An AI agent is software that you give a goal to — "answer this enquiry and get it booked" — rather than a list of steps. It works out the steps itself, and it is allowed to take real actions in real systems to get there: send the reply, read the calendar, write the appointment, update the record.
That last part is what separates an agent from everything else wearing the AI label. Plenty of software can produce text about your business. An agent changes something in it.
The three things that make it an agent
- A goal, not a script. You describe the outcome you want. The software decides the order of operations, and can take a different route when the situation is not what it expected.
- Tools it is allowed to use. Access to your calendar, your CRM, your inbox, your job management system. Without tools it can only talk.
- A loop. It acts, looks at what happened, and decides what to do next — rather than running once and stopping.
Take any one of those away and you have something else. A goal without tools is a chatbot. Tools without a loop is ordinary automation. A loop with no clear goal is a way to burn money.
What that looks like on a normal Tuesday
A form comes in at 9:40pm. An agent reads it, works out that it is a real enquiry rather than a supplier pitch, replies within seconds asking the two things you always have to ask, waits for the answer, checks which of your crews covers that area, offers three slots that are genuinely free, books the one the customer picks, and puts the job in your system with the notes attached.
None of those steps is clever on its own. The value is that the whole sequence happened at 9:40pm without anybody being awake for it, and the customer was not left waiting until Thursday.
The question that actually matters
Most conversations about agents get stuck on how capable the model is. In practice that is rarely what decides whether the thing is useful or dangerous. What decides it is scope: what is this allowed to do, and what does it do when it is unsure?
An agent with narrow permissions and a clear escalation rule is worth more than a clever one with the run of your systems.
A well-built agent has a short list of actions it can take, a shorter list of things it must never do without a person, and an obvious hand-off when it reaches either edge. Booking an appointment is a good candidate: the outcome is checkable, and a mistake is a phone call to fix. Agreeing prices or issuing refunds is not, at least not to begin with.
Where agents are genuinely weak
They are poor at open-ended work with no definition of done. They are poor at commercial judgement — what this customer is worth, whether to hold your price, when to walk away. And they are poor anywhere being wrong is expensive and hard to reverse, because the loop that makes them useful also means a bad decision can be acted on before anyone sees it.
The sensible pattern is to start with one job that has a clear finish line, watch it for a few weeks, and widen only once you have seen how it behaves on the odd cases. Anyone selling you an agent that runs your whole operation from day one is selling you a demo.
The difference between an agent and the chatbot you may already have: AI agents vs chatbots
What we build, and the limits we build in: AI agent development