Chatbot, assistant, agent: the useful distinction
A chatbot answers questions. You ask, it responds, and nothing in the world changes. Useful for support pages, limited everywhere else.
An assistant helps you do something. It drafts the email, summarises the document, suggests the reply. You remain the one who acts.
An agent takes actions on your behalf. It books the appointment, sends the invoice, updates the record, files the follow-up. The difference is not intelligence, it is permission and access.
That is the whole distinction, and it is why the third category is both far more valuable and far more consequential to get wrong. A chatbot that misunderstands you wastes thirty seconds. An agent that misunderstands you sends something to a customer.
What they are genuinely good at
Work that is repetitive, rule-shaped, and time-critical. Answering a call at 11pm, sending a reminder at the right hour, chasing a quote that has gone quiet for four days. Not hard work, but work that requires someone to be present and to remember.
Work that falls between the cracks. Every small business has tasks that are nobody's job: the follow-up nobody sent, the review nobody asked for, the record nobody updated. Agents are good at exactly this because they do not get busy.
Work with a clear definition of done. Book the slot, send the summary, log the call. Where success is unambiguous, an agent can be trusted with it and you can verify it cheaply.
And work at inconvenient volume. Contacting eleven waitlisted customers individually within the hour of a cancellation is trivially easy for software and practically impossible for an owner mid-service.
Where they cost you money
Anything requiring judgement about a person. Whether this customer gets a discount, whether that complaint needs an apology or a refund, whether this caller is in genuine distress. Agents have no read on any of it.
Anything where being confidently wrong is expensive. Quoting a price sight-unseen in a trade where the job varies wildly. Committing to a deadline you cannot see the workload behind. Stating a policy that has an exception you never wrote down.
Anything unsupervised in the first weeks. The failure mode is not dramatic, it is quiet: an agent that has been sending slightly wrong information to customers for a month while everything looked fine. This is why a draft-and-approve step is worth the small friction until you have watched it behave.
And anything nobody checks. An agent running unattended against stale knowledge will keep working perfectly and produce steadily worse outcomes, and by design nothing will alert you.
What this looks like in a real small business
The phone is the obvious first one, because it is where the money enters and where the gaps are most expensive. An agent that answers, qualifies, books, and writes you a summary is the highest-value single deployment for most service businesses. That is covered properly in what is an AI receptionist.
Follow-up is the second. The quote that was never chased and the customer who has not been back in eleven weeks are both pure margin, and both are lost purely to nobody having time.
The admin behind the work is the third: invoices drafted from completed jobs, records updated, appointments confirmed. Individually trivial, collectively the reason owners work evenings. I looked at the invoicing piece in automating your invoices.
Notice that none of these replace the skilled work. They remove the work around the work, which is the part that expands to fill every hour you do not defend.
How to start without regretting it
Pick one job with a clear definition of done, and preferably one that currently does not happen at all. Something that is not happening cannot get worse, which makes it the safest possible starting point.
Keep a human approval step on anything that reaches a customer, and keep it until you are bored of approving things. That boredom is the actual signal that it is ready to run unattended.
Make failures loud. The dangerous failure is the silent one, where everything looks fine and nothing is working. Build in a way that you would notice within a day.
And write down what it knows. Your prices, your service area, what you refuse to take on, and your exceptions. Almost every agent that behaves badly is doing so because it was given vague inputs, not because the technology failed.
The honest summary
AI agents are very good at being present, consistent, and tireless. They are poor at judgement, context, and knowing what they were not told.
That maps neatly onto the shape of most small businesses, where the expensive gaps are almost all availability gaps rather than skill gaps. You are not losing work because you are bad at it. You are losing it because you were on a ladder.
So the sensible ambition is not to replace anything. It is to stop the leaks that exist purely because one person cannot be in two places, and to hand back the evenings currently spent on admin.
If you want to hear one working before forming a view, ask me for a live demo. It answers live, it books against a real calendar, and it is a more useful test than any amount of reading.