A template chatbot follows a pre-written script: it recognises the type of question and returns the answer or decision branch assigned to it. An AI agent is a language model that decides for itself which steps and which tools (order management, inventory, calendar, email) it uses to complete a task. Most businesses need neither extreme. What they need sits in between: a tightly scoped process wired into their own systems, where the AI only makes decisions where it genuinely has to.
The typical story goes like this: someone switches on an off-the-shelf “AI assistant” on their website, and a few weeks later realises the bot is friendly and fluent but knows nothing that matters. It can’t see where an order is, it doesn’t know whether a product is in stock, and every now and then it promises something the company doesn’t offer. That isn’t the AI’s fault. It’s the wrong tool for the job.
- Template chatbot: predictable and cheap to maintain, but it only knows what someone wrote into it.
- Off-the-shelf AI assistant: a language model sitting on top of your FAQ. It talks well, but usually has no access to your real systems.
- AI agent: calls tools and plans steps. Flexible, but costlier and harder to control, so it needs strict permission management.
- In most cases the right answer is a predefined, AI-assisted workflow, not a fully autonomous agent.
- You are responsible for what your bot says, as a Canadian tribunal made clear in 2024.
What’s the difference between a template chatbot and an AI agent?
Everything on the market gets an “AI” label these days, so it helps to separate three levels. The dividing line isn’t whether there’s a language model inside. It’s who decides the next step and what the system can access.
1. Rule-based chatbot (decision tree)
Buttons, keywords, pre-written branches: “Shipping” → “How long?” → canned answer. It says exactly what you wrote, nothing more and nothing less. When a visitor steps off the script, the bot gets stuck, but at least it doesn’t make things up.
2. Off-the-shelf AI assistant
This is what most packaged products are today: a large language model answering from the documents you uploaded or from your website. It speaks naturally and understands a wide range of questions, but typically it only reads, and only what it was given. It usually has no real, live connection to your order system, your CRM or your booking calendar.
3. AI agent
Anthropic’s “Building effective agents” guide draws a useful line: workflows are systems where the language model and tools are orchestrated through predefined code paths, while agents are systems where the model dynamically directs its own process and tool use. The same guide notes that for many applications, a single well-optimised model call with retrieval and in-context examples is usually enough. That matches our experience.
| Aspect | Template chatbot | Off-the-shelf AI assistant | AI workflow / agent |
|---|---|---|---|
| Who decides the next step? | The script | The model, but it only writes text | The code (workflow) or the model (agent) |
| What can it access? | Only the answers written into it | Uploaded documents | Your systems, via APIs |
| Predictability | Complete | Medium | Workflow: high; agent: lower |
| Typical failure | “Sorry, I didn’t understand” | Confident but wrong answer | Wrong action with real consequences |
When do you need which?
A template chatbot (or nothing) is enough if…
- questions boil down to 10–20 recurring topics (opening hours, shipping, returns);
- the answer doesn’t depend on the specific customer or order;
- you work in a regulated field where a single badly worded sentence can cause trouble.
Honestly, in these cases a well-written FAQ page often beats any bot. Search engines and AI search can quote it, too, which they can’t do with content locked inside a chat widget. We cover this in more detail in our piece on technical SEO and AI discoverability.
You need an AI workflow if…
- the answer requires live data: order status, stock levels, available slots;
- the steps are known and only the input varies (for example, pulling the order number and reason out of a free-text return request);
- a real action happens at the end, but rules or a person sign it off.
This is the typical case for online shops and service businesses, and in our view it’s where most of the untapped value lies. The model understands the text; the code guarantees that only permitted steps happen.
You need a genuine agent if…
- the order of steps can’t be written down in advance because what comes next depends on the result;
- information has to be gathered from several sources and weighed;
- the task is mainly internal (analysis, reporting, research), where a mistake can be corrected and doesn’t happen in front of a customer.
Where do off-the-shelf “AI assistants” fail?
They’re disconnected from reality
The most common failure is mundane: the bot works from last year’s PDFs and the copy on your website, while the customer asks where their parcel is and whether size 42 is available. If the system can’t reach that data, it can do one of two things: hand off to customer service (a template chatbot could have done that) or make something up.
They’re confidently wrong, and you pay for it
In Moffatt v. Air Canada (2024 BCCRT 149), the airline’s chatbot gave a passenger incorrect information about bereavement fares. The company argued it couldn’t be held liable for what its chatbot said. British Columbia’s Civil Resolution Tribunal rejected that and found Air Canada liable for negligent misrepresentation. The lesson is simple: to a customer, the bot is your company’s voice, and you answer for the promises made on your website.
They’re given too many permissions
Once an off-the-shelf tool gains “agentic” features (issuing refunds, sending emails, changing records), a new class of risk appears. The OWASP Top 10 for LLM Applications again put prompt injection in first place in its 2025 edition: a user, or a document the model reads, slips in instructions that the model then executes as commands. The list separately names excessive agency, where the model gets more tools, permissions or autonomy than its job requires. With a packaged product, you often can’t even see exactly what it has access to.
You can’t measure or improve them
Without access to conversation logs, you don’t know where the bot goes wrong, which questions it can’t answer, or when visitors give up. In a custom system it’s the first thing we build in, because without it there’s nothing to improve.
They forget to say they’re machines
Article 50 of the EU AI Act (Regulation (EU) 2024/1689) requires AI systems that interact directly with people to be designed so users know they’re dealing with an AI, unless that’s obvious from the context. The obligation applies from 2 August 2026. If your bot introduces itself as “Kate from customer support” and nothing reveals it’s an AI, now is the time to review it.
How do you decide what you need?
Before choosing any tool, answer these questions, ideally based on real customer messages:
- What do people actually ask you? Go through a few hundred emails or chat messages and group them. You’ll see how much is routine and how much is one-off.
- Does the answer need live data? If so, an off-the-shelf knowledge-base bot won’t cut it.
- Does a real action take place? If so, write down in advance what the system may do on its own and what needs human approval.
- What’s the worst thing it could say? If that’s a legal or financial risk, narrow its scope and give it a clear “I don’t know, let me pass you to a colleague” exit.
- Who reads the logs? If nobody does, the bot will repeat the same mistakes for months.
If your answers point to a process rather than a conversation, you don’t need to buy a chatbot. You need to build an integration. For where AI pays off in a business more broadly, see AI in your business in 2026. For a concrete example running on a company’s own data, read our piece on building your own AI marketing assistant.
Frequently asked questions
Is a ChatGPT-based chatbot already an AI agent?
Not necessarily. If it only answers from your uploaded documents, it’s a well-spoken knowledge-base bot. It becomes an agent when it uses tools and decides the order of steps itself.
Is a template chatbot more reliable than an AI solution?
It’s more predictable, because it can only say what you wrote. In return, it hits its limits quickly. Often the best setup is a mix: fixed answers for sensitive topics, AI for the rest.
Do I have to disclose that my chatbot is AI?
From 2 August 2026, Article 50 of the EU AI Act requires that users know they’re interacting with an AI unless it’s obvious. The obligation falls primarily on the system’s provider, but it shows up on your website, so it’s worth checking.
How safe is it to let an AI agent take actions in my online shop?
As safe as its permissions are narrow. The agent should only access the operations its task requires, and irreversible steps should need human approval.
If it turns out that your conversation needs a real process behind it, we build that as custom software: connected to your own systems, with logging and clearly scoped permissions. If you’d first like to see how your business shows up in search and AI answers, request a free SEO and AI visibility audit.