What an AI agent actually is
The distinction that matters, and it is simpler than the marketing suggests.
| What it does | Who decides the steps | |
|---|---|---|
| Chatbot | Answers questions | Nobody — it responds |
| Automation | Runs a fixed sequence | You, at build time |
| Agent | Pursues a goal using tools | The model, at run time |
That third row is the whole thing, and it is both the appeal and the risk. An agent can handle a case you did not anticipate. It can also take a path you did not anticipate, which is why the tools it is given must be narrow and the things it may do alone must be few.
Do you actually need an agent?
Usually not, and this is the most expensive misjudgement in the category.
The test: do the steps genuinely vary between runs? If the same five things happen every time in the same order, that is an automation — build it as one. It will be cheaper, faster, entirely predictable, and you will be able to say exactly what it did. Dressing it up as an agent adds cost and removes certainty.
Where an agent is right: research and gathering across sources that differ every time; triage where the next step depends on what was found; and work where the goal is stable but the route is not. Those are real, and they are a minority of what gets asked for.
Cost, and the risk that comes with it
| Scope | Build | Running |
|---|---|---|
| Single-purpose agent, few tools | £4,000 – £8,000 | Usage, typically low |
| Multi-tool agent with human approval | £8,000 – £20,000 | Usage plus review time |
| Agent acting without approval | Not recommended for most businesses | — |
| Cybrial | Priced on application, from the hours | Your own accounts |
Source: UK market ranges, September 2026.
The third row is deliberate. An agent that can act on the outside world without a person approving the action is a category of risk most small businesses have no reason to take, and I would talk you out of it. Human approval costs a few seconds per run and removes almost all of the downside.
