The short version: A custom GPT trained on your own documents can save a small business real hours a week on repetitive writing and client questions, but only if you feed it your actual material and check its work every single time. Build it wrong and it will confidently invent facts in your name. I have done both, and I will show you which is which.
What a custom GPT is (and why most people build them wrong)
A custom GPT is a version of ChatGPT that you set up yourself, give a narrow job, and feed your own documents so it answers using your material instead of the general internet. You do not need to code anything. If you have a ChatGPT Plus or Team account, which runs about £20 a month, you can build one in an afternoon.
Most small business owners hear “custom GPT,” get excited, upload one PDF, and expect a digital employee. That is where it goes wrong. A GPT is not a hire. It is closer to a very keen junior who has read some of your files and will guess confidently when it hasn’t. The businesses getting real value out of this are the ones that treat it like training a new starter, not switching on a robot.
The two GPTs I use every week
I have built six of these over the past year. Two get used constantly. The other four are gathering dust, which tells you something on its own.
The first is trained on roughly 200 of my own newsletters, blog posts and old talk transcripts. Its job is one thing only: draft a first pass of a blog post in something close to my voice. Before this, a first draft took me 60 to 90 minutes. Now it takes about 25 minutes to get a rough draft, then another 45 minutes of me pulling it apart, checking every claim and rewriting the flat bits. It has not cut my total writing time in half, whatever the LinkedIn posts about “10x productivity” want you to believe. It has cut the blank-page part, which for me was always the worst bit.
The second is for a client of mine, a small accountancy practice in Kent with four partners and no marketing person. We trained a GPT on their engagement letters, their FAQ document, and about 40 real client emails with names stripped out. Its job is to draft replies to the routine onboarding questions new clients ask every week: what documents do you need, when is the deadline, how does payment work. The senior partner reckons it saves him close to four hours a week that used to go on typing near-identical emails. That is not nothing for a four-partner firm.
The uncomfortable bit nobody tells you before you start
Here is the part that most guides on this topic skip. In my third month of using my own writing GPT, it drafted a paragraph citing an email open rate statistic, presented with total confidence, complete with a specific percentage. It sounded exactly like the kind of stat I’d quote. I nearly used it. I only caught it because the number felt slightly too neat, and when I went looking for the source, there wasn’t one. It had made it up wholesale.
That is the part nobody puts in the headline because it undercuts the sales pitch. These tools do not know the difference between something they read and something they invented. They will state both with identical confidence. If you are a small business owner publishing under your own name, and most of you are, that means every single fact, number or quote a GPT produces needs checking against a real source before it goes anywhere near a client, a newsletter, or your website. Skip that step once and you are one wrong statistic away from an email that damages the trust you have spent years building.
How to build one, step by step
- Pick one narrow job, not “help my business.” “Draft replies to onboarding questions” works. “Be my marketing assistant” does not.
- Gather 10 to 20 real documents. Not a mission statement. Real emails you’ve sent, real proposals, real transcripts, real FAQ answers. The GPT is only as good as what you feed it.
- Go to Explore GPTs, then Create, inside ChatGPT. You need Plus or Team access.
- Write the instructions in plain English. Tell it what it’s for, who it’s talking to, what tone to use, and what it must never do (invent statistics, promise prices, mention competitors).
- Upload your documents under Knowledge. This is the step people skip or half do, then wonder why the output is generic.
- Test it with real questions from real people, not the polite hypothetical ones you’d ask yourself. Give it to a colleague and let them try to break it.
- Set a recurring reminder, monthly or quarterly, to refresh the documents. Rules change, prices change, your own opinions change. A stale GPT giving out last year’s information is worse than no GPT at all.
- Decide who can use it and for what. A drafting tool for your team is a different risk level to something client-facing.
The accountancy firm learned that last point the hard way. Tax rules shifted in one narrow area, and the GPT kept citing the old rule for about three weeks because nobody had gone back in to update the source document. Nothing serious came of it, but it was an uncomfortable phone call to a client who’d been given outdated information with total confidence. Build the update habit in from day one.
Why voice and personality still matter more than people admit
The businesses that get the most out of this have a strong, recognisable voice to begin with. It is the same principle brands like Duolingo built their whole marketing strategy around, a personality so distinct that even an automated system stays recognisably them. If your brand voice is already vague, a GPT trained on vague material will produce more vague material, just faster. Calm’s brand voice works because it is consistent everywhere, not because a tool made it consistent. Fix your voice first, then train the GPT on it, not the other way round.
Where this fits with the rest of your marketing
A custom GPT handling FAQs is doing the same job as a good interactive tool on your site, it is answering a question a visitor already has instead of making them wait for a human. If you like the idea of tools that keep people engaged and answering their own questions, interactive calculators do a similar job on your website and are worth pairing with this. And if you want the GPT’s output to trigger something automatically, like a drafted email landing in a folder for review, that is exactly the territory IFTTT built its whole product around, small automated triggers doing the boring connecting work between apps.
If none of this appeals and the whole thing sounds like a slog on top of running a business, that is a fair reaction, and it is why some owners bring in outside help rather than building it alone. If that is you, it is worth looking at what an AI implementation coach does before deciding, because the setup work is a few focused hours, not weeks.
What it will not do, whatever the adverts say
It will not replace judgement. It will not know your client has had a bad week and needs a softer email than usual. It will not notice that a prospect’s question is really about something else entirely, the way a person answering the phone would. The businesses that treat their custom GPT as a first draft machine, always checked, always edited, get real hours back. The ones that treat it as a replacement for a person end up publishing things they wish they hadn’t. I have been on both sides of that line in the same year, which is exactly why I am telling you now rather than after you’ve made the same mistake.
Frequently asked questions
Do I need to know how to code to build a custom GPT?
No. If you have a ChatGPT Plus or Team subscription, building one is done through a simple interface, you type instructions in plain English and upload documents. No coding involved.
How much does it cost to build a custom GPT for a small business?
The main cost is a ChatGPT Plus subscription at around £20 a month, or Team pricing if several people need access. There is no extra charge to build the GPT itself. The real cost is the time it takes to gather and organise your source documents.
Can a custom GPT make mistakes even when trained on my own documents?
Yes, and this is the part people underestimate. It can still invent facts, misquote your own documents, or give outdated information if you haven’t refreshed the source material. Every output needs a human check before it reaches a client or goes on your website.
What’s the fastest way to see if a custom GPT is worth it for my business?
Pick one repetitive task you or your team do at least five times a week, like answering the same client question or drafting similar emails, and build a GPT for that single job first. If it saves real hours within two weeks, expand from there. If it doesn’t, you have lost an afternoon, not a business.
Related reading: Why AI Pricing Advice Is Quietly Training Small Business Owners to Undercharge and Why Your AI Meeting Notetaker Might Be Breaking the Law (And Killing Your Sales Calls).
