How-To Guide

How to Build a Custom AI Assistant for Your Business (No Code Needed)

Most teams keep retyping the same long prompt into a blank chat box — here is how to turn it into a named, shareable assistant, in five blocks, tested on ten real cases.

B Biztrategy Published 2 November 2026 · 7 min read
An individual writes notes on green sticky notes placed on a laptop in a work setting.

Somewhere on your team, someone has a prompt saved in a notes app. It is four paragraphs long, it took them a fortnight to get right, and they paste it into a blank chat window three times a week — tweaking a line here, forgetting a line there. When they are on holiday, nobody else knows it exists. This is the stage most small businesses get stuck at: the AI works, but only in the hands of one person, and only when they remember exactly what to type.

A custom assistant fixes that. It takes the prompt that already works, bolts your reference documents to it, gives it a name, and puts it in front of everyone who needs it. No code, no developer, no new subscription — and roughly three or four hours of setup. Here is how to build one properly, and how to tell whether it is any good before you let the rest of the team near it.

What a custom AI assistant actually is

A custom assistant is a saved bundle of three things: a set of standing instructions, a small library of reference files, and a name your team recognises. Instead of opening a blank chat and re-explaining your business, you open "Quote Drafter" or "Enquiry Triage" and paste in today's input.

Every major platform now offers this under a different label. ChatGPT has Custom GPTs and Projects. Claude has Projects. Microsoft 365 Copilot has agents. Gemini has Gems. The mechanics differ slightly, but the concept is identical, and the instruction file you write for one will transfer to another with minor edits — which is exactly why you should write it as a document you own rather than typing it straight into a text box.

Be clear about what it is not. It is not a model trained on your data; the underlying model is unchanged and your documents are simply read at the moment you ask. It is not an integration — it will not reach into your CRM unless you connect one separately. And it is not an autonomous agent that acts on your behalf. It is a very well-briefed colleague who has read your handbook and forgets nothing between conversations.

The payoff is consistency. Once the brief lives in the assistant rather than in one person's head, the quality of the output stops depending on who typed the prompt.

Start with one job, not a job description

The single most common mistake is scoping too wide. "A marketing assistant" is a job description, and an assistant built to that brief will be vague at everything. "Turns a 40-minute discovery call transcript into a one-page proposal brief in our format" is a job, and an assistant built to that brief will be genuinely useful by Thursday.

Use three questions to test a candidate task:

  • Can you name the input and the output? If you cannot say "it takes X and produces Y", the task is not ready to be delegated — to a person or to a model.
  • Does it happen at least weekly? Below that frequency, the setup time will not pay back and nobody will remember the assistant exists.
  • Do you already have ten good past examples? You need them for the worked example in the instructions and for the test in step five. If you have none, the output shape is not settled enough yet.

Tasks that consistently clear all three bars in small businesses: drafting quotes and estimates from a site visit note; triaging inbound enquiries and drafting a first reply; turning call notes into a proposal brief; writing the monthly commentary that sits on top of your numbers; producing a job ad plus screening questions from a short role description.

Pick the task your best person does well and your newest person does badly. That gap — not the glamorous work — is where a custom assistant pays back fastest.

The five-block instruction file

Write the instructions in a document first, then paste them in. Five blocks, in this order, will cover almost any business task:

  1. Role and scope. Who the assistant is, what it handles, and — critically — what it refuses. "You draft quotes for domestic plumbing jobs in the Barcelona area. You do not quote commercial work or anything requiring a structural survey; for those, say so and stop."
  2. Inputs. What the user will give it, and what to do when something is missing. "You will receive a site visit note and photographs. If the note does not state the property age or the boiler model, ask for them before quoting."
  3. Output shape. The exact structure, in order, with a length limit. Specify the sections, the tone, the units, the currency, the spelling convention. Vagueness here is what produces output that needs rewriting.
  4. Rules and red lines. The non-negotiables. Never invent a price not present in the rate card. Never promise a lead time. Never name another client. Always flag assumptions in a closing list. State the escalation rule explicitly: if confidence is low, say so rather than guessing.
  5. Worked example. One real input and the ideal output, pasted in full. This single block does more for quality than the other four combined, because it shows rather than describes. Strip the client name first.

Keep the whole thing under roughly 600 words. Longer instruction files are not more precise; they are just harder to maintain, and contradictions creep in. If you need more structure in how you phrase these blocks, our guide to prompt engineering for small business covers the underlying pattern in more detail.

Give it the right context — and nothing else

Most platforms let you attach files the assistant can read every time. The instinct is to upload everything. Resist it: five to fifteen well-chosen documents beat two hundred, for three reasons. Relevant detail gets diluted in a large pile. Old versions get quoted back at you as current policy. And every extra document widens the confidentiality surface for no benefit.

Four rules that prevent most of the trouble:

  • One canonical version of each document. Delete the drafts. If two price lists are attached, you will get answers from both.
  • Date-stamp the filenames. rate-card-2026-10.pdf tells the assistant — and you — how fresh it is.
  • Strip personal and client-identifying data unless you are on a business or team plan whose terms say your inputs are not used for training. Check that before anything sensitive goes in.
  • Put a review date in the diary. Quarterly is enough for most reference packs; monthly if your prices move.

If the reference pack is growing past twenty documents, the problem is no longer the assistant — it is that the underlying information is not organised. Our walkthrough on building an AI knowledge base covers that step.

Test it on ten real cases before anyone else sees it

This is the step almost everyone skips, and it is the one that decides whether the assistant survives contact with your team.

Pull ten real inputs from the last few months — ones where you already know what a good output looks like. Run each one through the assistant and score it: usable as-is, usable after light editing, or wrong. Keep the scores in a simple sheet with a one-line note on what failed.

The bar to clear before rollout is eight out of ten usable with light editing, and zero confidently wrong answers. A wrong answer delivered with assurance is far more expensive than a vague one, because nobody checks it — see our notes on preventing AI hallucinations in client work for the guard rails that help here.

When something fails, fix the instructions, not the model. Nine times in ten the fault is a missing rule or an ambiguous output spec. Re-running the full ten-case sheet after a change takes about five minutes, which makes it cheap enough to do every time you edit the brief.

Roll it out, name an owner, and prove it paid

Before you share the link, time the task the old way on five real instances and write the average down. Without that number you will never be able to tell whether this worked, and "it feels faster" does not survive a budget conversation.

Then name one owner. Not a committee — one person whose job includes keeping the instructions current, collecting complaints, and re-running the test sheet when prices or policies change. Assistants without an owner quietly rot: the rate card moves, the outputs go stale, and within six weeks the team is back to the blank chat box.

The economics are usually straightforward. The assistant itself costs nothing extra beyond the seat you already pay for — roughly €18–30 per user per month on a business plan. Setup is about three to four hours: two on the instructions, one on assembling the reference pack, one on testing. If the task takes 45 minutes today, the assistant gets it to 10, and it happens three times a week, you are recovering around seven hours a month from a one-off afternoon of work. That is the whole case.

Finally, show the team how to use it rather than announcing it in a message nobody reads. A fifteen-minute walkthrough on a real example, with the red lines explained, converts far better — our guide on training your team to use AI covers the rollout pattern.

The bottom line

A custom AI assistant is not a technical project. It is the act of writing down something your business already knows how to do, in enough detail that a capable stranger could do it — which is a useful exercise even if you never paste it anywhere. Pick one weekly task with a fixed output. Write the five blocks. Attach ten documents, not two hundred. Test on ten real cases and refuse to ship below eight out of ten. Name an owner and a review date.

Do that once and you will have something more valuable than a clever prompt: a repeatable process that works the same whoever is at the keyboard, and a template for the next five assistants you build.

Where does your business stand on AI?

Take the free 3-minute AI Readiness Quiz and get a personalised score with your next steps.

Take the Free Quiz →