How to build a custom AI assistant without a developer.

"Custom AI assistant" sounds like a development project: a spec, a budget, a developer, three months. It is not. The main business AI platforms now let you package up instructions and documents into a reusable assistant that colleagues can open and use, subject to your plan and sharing settings. If you can write a clear brief and drag files into a window, you can build one.

What you cannot skip is the thinking. The build takes an afternoon; knowing what the assistant should do, and what good output looks like, is the actual work.

This is about a knowledge or drafting assistant: it reads, it writes, and a person decides what happens next. Once you want one that updates systems, sends messages or acts on decisions, the job turns into integration, permissions and governance, and that is the point to bring in technical help.

What a custom assistant really is

Strip away the branding and the simplest versions have the same three ingredients:

• Standing instructions. A written brief the assistant reads every time: what its job is, how to respond, what tone to use, what never to do.
• Reference documents. The files it should answer from: your price list, your policies, your product sheets, your best past examples.
• A name and a front door. So a colleague can open "Quote Drafter" and start typing, instead of reinventing the setup in a blank chat.

An assistant is just a good brief and the right documents, saved so nobody has to set them up twice.

Where to build it

• ChatGPT: custom GPTs and workspace agents. A custom GPT is still the straightforward way to package up instructions and reference material. Where workspace agents are switched on, they go further, using connected apps and running repeatable workflows. What you can share, and with whom, depends on your plan and your admin settings.
• Claude: projects. A project holds your instructions and documents, so every new conversation starts with the same context. Projects can be shared across an organisation on Team and Enterprise plans. Less packaged than a named assistant, but it works well for a shared body of knowledge.
• Microsoft 365 Copilot: Agent Builder. The afternoon build, and the closest Microsoft has to a custom GPT. You describe what you want in plain English and point it at content you already hold in SharePoint, OneDrive, Teams and other approved sources. It comes with the paid Copilot add-on and appears in the Copilot app beside Microsoft's own agents, which makes it the right place to start when your material is documents.
• Microsoft 365 Copilot: Copilot Studio. Same vendor, bigger tool, different place to work. The brief carries over, because you still start from instructions and knowledge, but Studio reaches past the files you point it at and into the wider Microsoft estate: Dataverse, Microsoft Graph connectors, and the Power Platform connectors into the line-of-business systems you already run. It publishes to Teams, to Microsoft 365 Copilot and to a website, and it is governed the way a Power Platform app is, through environments and data loss prevention policies. For licensed users it comes with the same add-on rather than a separate bill, which is not what most people assume; we costed that in Microsoft Copilot for small businesses. The common path stays no-code, but go far enough and you are in Power Automate, which is where it stops being an afternoon.

Those two are easy to confuse, and Microsoft's naming does not help. The distinction worth holding is not the label: it is that Agent Builder is the afternoon this article describes, while Copilot Studio is a build, with an owner and a governance conversation attached. Reach for Studio when the assistant needs something that does not live in a document, and expect your IT people to want to be in the room rather than told afterwards.

Pick whichever tool your team already uses. The skills transfer; the assistants themselves do not transfer cleanly between platforms, which is worth remembering before you build twenty of them in one vendor's walled garden.

A worked example: the enquiry answerer

Say the same ten questions arrive by email every week. An afternoon's build:

• Gather the source material (30 minutes). Your price list, your FAQ, and half a dozen past replies you were proud of.
• Write the brief (30 minutes). "You draft replies to customer enquiries for [business]. Answer only from the attached documents. If the answer is not in them, say so and suggest the customer calls, never guess a price. Match the tone of the example replies. Keep it under 150 words. Always close by offering a call."
• Test it against reality (an hour). Feed it the last twenty real enquiries and compare its drafts with what you actually sent. Use an approved business account, and strip out personal information the test does not need. Keep those enquiries separate from the example replies you gave it, then score each draft for factual accuracy, tone, invented claims, and whether it escalated properly when the answer was missing. Every miss points at a missing source, an unclear instruction, or a limit the assistant will not reliably get past. Fix what you can, write down what stays unreliable, and repeat.
• Share it, with one rule (10 minutes). The rule: drafts get read before they get sent. It is an assistant, not an employee.

What keeps assistants alive

Most custom assistants die quietly of stale documents. Prices change, policy changes, and the assistant keeps answering from last year's files, confidently. Two defences:

• Give it an owner. One named person who updates the documents when reality changes. If nobody owns it, retire it.
• Feed it from a living source. This is the deeper fix: a shared store of what your business knows that connected AI tools can draw on, so you maintain the knowledge once rather than updating a separate uploaded snapshot in every assistant.

Start with one

Pick the single most repetitive, text-shaped job in your business and build an assistant for that. One assistant that a team genuinely uses teaches you more than ten built for the demo. In training we build one live with your team, on one of your real jobs, which is usually the moment this stops sounding like technology and starts sounding like Tuesday afternoon.

Written with Claude, then steered, cut and checked by me. I answer for it.

Finlay Carmichael

Finlay Carmichael is the founder of The AI Lab, helping teams put AI to practical use through hands-on training and clear advice. He brings nearly 30 years' experience building and leading a Microsoft Business Solutions partner, working with hundreds of clients across the UK. Request a call