Build your own AI assistant for your team

An assistant waits until you ask something and only works in the conversation. A workflow starts by itself, for example on an incoming email, and carries out fixed steps in your systems. An agent gets a goal and chooses the steps itself. Your team builds an assistant itself; workflows and agents you have built.

Side by side:

• Assistant. You start the conversation. It reads, writes and thinks along, but does nothing in your ERP or CRM. You copy the result to where it needs to go.
• Workflow. Starts on an event: an email, a form, every morning at eight. Pulls data from one system, lets AI do the work and puts the result in the other. A human checks where it matters.
• Agent. Gets a goal instead of a route, chooses which systems to consult and puts the result to a human.

Watch the words. In Microsoft Copilot, your own assistant is called an agent. In our classification it stays an assistant as long as you start the conversation and it does nothing in your systems. More about the difference in what is an AI agent?, and what agents do in practice on AI agents.

Choose the tool your team already uses, with a business account. An assistant in a tool nobody ever logs into won't get used. If everyone works in ChatGPT, make a custom GPT. If your team works in Claude, make a Project. If your organization lives in Microsoft 365, look at an agent in Copilot.

At heart they do the same thing: you give it a name, instructions and files, and you share the result with colleagues. The differences are in the details, and those change every few months.

• ChatGPT: custom GPT. An assistant with its own instructions and files, which you can share with colleagues.
• Claude: Project. A workspace with fixed instructions and documents, in which you have your conversations. With skills you can also teach Claude a fixed way of working, such as how you structure a quote.
• Microsoft Copilot: agent. An assistant within Microsoft 365. Handy if the files it needs are already in SharePoint or OneDrive.

Whether you can create and share an assistant, and with whom, depends on your subscription and on what your administrator has switched on. Check that before you start. And don't use a personal account: what you put in an assistant is company knowledge.

Choose one clear goal, write the instructions, add only the files it really needs, test with ten real examples and share it with an owner who keeps it up to date. The first version is up and running in an afternoon. It gets good in the weeks after that, by using it and refining it.

1. Choose one goal

One assistant, one task. 'Drafting a quote for the inside sales team' is a good goal. 'Everything for sales' isn't. The narrower the goal, the better the answers, and the easier it is to see whether it works well. Choose a task that comes back every week, and for which someone currently has to give the same explanation over and over.

2. Write the instructions

The instructions are the fixed prompt that goes along with every question. Use the same building blocks as for a good prompt: role, context, task, example and format. Also write down what it does not do, and what it does when it doesn't know. A setup to copy:

You are the quote assistant of [company]. You help the inside sales team draft a quote from a customer request. Only use the prices from 'Price list 2026' and follow the structure of the sample quotes. Never make up prices, discounts or delivery times. If something isn't in the files, put TO CONFIRM in the text. First ask about what's missing from the request. Write in a formal, businesslike and brief style.

3. Add knowledge files

Only what it really needs. For a quote assistant: the current price list, a few good quotes as examples and the general terms and conditions. Not the whole 'Sales' folder. More files doesn't mean better answers: old versions and half-finished drafts actually cause mistakes. Give the files clear names with a date, so you can see which version is in there. And put nothing in it that not everyone with access is allowed to see.

4. Test with real examples

Take ten requests, emails or questions from the past few weeks, remove the names and let the assistant handle them. Put the result next to what you made yourselves. Where does it go wrong? Adjust the instructions and test again. Also test the edge cases: a request for something you don't sell, a question whose answer isn't in the files. Does it then honestly say it doesn't know? Great.

5. Share it and appoint an owner

Share the assistant with the team, with a short explanation: what is it for, what isn't it for, and who is the owner? That owner keeps the files up to date, collects feedback and adjusts the instructions. Put it in the calendar: fifteen minutes every month to check that everything is still correct.

Almost always on one of these three: too much in one assistant, outdated files and no owner. You don't see them in the first week, but you do after three months, when nobody uses it anymore.

Too much in one assistant

The 'sales assistant' that writes quotes, makes LinkedIn posts, answers customer questions and does the weekly schedule as well. The instructions turn into a wall of text full of contradictions, and it does everything a little. Better to make three small assistants that each do one thing well.

Outdated files

Last year's price list is still in there. So is the new one. The assistant picks one, or mixes them. An assistant is as good as its files. Agree on this: whoever changes the price list also updates the assistant.

No owner

Someone made it enthusiastically in a workshop, shared it with the team and went back to business as usual. Six months later nobody knows which version is in it. Without an owner, every assistant dies a quiet death.

And three smaller ones

• Not tested. In the demo it looks good, but with the first real request it makes up a discount.
• Too many assistants. Twenty assistants that look alike, and nobody knows which one is right. Keep a list with the goal and the owner of each assistant, just like the AI register from your AI policy.
• Sensitive files. Anyone allowed to use the assistant can in principle ask anything that's in the files. So put nothing in it that not everyone with access is allowed to see.

As soon as the work no longer fits in the chat. Do you keep retyping the assistant's result by hand into another system, does the work come by dozens of times a day, or is the knowledge too large or restricted per user? Then an assistant is no longer enough, and you need a workflow, an agent or a RAG application.

Five signals:

• You copy between systems. The request comes in by email, you paste it into the assistant and retype the quote into the CRM. A workflow does that in one go, and you check.
• The volume is high. Ten quotes a week work fine through an assistant. Two hundred customer emails a day don't.
• It has to start by itself. An assistant waits for you. If the work has to begin as soon as the email comes in, you need a workflow.
• The knowledge is large or sensitive. Hundreds of documents, or permissions that differ per employee. Read what is RAG?
• You want to measure and look back. What went in, what came out, who approved it? With an assistant in the chat, you have little grip on that.

In the APO method those are build layer 3, workflows with n8n, and build layer 4, custom work such as agents and RAG. We build those, within the AI Partnership, from € 4,500 per month, excluding VAT. What agents do in practice is on AI agents.

A good assistant is never wasted work. The instructions, examples and test cases you made for it are exactly what's needed to build a workflow or agent later. Want to teach your team to build assistants themselves first? We do that in a workshop on your own work.