Do I need a custom AI model?

No, almost certainly not. For a business of two to ten people, what you really want is an AI tool that follows your house rules and can look things up in the systems you already use, and neither of those involves training a model.

It's worth being clear on this before you sign anything, because "your own AI, trained on your business" is still a phrase you'll hear in sales pitches, while the big providers have stepped back from offering it. On 7 May 2026, OpenAI stopped letting new customers fine-tune its models. Its pricing page says so in plain words: "OpenAI is winding down the fine-tuning platform. The platform is no longer accessible to new users, but existing users of the fine-tuning platform will be able to create training jobs for the coming months." Anthropic's glossary is just as direct about Claude: "The Claude API does not currently offer fine-tuning."

What does "trained on your data" usually mean?

Two quite different things have ended up sharing one phrase.

Training, often called fine-tuning, changes the model itself. It's a bit like sending a new member of staff on a long course: they come back with new habits and a particular style, but they haven't memorised your price list. You feed the model worked examples, and its internal settings (known as weights) shift so its answers start to look like those examples. It's permanent and expensive, and the model comes away with a style and a habit rather than facts about you.

Access leaves the model exactly as it is and hands it your documents at the moment you ask. It's like giving a capable temp the key to the filing cabinet. They read what they need, answer your question, and nothing about them changes.

Vendors often say "trained on your data" when they mean access. That's usually loose wording rather than dishonesty. But the difference in cost, complexity and how tied-in you'll be is huge, so it's in your interest to know which one you're being quoted for.

It helps to think of the options as four steps, from cheapest to most expensive. Most small businesses only need the first two.

Step one: a well-written set of instructions

The cheapest useful thing is a properly written set of instructions that fixes the rules the AI works within: your tone, your format, your definitions, the things it must never say and the shape of the answer. Think of it as the briefing sheet you'd give a new starter. You write it once and reuse it every time, as a project's custom instructions, a saved prompt, or a template your team opens rather than starting from a blank page.

That sounds almost too simple, but OpenAI's own model optimisation guide puts it strongly: "The prompt engineering process may be all you need in order to get great results for your use case." Prompt engineering just means working out and testing the wording of those instructions.

The cost is an afternoon of someone's attention.

The limit is that it can't know anything you haven't written down for it. It won't know today's stock level, last month's invoices, or what you said to a customer in March.

If you'd rather not write and test those instructions yourself, we're happy to set them up for you in the AI tools you already use.

Step two: letting AI look inside the systems you already pay for

This is where most of the value sits for a small business, and it's the step that most often gets mislabelled as training.

The main way of doing it now has a name. The Model Context Protocol, or MCP, describes itself in its own documentation as "an open-source standard for connecting AI applications to external systems", such as a calendar, a Notion workspace or a database. We've explained how it works, with a worked example, in what APIs and MCP are. The point that matters here is that nothing about the model changes: it's handed a way to look things up, not retrained.

That's usually what people picture when they imagine a custom AI. Not a model that has memorised your business, but one that can go and look at your accounting system, your shared documents and your job records when it needs to.

There are some awkward details worth knowing. Anthropic's MCP connector is still marked as beta, which means it's still being finished. It supports only tool calls rather than the full protocol, and it needs the systems it connects to be reachable securely over the public internet, so ones running only on your own computer can't be connected directly. It isn't covered by Anthropic's zero-data-retention arrangements (an agreement that the provider won't keep what you send), so data passing through those connections falls under standard retention. And the documentation recommends switching off tools that can change or delete things when you want a person to confirm before anything changes.

That last point is the one we'd stress most. Letting AI read your records and letting it change them are two different decisions, and it's worth making them separately. If you'd like to know more about what happens to the information you share, our article on whether AI tools train on your business data covers it.

If connecting AI to your own systems is the part you'd like done for you, we're happy to set up those connections, starting with read-only access.

Step three: a searchable library for larger amounts of material

If you have more material than fits in one conversation, such as years of quotes, a policy library or every support email, you can put it in a searchable store that the AI looks things up in as needed. It works rather like the index at the back of a book: the AI finds the right pages first, then reads them to answer. The technical name is retrieval, or retrieval-augmented generation (RAG). OpenAI's version, file search, "enables models to retrieve information in a knowledge base of previously uploaded files through semantic and keyword search".

The published price helps set expectations. Storage is $0.10 per GB per day with the first gigabyte free, and searches are $2.50 per thousand. Five gigabytes of documents means paying for four: forty cents a day, or roughly twelve dollars a month before any searching. That's a running cost more like a utility bill than a project.

Anthropic's own glossary gives the honest caveat: "the effectiveness of RAG depends on the quality and relevance of the external knowledge base and the knowledge that is retrieved at runtime." Put simply, if your documents are a muddle, the answers drawn from them will sound confident and still be muddled.

Step four: fine-tuning, where the numbers don't work for a small business

OpenAI's supervised fine-tuning guide says the minimum is ten examples, that improvements show from 50 to 100, and recommends "starting with 50 well-crafted demonstrations". Fifty sounds manageable until you look at what an example is: a complete question paired with the exact answer you want, written by someone who knows the right answer, and consistent with the other forty-nine. That's skilled work, and it would be your staff doing it.

It's also worth noticing what OpenAI lists as the good uses: sorting things into categories, nuanced translation, producing content in a specific format, and fixing cases where the model doesn't follow instructions. Knowing things about your business isn't on that list, because fine-tuning isn't how a model learns facts.

The pricing points the same way. Providers charge by the token, which is a small chunk of text, so think of it as paying by the word. Training gpt-4.1 costs $25 per million tokens, which is manageable as a one-off. The ongoing cost is the problem: a fine-tuned gpt-4.1 costs $3.00 per million tokens going in and $12.00 coming out, against $2.00 and $8.00 for the ordinary model. That's fifty per cent more on every request, for as long as you use it. On the smaller mini model the gap is bigger in proportion: $0.80 and $3.20 fine-tuned, against $0.40 and $1.60 standard.

It also goes out of date by design. A fine-tune is tied to a dated snapshot of a model, a bit like a particular edition of a textbook, and OpenAI states that "inference on fine-tuned models will continue to be available until the base models are deprecated". In other words, when the underlying model is retired, your fine-tuned version goes with it. The closure itself runs to a published timetable: organisations that had never fine-tuned were locked out on 7 May 2026, inactive ones on 2 July 2026, and even active existing customers lose the ability to create new fine-tuning jobs on 6 January 2027.

OpenAI's guidance on when to bother puts testing first: only invest in fine-tuning once you've set up evals, which are simply tests that measure whether the answers are good enough. If nobody has asked you how you'd measure whether a custom model is actually better, that first step hasn't been done.

What should you ask a vendor who offers a custom model?

Ask one question first: does this change the model, or does it give the model access to our files?

If the answer is access, that's the right answer for nearly everyone. Now price it as access. Ask which systems it connects to, whether it can change things as well as read them, where your data goes, and what you're left with if you cancel.

If the answer is training, ask which model, given that OpenAI has closed fine-tuning to new customers and Anthropic doesn't offer it through the Claude API. Ask who writes the examples and how many there will be. Ask what happens when the underlying model is retired, and what each request will cost afterwards.

For a business of two to ten people, the honest answer is almost always the first two steps, done properly: instructions that fix the rules, and connections into the systems you already pay for. We'd say that, since setting those up is the work we do. The reason to trust it anyway is that those are the cheap steps, and we're steering you away from the expensive one.

If nobody is trying to sell you a custom model, there's nothing you need to do. If you'd like help with the first two steps, get in touch and we'll talk it through.

Sources

• Pricing - OpenAI
• Deprecations - OpenAI
• Supervised fine-tuning - OpenAI
• Model optimization - OpenAI
• File search - OpenAI
• Glossary - Anthropic
• MCP connector - Anthropic
• What is the Model Context Protocol? - Model Context Protocol

Checked on 22 October 2026. Provider pricing and fine-tuning availability have both moved sharply this year, so check the links rather than trusting the date on this page.