Custom GPTs versus a properly built agent: the practical difference

You built a Custom GPT. It took an afternoon, it does something useful, and people in your team actually use it, which is more than can be said for most internal tools.

Now somebody has told you that you need "a proper agent", and the estimate has a number on it that your Custom GPT did not. You suspect you are being sold something.

Sometimes you are. A Custom GPT is the right tool more often than a firm charging for engineering will volunteer, and much of what used to require a build now ships in the product you already pay for. This piece sets out where the boundary actually falls, so you can tell which side of it you are on before anybody quotes you.

A Custom GPT is a good answer to a specific question

OpenAI describes GPTs, also called custom GPTs, as versions of ChatGPT configured for a specific purpose. The configuration surface is deliberately small: instructions that define behaviour and tone, knowledge in the form of up to 20 uploaded files at up to 512 MB each, conversation starters, a recommended model, capabilities such as web search and data analysis, and then either apps or actions to reach outside services. It is one or the other, not both at once.

That is a genuinely good fit for a real category of work: a person-driven assistant, used by somebody who is present, reading the output and deciding what to do with it. Drafting in a house style. Answering questions against a policy document. Structuring a first pass at something a human then finishes.

If that is your job to be done, you are finished. Do not let anyone tell you otherwise. The cost is a paid seat and an afternoon, and there is no engineering to maintain afterwards.

The limits are boundaries in the product, not defects in your build

This is the part worth being precise about, because vague warnings about "scale" and "robustness" are how upsells are made. OpenAI states the boundary plainly in its own documentation: GPTs are designed to work in ChatGPT, they are not a way to embed ChatGPT in an external website or application, and to build an assistant into a product you use the API.

Underneath that headline sit six specifics.

Nothing happens unless a person makes it happen. There is no schedule and no trigger anywhere in a GPT's configuration. It starts when somebody opens it and types.

The model underneath can change without you touching anything. OpenAI's guidance is that if a model is no longer available, GPTs are automatically switched to a similar current model. Not hypothetical: on 13 Feb 2026, GPT-4o, GPT-4.1, GPT-4.1 mini, o4-mini and GPT-5 in both Instant and Thinking forms were retired from ChatGPT. Any GPT tuned against one of those was quietly moved onto something else.

There is no error handling, because there is nowhere to put any. No retry, no fallback, no alternative branch when an action call fails. The person in the chat sees something go wrong and decides what to do next, which is fine when there is a person and useless when there is not.

You cannot see what it did. GPT builders cannot view individual conversations that users have with their GPTs, so the person who built the thing has no view of what it has been telling people. Conversations are available through OpenAI's Compliance Platform, but that is an Enterprise and Edu capability. Know which tier you are on before you promise an auditor anything.

Changes ship straight to whoever uses it. There is version history, and you can restore an earlier version, which is better than most people expect. What there is not is a test suite, a staging copy or any gate between an edit and the people relying on it. Testing means a person trying it in Preview and forming an opinion. One sharp edge: restore a version that uses actions and you may need to reconfigure its authentication afterwards.

Ownership is better handled than the scare story suggests. You will hear that a Custom GPT walks out of the door with whoever built it. In a managed Enterprise or Edu workspace that is not true: if the owner is deactivated or removed through the normal flow, ownership transfers to a workspace owner and the GPT is flagged as unassigned for reassignment. On a personal account the scare story is exactly right, and personal accounts are where most Custom GPTs live.

Before you commission anything, check what your subscription already does

Here is the part that most articles on this question are too out of date to tell you, and the reason your suspicion may be well founded.

On 22 Apr 2026 OpenAI introduced workspace agents in ChatGPT, describing them as an evolution of GPTs. That announcement put them in research preview on ChatGPT Business, Enterprise, Edu and Teachers. OpenAI's help centre no longer uses the research-preview wording, and has not positively declared general availability either, so treat the current status as unsettled and check before you rely on it. Availability varies by plan, and agents run in the cloud, so they keep working when nobody is logged in. Read their capabilities against the limits above, because they close most of them without an engineer:

• The person-present limit closes. They run on a schedule, or in Slack, or from an API trigger.
• The release limit closes. They have version history, and a preview to test against before publishing.
• The visibility limit closes on Enterprise. The Compliance API gives admins visibility into every agent's configuration, updates and runs, and admins can suspend an agent.
• The ownership limit closes in a managed workspace. They have named owners, editors, group-based sharing and transferable ownership, and each connection can authenticate as the end user or as the agent, with OpenAI recommending a service account rather than somebody's personal login.
• Write actions are gated by default. Approval is asked for during a run, and you can narrow what a connector is allowed to do.

Pricing moved to a credit-based model on 06 May 2026, so confirm current terms with OpenAI directly rather than trusting any figure in an article, including this one.

OpenAI has also said GPTs remain available and that converting a GPT into a workspace agent will be made straightforward. So your afternoon's work is not stranded either way.

The practical consequence: if the pitch you have been given is "your Custom GPT cannot run on a schedule, so we will build you one", the correct first question is whether a workspace agent does it for the price of a subscription you may already hold. If the answer is yes and the pitch does not mention it, that tells you something about the pitch.

Where engineering genuinely starts

There is still a real line, and it is sharper than "when things get serious". OpenAI's documentation for workspace agent API triggers, checked 08 Aug 2026, is explicit about the limit that matters: the agent's response cannot currently be retrieved through the API. You can start an agent run from another system and confirm it ran, and a person can read the answer in ChatGPT, but your systems cannot collect it. Anything downstream that has to receive the output, act on it and be correct sits outside what the in-app tooling does, and no amount of prompt-writing changes that.

Two others in the same register. Connector constraints govern what an agent may ask a connector to do, not what data the connector hands back, so they are not a data-loss control. And shared editing does not merge simultaneous changes: two people editing one agent produces a save conflict and somebody loses work.

Against that, what a built agent buys is unglamorous and specific:

The right-hand column is what you are paying for: a trigger you control, code under version control with tests that run before a change ships, credentials scoped to a service identity rather than a person, output constrained to a schema so an invalid write is rejected rather than saved, logs in your own store on your own retention schedule, and failures that arrive somewhere a human is already looking. If a prototype has proved the idea and the question is how to run it for real, taking a ChatGPT prototype into production covers that transition.

Five questions that settle it

Answer honestly.

1. Does it need to run when nobody is there?
2. Does the same input have to produce the same output, every time?
3. Does anything downstream depend on the result being correct and machine-readable?
4. Would you have to show a client, an auditor or a regulator what it did and why?
5. Does it write into a system of record?

Mostly no: keep the Custom GPT. You have the right tool and someone is selling you a bigger one.

Yes to 1 or 2 only: look hard at a workspace agent first. This is the case that used to justify a build and mostly no longer does.

Yes to 3, 4 or 5: you need software. Not because Custom GPTs are inadequate, but because you have left the space they are designed for. If you have not built anything yet and are weighing seats against a commissioned build, the API versus ChatGPT question for business is the better starting point.

One governance point worth an hour of somebody's time

Custom GPTs multiply quietly. They are easy to make, each one holds whatever material somebody pasted into it, and almost nobody keeps a list. A year in, most organisations cannot say how many exist, who owns them, what is in their knowledge files or which ones reach an external API.

That is an inventory problem rather than a technology problem, and it is the same exercise as auditing the shadow AI already happening in your business. Worth doing before anyone approves a build, because it usually turns up something that already does the job.

The decision in front of you

We are an OpenAI Select Partner and we build on the platform. We also resell nothing and take no margin on your usage, so we have no financial reason to talk you up a rung, and we run the same work model-neutral when the honest answer is a different vendor or the AI already inside software you pay for. More of these conversations end in "keep what you have, add a schedule" than end in a build.

So run your Custom GPT through the five questions. It takes ten minutes and gives you a defensible answer either way. Mostly no and you are done, with a reason to decline the upgrade. Yes on 3, 4 or 5 and the next step is a scope rather than a build: what triggers it, what it writes, what happens when it fails, and what you need to prove afterwards.

Where this fits the rest of the picture is set out in the complete view of ChatGPT and the OpenAI API for UK businesses.

If you would rather have that scoped properly, including an honest view on whether your subscription already covers it, request a quote or read how we approach OpenAI implementation.