You want an agent that drafts your replies to customers, sorts your emails, or prepares your reports, and you already have a ChatGPT account. Good news: you don't need to learn to code. OpenAI has built several paths for creating an agent from its tool, from the simplest to the most technical. The trick is knowing which one to pick, and above all what each one can't do.
Key Takeaways
- ChatGPT offers three paths for building an agent: custom GPTs (no code), agent mode (launched in July 2025), and the API for technical teams.
- A custom GPT is set up in minutes with instructions and around twenty reference files, but it stays locked inside ChatGPT.
- OpenAI commits to not training its models on business accounts' conversations, a key point for GDPR.
- ChatGPT is a cloud service: for data that must never leave the company, you need a different architecture.
- OpenAI announced in August 2025 more than five million paying business users on ChatGPT.
The three paths to building an agent with ChatGPT
The first thing to understand is that "building an agent with ChatGPT" doesn't mean just one thing. OpenAI offers three entry points, and they aren't aimed at the same audience.
Custom GPTs are the simplest entry point. A GPT is a specialized version of ChatGPT: you give it instructions (its role, its rules), you feed it reference documents, and you can connect it to external tools through what OpenAI calls actions. It's all set up in a visual interface. Launched in January 2024 alongside the GPT Store, they remain the most accessible entry point for an SME.
Agent mode is more recent. OpenAI launched it on July 17, 2025, under the name "ChatGPT agent." Here, you don't build anything in advance: you describe a complex task and ChatGPT carries it out on its own, browsing websites, running code, and producing finished files. It asks for your approval before committing actions, and you can interrupt it at any time.
The Assistants API is the technical door. It lets developers build a custom agent connected to the company's internal tools. It's the path an SME takes once a custom GPT isn't enough anymore.
The right choice comes down to a single question: is the task simple and repetitive, or complex and one-off?
Building a custom GPT, step by step
A custom GPT can be built in half an hour, and it's the best starting point. Here's how.
Step 1: define the task
Before opening the tool, write one sentence: what exact job should the GPT do? "Draft replies to quote requests" is a good definition. "Help the sales rep" is too vague.
Step 2: write the instructions
This is the step that makes the quality. The instructions describe the role, the rules, and the expected format. Example:
You are the assistant of an accounting firm. Reply to clients in English, in a courteous and precise tone. Never give firm tax advice without referring the client to the accountant. End every reply with the next steps.
Step 3: provide reference documents
You can attach files: a price list, a company brochure, an internal procedure. The GPT draws on them to answer. The practical limit is around twenty files, which covers most needs.
Step 4: connect actions, only if needed
Actions let the GPT query an external tool, for example an appointment calendar. This is more technical: each action corresponds to an API call. Most SMEs skip this at first.
Step 5: test before you share
The GPT can stay private, be shared by link, or be published in the GPT Store. Either way, test it on real cases before handing it to your teams.
The result is a consistent assistant that applies the same rules to every conversation. That's its strength, and its limit: it only does what you described.
What agent mode can do on its own
Agent mode requires no prior setup. You describe a complete piece of work, and ChatGPT handles it.
OpenAI built this capability by merging three existing building blocks: Operator, which interacts with websites, deep research, which synthesizes information, and ChatGPT's conversational intelligence. The result is a system that browses, analyzes, and produces finished deliverables: a spreadsheet, a presentation, a report.
For an SME, the uses are concrete. You can ask the agent to compare three suppliers and produce a summary table. To prepare a client brief from your recent emails. To update a dashboard with new data while keeping the formatting. Scheduled tasks even let you request a report every Monday morning. Other concrete examples appear on our use cases page.
Two safeguards make the use safer: the agent asks for permission before committing actions, and you can take back control at any time. That's an important difference from an agent running unsupervised.
The limits to know before you commit
Transparency has to run both ways. ChatGPT has limits a business leader should know.
The first is dependency on a single provider. A custom GPT lives inside ChatGPT. You can't deploy it on your website, in your internal messaging, or on WhatsApp. Your staff have to open a ChatGPT account to use it. If OpenAI changes its prices or its terms, you don't have an easy plan B.
The second is limited memory. A GPT remembers the files you give it, not past conversations with your customers. For an agent that follows a case over time, that's not enough.
The third is the depth of integrations. Connectors to tools exist, but plugging into an ERP or French software like Sage or Cegid almost always requires development. The GPT alone doesn't talk to your information system.
The fourth is data location. This is the most sensitive point, and it deserves its own section.
Confidentiality: what's guaranteed, and what isn't
For a French SME, the data question comes before the question of features.
What's guaranteed: on business plans, OpenAI commits to not training its models on your conversations or your data. The official wording, in place since the Team plan launched in January 2024, is clear: you own and control your data, and it isn't used for training. OpenAI also provides a data processing agreement and lets you choose hosting in Europe.
What isn't guaranteed: your data passes through OpenAI's servers. This isn't a local installation. For a law firm handling files covered by professional secrecy, or an accountant processing financial statements before they're filed, a contractual promise alone may not be enough. The CNIL, France's data protection authority, keeps a close eye on ChatGPT's use in business, and an impact assessment is still needed as soon as you process sensitive data.
This is exactly where the "local" angle changes things. An agent can run on your infrastructure, with a model hosted at your premises or within the European Union, without ever sending your data to an American service. ChatGPT remains a remarkable tool for generic tasks. For sensitive data, other architectures are called for. We dug into this same question with Claude by Anthropic, a model that pushes reliability further, in our article on Claude.
How much it costs
ChatGPT's public prices move fast, and they depend on the plan. As of this writing, three tiers take shape for an SME.
The individual subscription (Plus) costs around twenty euros a month, and gives access to custom GPTs and agent mode. That's enough for testing, but not for a team: there's no access management and no business-grade data guarantees.
The Business plan (formerly Team) is billed per user, around twenty-five to thirty euros a month per person, with a minimum of two seats. This is where confidentiality guarantees and workspace administration begin.
Beyond that, the Enterprise plan is negotiated directly with OpenAI, with a minimum of several dozen users. It doesn't concern the typical SME.
On top of these subscriptions comes the real cost: the time to configure the GPT, maintain it, and check that it answers correctly. This time is often underestimated. An audit of your needs can price all of this out before you commit.
FAQ
Do you need to know how to code to build an agent with ChatGPT?
No, for custom GPTs and agent mode. Configuration happens in a visual interface, in your own language, with no line of code. Only creating actions (connecting to external tools) requires a minimum of technical skill, and it's optional.
Is my data used to train ChatGPT?
Not on business plans. OpenAI commits to not training its models on your conversations or your data. On the individual subscription, however, the default settings are less protective: you have to turn off training in the settings.
Can a custom GPT work with my business tools?
Partly. Actions let you connect to APIs, but integrating with an ERP or French business software almost always requires development. The GPT alone doesn't connect to Sage or Cegid.
Can a ChatGPT agent run on our own servers?
No. ChatGPT is a cloud service: everything goes through OpenAI's servers. For local execution, you need a different approach: a model hosted at your premises or within the European Union, paired with an orchestration tool. This is an architecture NexeAI designs for regulated professions.
ChatGPT or a custom agent, where do you start?
Start with a custom GPT on a single, well-defined task. If the task touches sensitive data, or needs to integrate with your tools, a custom agent is a better fit. The factor that matters most isn't the tool, it's how clearly the task to automate is defined.
Taking action
ChatGPT is the fastest way to build a first AI agent without coding, and for many generic tasks, it's enough. Its limits show up when you require data confidentiality, integration with business tools, or an agent that works over the long term.
Start with a repetitive task, test it with a custom GPT, and measure the time saved. If the result convinces you and the task touches sensitive data, a custom agent can take over.
Let's talk about your AI agent project: we help you choose between a GPT, a custom agent, or local execution, based on your data and your tools. Also check out our AI agent development offer and our AI agent training to make your team self-sufficient.
