA on that old contract?" looks nothing like the same question written out formally, and the assistant gets the first version.
03
Where the assistant stands
It goes where the team writes every day, with no separate login. Who introduces it to people is worth settling too, along with the words they use. That part is done by somebody inside the company, with a name and a face.
04
Who shows it to the next group
The second group gets the assistant from somebody in the first. That person already knows what is worth asking, and will say it in ordinary language. I ask the first group which of them wants to take that on.
Technologies
The assistant works in a tool the team already knows
A new browser window is one more thing to remember, so I begin where people already write to each other. An interface of its own comes later, when somebody genuinely needs one. The rest of the list is plumbing nobody on the team will ever see.
- Slack / Microsoft Teams
- OpenAI API
- Claude / Anthropic
- PostgreSQL
- Python
- Node.js
- AWS
A month in
An assistant can work correctly and still stop being used
The channel that went quiet
Nobody announces it. One week the questions simply stop coming. Behind that there is usually one sentence said over coffee by somebody the team listens to, and that sentence travels faster than any answer.
An answer that was true six months ago
I ask whose task list carries the job of keeping this current, and what happens to that duty when the person goes on holiday. The answer that comes back is sometimes "somebody on the team", and then I ask again, this time for one particular person.
One mistake in front of the wrong person
A bad answer sometimes lands with somebody who knows immediately that it is bad. That person tells the rest of the team, and does it the same day. Winning their trust back takes longer than correcting the answer, so I take the first test questions from the most experienced people on the team.
FAQ
Keeping answers current and getting the team to use it
Is this not just RAG?
Underneath, usually yes. RAG is the mechanism, and how the model gets to the document and what comes back alongside the answer is written up under RAG for business. This page covers what happens around it: whether anybody in the company asks a second question after getting an answer to the first. A faultlessly built RAG can sit underneath an assistant that nobody has opened in a month.
Who in the company keeps the answers current?
Somebody with a name, who has that duty written into the job. Bolting it onto a person's evenings ends the way every other job without an owner ends. Material ages faster than anyone assumes at launch, and an answer from six months ago sounds exactly as confident as yesterday's. When no such person is anywhere in sight, I say so before the estimate, because without them the rollout has a shelf life.
Will the team stop talking to each other?
The questions that disappear are the ones with a single correct answer, which people asked reluctantly anyway. Where a case needs somebody's judgement the assistant goes quiet, because it has nothing to judge with. One thing is worth watching separately. People sometimes take a question to a machine when what they need is a decision from a manager. That puts the problem in the team, and the tool has merely made it visible.
Who is allowed to ask what?
That is a board decision, not a configuration setting. The person those questions go to today filters them in their head, and nobody ever asked them to write the filter down. The assistant needs a list: who belongs to which group, and which group may ask about salary bands or disciplinary cases. I ask for that list before anybody gets access.
The question list
Start with the questions that come up most often
Their first name and what people bring to them are enough to open the conversation. Those two things already show which part of the questions can go to a machine and which part stays with a person, whatever gets built. At this stage I am not collecting any list of documents.