The honest answer is narrower than the marketing and wider than the scepticism. An AI agent can carry out a defined job end to end inside your systems — read something, decide what it is, do the next thing, and record it. What it cannot do is exercise judgement about your business.
That distinction does more work than any capability list, so it is worth getting straight before the list.
What it does well
- Read an incoming message and work out what it is actually about, including when it follows no template.
- Ask the follow-up questions a person would ask, and wait for the answers.
- Look something up across systems — a job status, an availability, a past order — and answer from what it found rather than from memory.
- Take a bounded action: book the slot, create the record, send the confirmation, update the stage.
- Pull structured information out of an unstructured document and put it where it is needed.
- Notice that something has not happened — a quote unanswered, a confirmation missing — and chase it.
- Hand over to a person, with the context attached, the moment it hits something outside its rules.
Every one of those is checkable after the fact, which is not a coincidence. The jobs agents do well are the jobs where you can look at the result and say whether it was right.
What it does badly
- Anything with no definition of done. "Improve our marketing" is not a task an agent can finish.
- Commercial judgement — what this client is worth, whether to hold your price, when to walk away.
- Work where being wrong is expensive and hard to reverse, because the loop that makes it useful means a bad step gets acted on before anyone sees it.
- Situations needing a human relationship: an apology, a negotiation, a complaint that has gone past the facts.
- Anything requiring knowledge it has no access to. An agent cannot know what was agreed on a phone call nobody wrote down.
The test
Ask two questions about the job. Could you write down what "done correctly" looks like? And if it got it wrong, would you find out, and could you fix it?
Two yeses and it is agent-shaped. A no to either and you want a person, or a person checking the output before it goes anywhere.
Booking an appointment passes: done means the right slot in the right calendar, and a mistake surfaces immediately and costs a phone call. Deciding what to charge fails the second question badly — a wrong number lands in a client inbox and no apology fully retrieves it.
What changes the answer
Capability is mostly set by access, not by the model. An agent that cannot see your calendar cannot book anything, however clever it is. Most of the difference between a useful agent and a disappointing one is how much of your actual business it can reach — which is an integration question, and an unglamorous one.
The second thing that changes the answer is permission. The same software given a short list of allowed actions behaves predictably; given the run of your systems it eventually does something nobody sanctioned. Narrow is not a limitation here, it is the design.
The definition, and why the permissions matter more than the intelligence: What is an AI agent?
A fuller inventory of the jobs worth handing over: 25 business processes you can automate with AI