ChatGPT vs custom GPTs vs AI agents: what does your small business actually need?

The most common question I get from business owners exploring AI: "Should we just use ChatGPT, or do we need something custom?"

It's the right question, and most of the answers floating around the internet are wrong — either evangelizing custom development for problems that ChatGPT would solve in 20 minutes, or claiming everything can be handled with a free account and a good prompt.

Here's the actual framework I use.

First: understand what you're actually choosing between

Plain ChatGPT (or Claude.ai): A general-purpose AI assistant your team uses through a web interface. No custom setup. No integration with your systems. The AI doesn't know anything about your business except what you paste into the conversation.

Custom GPTs / Claude Projects: A configured version of the same AI with a system prompt that defines its role, uploaded knowledge (your docs, FAQs, SOPs), and specific instructions for how to behave. Still accessed through a chat interface, but tuned for your context. No code required to set up.

Custom AI agents: Software built by developers. The AI is the brain, but it has tools — it can read from databases, send emails, update CRM records, call APIs, trigger workflows. It can take action, not just produce text. This is where significant engineering is involved.

What is an AI agent, in plain English?

An AI agent is an AI model with permission to do things, not just say things. You give it a goal and access to some of your tools (your inbox, your calendar, your CRM, your accounting software), and it works out the steps on its own: read the new request, look up the customer, draft the reply, book the appointment, note what it did. A chatbot hands you an answer and stops there. An agent takes the next step itself, and that's exactly why it needs rules about what it may do alone and what it has to bring to a person first.

AI persona vs chatbot vs custom GPT vs AI agent: what's the difference?

These four words get used as if they mean the same thing. They don't, and the difference decides what you're actually buying.

AI persona: An AI persona is an AI assistant given a defined role, a name, a voice and clear boundaries: who it speaks for, what it will and won't talk about, and when it hands you to a person. It usually sits on top of a chatbot or a custom GPT, so the persona shapes how it sounds, not what it can do.

Chatbot: A chatbot is a chat window on your website, in your app or in Messenger that answers the questions people type into it, usually from your own documents. It answers and stops there; it doesn't go and do the work.

Custom GPT: A custom GPT (or a Claude Project) is ChatGPT or Claude set up with your instructions and your documents, so its answers fit your business. Your team usually uses it inside the ChatGPT or Claude app, and setting one up takes no code.

AI agent: An AI agent is an AI model with permission to act in your other systems, not just answer: update the CRM, book the appointment, send the email. It works out its own next steps, which is why it needs rules about what waits for a person.

Think about it this way: the persona is how it sounds, the chatbot is where it lives, the custom GPT is what it knows, and the agent is what it can do. They stack, too. A chatbot on your website can carry a persona, answer from your documents and, if it's built that way, act as an agent.

The three questions

Question 1: Does the AI need to access your data, or just general knowledge?

If your team is using AI for general tasks — writing, brainstorming, summarizing, editing — plain ChatGPT is probably fine. The AI doesn't need to know anything specific about your business to help draft a marketing email or summarize a report.

If the AI needs to answer questions about your business — your products, your policies, your pricing, your processes — you need some form of custom configuration. A custom GPT with your knowledge base uploaded, at minimum. If the data changes frequently or comes from live systems (inventory, customer records, order status), you need an agent with a database connection.

Question 2: Does the AI need to take action, or just produce information?

This is the clearest dividing line between a custom GPT and a full agent build.

If the AI needs to: send an email, book a calendar appointment, update a record in your CRM, submit a form, call a webhook, or do anything that changes state in another system — that requires code. You can't do this with a custom GPT through the standard interface.

If the AI just needs to answer questions and produce text — even complex text with specific formatting — a custom GPT can handle it. The bar for "needs an agent" is specifically: does it need to do something, not just say something?

Question 3: Who is the user — your team, or your customers?

If it's your internal team using the AI through a chat interface: a custom GPT is often the fastest and cheapest option. No deployment complexity, no website integration, your team already knows how to use chat interfaces.

If it's customer-facing — on your website, in your product, answering support tickets — you almost always need a custom build. You don't want customers using OpenAI's interface directly (no brand control, no audit trail, privacy concerns). You need an embedded widget or API integration, which means code.

The decision matrix

Here are the five options side by side. The first three columns are the three questions; the last is how much work each one takes to set up.

Based on those three answers:

• Team productivity, general tasks, no custom data needed → Plain ChatGPT or Claude.ai. Don't overcomplicate it.
• Team productivity, but needs your specific knowledge → Custom GPT or Claude Project. Configure it with your docs. Still no code required.
• Customer-facing, answers only, no live data needed → Simple chatbot build. Custom prompt, RAG on your docs, embedded widget. A contained build: one job, one audience.
• Customer-facing or internal, needs live data or system actions → Full agent build. API integrations, database connections, tool use. Real engineering, scoped once we've seen the process.

Where each option sits on the 5-level AI ladder

In our workshop series we use a five-level ladder to answer a simpler question: what level of AI does your business need this quarter? The three options above map onto it like this.

1. Personal productivity. You use plain ChatGPT or Claude for your own work: drafting an email, summarizing a document, prepping a quote.
2. Team templates. Shared prompts, or a custom GPT or Claude Project the whole team uses, so the output comes back consistent.
3. Business workflows. Something happens (a new lead, a missed call, a review), AI prepares the response, and a person approves it. This is the first level where AI does work without being asked.
4. Connected systems. Those workflows talk to your CRM, calendar, email, forms and documents, so the work flows through instead of happening one task at a time.
5. Autonomous agents. AI that completes multi-step tasks end to end, with approval gates. Worth it for high-value, repeatable, well-defined processes, and not before.

Plain ChatGPT is Level 1. A custom GPT is Level 2. Anything that acts inside your systems starts at Level 3, and a true agent, one that chooses its own steps, is Level 5. Most owners belong at Level 1 or 2 right now. Skipping levels is the expensive mistake: you end up paying to automate a process nobody has written down yet.

When you don't need an agent

For most first projects, you don't. If the work is writing, summarizing, or answering questions from your own documents, Levels 1 and 2 cover it. If the work is “when this happens, prepare that, and let someone check it,” that's a Level 3 workflow, not an agent. A missed-call text-back, a quote follow-up reminder, an intake form that becomes a CRM record: none of those need software choosing its own next step. They're cheaper to build, easier to test, and easier to trust. (There's a reason your first AI project should be boring.)

An agent starts to make sense when the steps genuinely change from case to case, the process is already well understood, and the work is valuable enough to pay for the build and for someone to supervise it.

If nobody in the business has the time to own that supervision, that's the kind of ongoing job a fractional AI officer takes on.

What an agent should never do without a person approving it

If you do build one, settle this before anything else. An agent can look things up, draft, fill in forms and prepare changes all day long. These four wait for a person:

• Sending. Emails, texts, quotes, or anything else that goes out under your name.
• Paying. Moving money, approving invoices, issuing refunds, changing someone's billing details.
• Deleting. Customer records, files, bookings, anything you'd have to rebuild by hand.
• Anything you can't undo. Submitting, cancelling, signing, publishing. If a mistake can't be walked back, a person makes the final click.

In practice the agent queues the action and someone approves it with one tap. It feels slower at first. It's also how you learn, safely, where the agent gets things wrong. Some approvals can loosen later, once the record shows it handles a case well. I'd keep the money and the deletions with a person for good.

And test it before it touches anything live. That's what our Workflow Proof Sprint is for: we run the workflow against real examples from your business, messy ones included, write down where the approval points sit, record what fails, and finish with a plain recommendation to deploy, revise, or stop.

Where people go wrong

My recommendation for most SMBs starting out

Start with the simplest thing that solves a real problem. If your team needs help with repetitive writing, give them Claude.ai accounts and a shared prompt library. That's a monthly subscription per person, and you'll know within 30 days whether the time savings justify going further.

If you want a customer-facing chatbot, build a proper one — not a half-measure that creates more problems than it solves. But scope it tightly. One function. One audience. Deployed, measured, tuned. Then expand.

The biggest waste I see: companies that pay for a complex agent build before they've checked that their team will actually use AI tools. Run the cheaper experiment first. Build confidence. Then invest in automation.

Not sure where your team stands? The AI readiness quiz is seven questions and about two minutes, and it points to a sensible first project. If you'd rather work it out with someone, an AI Strategy Day is where we map how the work actually moves through your business and decide which of these options each process needs, if any.

Quick answers