The phrase “AI agent” has been stretched to cover a chatbot with a nicer font, and that is why most small businesses cannot work out whether they need one.
Here is the distinction that actually matters. A chatbot produces text and stops. An agent takes an action: it books the appointment, issues the refund, updates the record, sends the email, and then checks whether that worked.
Everything useful about agents, and everything risky about them, follows from that one difference. Software that writes a wrong sentence wastes ten seconds. Software that takes a wrong action has done something to your business that somebody now has to undo.
The four kinds you will actually be sold
Vendors rarely say which of these they are. It is usually obvious within two questions.
| Type | What it does | Where it earns its money |
|---|---|---|
| Retrieval | Answers from your documents and data | Support, internal “where is the policy on…” |
| Workflow | Runs a fixed sequence across systems | Onboarding, invoicing, order routing |
| Triage | Reads incoming work and routes or drafts | Inboxes, enquiries, tickets |
| Autonomous | Decides its own steps toward a goal | Rarely, in production, in a small business |
The first three are worth money today and are largely a solved engineering problem. The fourth is where the marketing budget goes and where the failures live.
If a vendor is pitching the fourth for a job the third would do, that is not sophistication, it is a larger invoice and a wider blast radius.
What they are genuinely good at
Three properties, and a job that has all three is a good candidate.
High volume and low variance. Two hundred near identical enquiries a week is ideal. Six unusual ones is not; you will spend more time on exceptions than the agent saves.
A checkable output. The agent files the invoice against a purchase order, and the numbers either reconcile or they do not. Jobs where nobody can tell whether the output is right are jobs where the agent’s errors accumulate silently.
A cheap failure. A misrouted enquiry costs a minute. A mispriced quote sent to a customer costs a relationship. Start where being wrong is survivable.
Enquiry triage, appointment scheduling, document extraction, first line support against a real knowledge base, and routine data entry between systems that do not talk to each other. That is the honest list. It is less exciting than the pitch and it is where the returns are.
Where they fall over
The last ten per cent. An agent that handles ninety per cent of a workflow has not removed ninety per cent of the work. It has removed the easy ninety and left a human doing only the hard ten, with no warm-up and no context. Several businesses we have spoken to found their team slower after deployment, because the remaining work was uniformly difficult.
Confident wrongness. An agent does not hesitate. It does not flag that it was unsure, because it was not unsure. Anything it touches needs a check that does not rely on the agent reporting its own doubt.
Integration, which is most of the cost. The model is the cheap part. Connecting it to a booking system, a CRM, an accounts package and an email provider, each with its own auth and rate limits, is where the budget goes. We set out the full picture in what an AI agent actually costs a small business.
The wall. Prototype in a weekend, then discover the last twenty per cent needs real engineering. That pattern is common enough to have a name, and we wrote it up in the wall problem with AI app builders.
Data protection, which is not optional. If the agent touches personal data, UK GDPR applies exactly as it would to any other processing. The ICO’s guidance on AI and data protection ↗ covers lawful basis, transparency and the rules on automated decisions. A vendor who cannot tell you where your data is processed and for how long has answered the question.
Five questions that separate a real product from a demo
- Show me it failing. A vendor who has never seen their agent fail has not run it anywhere serious. Ask what it does when it cannot complete a task.
- What does a human see, and when? There should be a named point where work is handed back, with the reasoning attached. “It handles everything” is not an answer.
- Where is my data processed, and is it used for training? Get it in writing.
- What happens when your model provider changes the model? Behaviour shifts under you. Ask how they version and test.
- Can I see the audit trail for one real task, end to end? If there is no trail, there is no way to investigate the day it does something odd.
How to start, if you are starting
Pick one job. Measure it first: how many times a week, how long each, how often wrong today. Without that baseline you cannot tell afterwards whether the agent helped, and almost nobody collects it.
Run the agent alongside the human for a fortnight rather than instead of them. Compare. Then decide.
The businesses that get value out of this are not the ones that moved fastest. They are the ones that picked a boring, countable job and could prove the difference at the end of the month. If you want a hand working out which job that is for you, get in touch or look at how we work.
Frequently asked questions
What is an AI agent in business terms?
Software that takes actions in your systems to complete a task, rather than just generating text. The practical test is whether it can change something: book, send, update, refund. If it only produces words for a person to act on, it is an assistant, not an agent.
Are AI agents worth it for a small business?
For the right job, yes. High volume, low variance work with a checkable output and a cheap failure mode pays back quickly. Low volume, high judgement work generally does not, because the exception handling costs more than the automation saves.
What is the difference between an AI agent and a chatbot?
A chatbot answers. An agent acts. A chatbot that can also book the appointment, write to the CRM and send the confirmation has crossed into being an agent, and it should be assessed with the questions above rather than on how well it converses.
Do AI agents replace staff?
Rarely, in a small business. They remove the routine share of a role and leave the difficult share, which usually means the same people doing harder work rather than fewer people. Budget on that basis rather than on a headcount saving.
Is using an AI agent GDPR compliant?
It can be, and it is not automatic. You still need a lawful basis, transparency with the people whose data is processed, and care around automated decisions with legal or similarly significant effects. The ICO publishes specific guidance on this and it is worth reading before you sign, not after.
How long does it take to deploy an AI agent?
A demo takes days. A production deployment against real systems, with error handling, audit and a human handover path, is usually weeks to months, and almost all of that time is integration rather than anything to do with the model.
The short version
An agent is software that acts. Judge it on what happens when it acts wrongly, not on how well it talks.
Pick the boring job. Measure it first. Keep a person in the loop at a named point. Most of the cost is integration, and most of the risk is the ten per cent it hands back.