Custom GPTs are the right place to start when you want an AI assistant that writes, summarises or answers in a way you have set up, and a built AI agent is the right step when the job needs to look things up and change records in your own systems. A custom GPT mostly answers. An agent answers and then acts: it checks the diary, books the slot, updates the customer record and hands over to a person when it should.
A disclosure first: CentriWeb builds AI systems for businesses, so we have a stake in the second option. We have tried to describe both fairly, and the last sections cover who should not pay for a custom build.
What custom GPTs are
A custom GPT is a version of ChatGPT that someone has configured for a particular job. You give it standing instructions, such as the tone to use and the steps to follow, so you do not retype them in every chat. OpenAI also lets a GPT call outside services through what it calls Actions. According to OpenAI’s documentation on GPT Actions, these let ChatGPT users work with external applications through API calls, either to fetch information or to do something in another application. The developer specifies how the action authenticates, and the model chooses which call fits the request.
What each plan includes, who can build a GPT, who can use one and what limits apply are set by OpenAI and change. As at November 2026, check OpenAI’s own help pages for the current position before you plan around any of it.
Custom GPT vs AI agent: how they differ
The labels blur, because a GPT with Actions can behave like a small agent. The table describes the usual shape of each in a small business.
| Custom GPT | AI agent built for the business | |
|---|---|---|
| What it is | A configured ChatGPT assistant | A system that answers from your information and acts in your tools |
| Where people use it | Inside ChatGPT | On your website, by text, on the phone or inside your own software |
| What it touches | What you paste in or attach, plus any Actions someone builds | Your calendar, CRM, job system and inbox, with set permissions |
| Who maintains it | Whoever built it, usually you or a staff member | The builder, under an agreement, or your own technical person |
| How it is paid for | A subscription to the platform | A build and usually a monthly fee for hosting, usage and upkeep |
| Who it suits | Your own team doing repeated drafting and thinking work | Work where customers deal with it directly or records must change |
What custom GPTs do well
For work that stays inside your own team, a custom GPT can be the cheapest useful AI project you run. The common pattern is a task someone does over and over, where the output is a draft that a person reads before it goes anywhere.
- A bookkeeping practice sets up a GPT with its house style for client emails, so a junior can turn a rough note about a late BAS into a polite, consistent message that a senior reads before sending.
- An electrical contractor’s office manager pastes the site notes from a job and gets back the wording for a quote, in the format the business always uses.
- A clinic manager keeps a GPT that rewrites patient-facing instructions in plain language, then checks the result against the clinical source.
- A marketing owner keeps one for turning a long job write-up into three social posts in the voice of the business.
None of these touches a customer directly, and none changes a record. If the GPT gets something wrong, a person catches it before any harm is done. That is the sweet spot. It is also quick to test: write the instructions, run ten real examples through it, and see whether people actually use it for a month.
Where a custom GPT runs out of road
The limits show up when the job moves from drafting to doing, or from your team to your customers.
- Customers are not in your ChatGPT. A GPT lives in ChatGPT. If you want the assistant on your website, answering the phone or replying by text, you need something else around it. Check OpenAI’s pages for who can use a shared GPT.
- Acting needs plumbing. Actions can call another application’s API, but someone has to write and test that connection, handle sign-in, decide what the assistant is allowed to change and deal with failures. That work is the same work a custom build does. It does not disappear because the front end is ChatGPT.
- Knowing when to stop is your job to design. A business assistant needs rules for handing a conversation to a person: an angry customer, a medical question, a request for a refund. Instructions can describe those rules. They are harder to enforce and check than when the system is built around them.
- Answers can be wrong. Any model can state something confidently that is not in your documents. Our guide on reducing AI hallucinations covers the checks that help.
- Business information goes through someone’s account. Think about what gets pasted in, and whose login is attached. Our guide on whether ChatGPT is confidential enough for business data sets out the questions to ask.
What a built AI agent adds
An agent built for the business answers from your own information, then does the follow-on work. In practice that means a phone or chat assistant that books into the real calendar, writes the enquiry into the CRM with the details it collected, sends the confirmation and passes the conversation to a person when a rule says it should. Anthropic, which makes the Claude models, separates systems where fixed code paths direct the steps from agents where the model decides its own steps, and advises in its guide to building agents that you should look for the simplest approach that works and add complexity only when needed. We agree with that, and it is why many jobs do not need a custom agent at all. We cover the difference in AI automation vs AI agents.
A quick test for which one you need
- Who uses the output? If your own staff read it before it goes anywhere, start with a custom GPT. If a customer talks to it, you need a system around it.
- Does a record have to change? A new booking, a status update or a payment link means connecting to your software.
- What does one mistake cost? A clumsy draft costs a minute. A wrong appointment or a wrong price costs a customer.
- How often does it run? A task done twice a month may never repay a build. A task done forty times a day probably will.
When a built agent fits better, and when it does not
A system built and looked after for your business fits better when customers deal with the assistant directly, when it must write to your calendar or CRM, and when missing an enquiry costs you real money. For example, a plumbing business that loses jobs because nobody answers after hours has a problem a custom GPT cannot fix, because the caller never opens ChatGPT. That is a job for an AI phone and chat assistant that answers from the business’s own information, books the real calendar and hands over to a person when it should.
It does not suit everyone, and you should hold off if:
- the work is low volume, a handful of cases a week, where a checklist or a custom GPT does the job;
- you or a staff member have the time and interest to build and maintain a GPT or a simple automation yourselves;
- your budget is better spent on the plan you already pay for, because the standard features cover what you need; or
- the process is not settled. Fix the process on paper before you ask any system to follow it.
Most of our conversations start with a 30 minute call to work out which of these applies, and sometimes the answer is that an off-the-shelf tool is enough. The wider question of where to begin is in AI for small business: where to start.