An AI agent is software that can read a situation, decide what to do next, and take the action - inside limits you set. If it can't take the action, it's a chatbot. If it can't decide, it's an automation. Both are useful. Neither is an agent.
When we sit down for a time audit, the first question we usually get isn't about price or setup. It's some version of "so what actually is this stuff?" Fair question. The industry has done a terrible job of answering it.
The three-part test
Software earns the name "agent" when it can do three things without you in the loop:
- Read the situation. A new enquiry arrives. Is it a quote request, a complaint, or spam? An agent works that out from the message itself, not from which button someone pressed.
- Decide what to do. A quote request gets a reply with available times. A complaint gets flagged to you immediately. Spam gets binned. The decision changes with the input.
- Take the action. It actually sends the reply, books the slot, or updates the record - it doesn't just draft a suggestion and wait for applause.
Most of what's sold as "AI for business" fails the test. A chatbot on your website reads and decides, but the action is always the same: type words back. A Zapier-style automation takes actions, but it can't decide - the same trigger always fires the same steps. An agent is the overlap: judgement plus hands.
What agents are genuinely good at
The sweet spot is work that's repetitive in shape but variable in detail. Every missed call needs a text back, but each text should mention what the caller actually asked about. Every overdue invoice needs a nudge, but a first-time late payer and a repeat offender deserve different tones. That's exactly the territory where old-school automation is too rigid and hiring a person is too expensive.
What they're still bad at
Anything where the cost of a wrong answer is high and the answer isn't checkable: pricing a complex job, giving advice with legal weight, handling a genuinely upset customer. In our own businesses, those go straight to a human, every time. Any provider who tells you otherwise is selling the demo, not the Tuesday.
That's also why every system we build starts in draft-and-approve mode - the agent does the work, you sign it off, and it only gets autonomy once it's earned it.