A boutique consultancy quoted an operations director I work with £2,000 to build a custom GPT and £500 a month to maintain it. She asked me to look at the proposal. The build was a system prompt of about 600 words and twelve uploaded PDFs. She could have done it herself in an evening for the price of a ChatGPT Plus subscription. The vendor was not lying about what they had built. They were charging consultancy rates for it.
The phrase custom GPT now covers a wide spectrum, from valuable bespoke configurations to long prompts in expensive packaging. The plain-English version lets you tell the two apart.
What is a custom GPT?
A custom GPT is a configuration on top of an existing AI model. You write a system prompt that gives the model standing instructions, you upload a knowledge base of reference documents, and you optionally connect it to a few external tools. The underlying model, GPT-5, Claude Opus, Gemini, is unchanged. What changes is the instruction set the model reads before each conversation and the documents it can look up while answering.
OpenAI introduced GPTs in late 2023 and opened the GPT Store in early 2024. Anthropic launched the equivalent under the name Projects, accessed inside Claude. Google’s version is called Gems and lives inside Gemini for Workspace. The names differ, the mechanism is essentially the same.
The retrieval part is worth pausing on. When you upload PDFs to a custom GPT, you are not retraining the model. The configuration uses retrieval-augmented generation, which means it searches your uploaded files for relevant passages at the moment a question is asked and includes those passages in the prompt sent to the model. The model writes the answer. Your documents shape what it has to draw on. They never become part of the model itself.
That distinction is what separates a custom GPT from fine-tuning. Fine-tuning is a formal training process that adjusts the model’s internal weights using your data, costs hundreds to thousands of pounds per run and produces a new model variant you then have to host and maintain. A custom GPT is reversible, instant and almost free. Vendors sometimes blur the two with phrases like “trained on your data”. For most situations they mean configured, not trained.
Why it matters for your business
The first thing it changes is how quickly you can stand up a useful internal tool. A team that spends three hours a week digging through old contracts, proposals or compliance documents to answer recurring questions can replace that with a fifteen-minute conversation against a custom GPT loaded with the same files. The cost of the configuration is close to nil if you do it inside an existing ChatGPT Plus or Claude Teams subscription. The cost of not doing it is the salaried hours you keep paying for retrieval that an automated tool can do better.
The second is reversibility. Because a custom GPT is a configuration rather than a trained model, you can delete it, modify it or replace it without sunk cost. That makes it the right shape of tool for experiments. If a use case fails, you scrap the configuration and try a different one. If it works, you formalise it and bring it under governance. The cycle costs hours, not weeks.
The third is governance debt, which arrives quietly. Once a team starts building custom GPTs, they tend to multiply. A ten-person team can end up running twenty or thirty configurations, each with a different owner, a different set of files and a different freshness state. The original creators move on, the knowledge bases go stale, and someone in customer service ends up quoting last year’s policy because that is what their custom GPT still has loaded. The configurations are cheap to build. They are not free to keep accurate.
Where you will meet it
You will meet custom GPTs in two shapes. Internal: someone on your team has built one to solve their own problem. By 2026 a fair number of knowledge workers have built at least one. External: a vendor pitches a custom GPT as a productised service, usually with a build fee and a monthly retainer. The pitch is consistent. “We will build you a custom GPT for X”, where X is customer support, proposal drafting, lead qualification or compliance summarising.
You will also meet them in the OpenAI GPT Store and the equivalent listings inside Claude and Gemini. Some are free, some charge a subscription. Quality varies widely. A custom GPT for UK employment law may have been built by a thirty-year practitioner or by someone following a Sunday tutorial. The store does not separate them.
The most useful place to meet the term is when a vendor’s price tag does not match what they have actually built. A bespoke custom GPT worth paying for integrates with your CRM or helpdesk, runs under a Data Processing Addendum, has a refresh schedule, and surfaces sources in its answers. A weak one is a long prompt and a few PDFs sold at consultancy rates. Both call themselves the same thing.
When to ask about it, when to ignore it
Ask about a custom GPT when the task is information retrieval and synthesis from your own documents, the work is repetitive across several people in your team, and you have a body of source material the configuration can draw on. Those three conditions together are what make the maths work. A custom GPT loaded with your past proposals, methodology notes and case studies, used by a small bid team, will pay for itself in weeks. The same tool used by one person occasionally for novelty will not.
Ask hard questions when the task involves reasoning, judgement or pattern recognition rather than retrieval. A custom GPT cannot be “trained for clinical diagnosis” or “configured for strategic advice” by a system prompt and a folder of PDFs. The underlying model either can do the task already or it cannot. Configuration does not change that.
Ask harder questions when the configuration touches personal data. The ICO’s guidance is unambiguous: if your custom GPT processes employee or customer data, you remain the controller and you owe the standard obligations, lawful basis, data minimisation, a Data Processing Addendum, and consideration of where processing happens. OpenAI’s enterprise tier offers European data residency; Claude Teams offers UK residency at a more SME-friendly price; Google’s Gems sit inside Workspace. Standard ChatGPT Plus does not give you those guarantees.
Ignore the term when the offering is a long prompt and a few PDFs sold at vendor rates. You can build the same thing for the price of a Plus subscription. Ignore it too when the work is already handled well by a single good prompt typed into the chat. The wrapper adds nothing then except recurring fees and lock-in.
Related concepts
System prompt is the standing instruction set you write at the top of the configuration. It is what shapes the assistant’s tone, scope and guardrails across every conversation. The quality of the system prompt is usually the single biggest determinant of whether a custom GPT is useful or generic.
Knowledge base is the folder of files the configuration can reference. PDFs, spreadsheets, internal docs. The custom GPT does not memorise these; it looks them up at query time and passes the relevant passages to the model.
Retrieval-augmented generation, or RAG, is the mechanism underneath the knowledge base. It is what lets the custom GPT answer using your documents instead of inventing answers from training data. RAG has its own explainer in this series.
Fine-tuning is the alternative path, formal model training on your examples to bake patterns into the model itself. Slower, costlier and harder to reverse than a custom GPT, but the right call when you need consistent behaviour or domain pattern recognition that no amount of configuration can deliver.
Prompt engineering is the craft of writing the system prompt and the user-side instructions. A well-engineered system prompt can double the usefulness of a custom GPT. A weak one produces generic answers regardless of the model underneath.
The honest test of any custom GPT, internal or vendor-built, is the question your documents do not cover. Ask the configuration something it should not know. The version worth keeping says it does not know. The version worth deleting confidently invents an answer.
