What is an AI agent, and do you actually need one?

The vocabulary in this market is deliberately blurry, and the blur is expensive: “agent” on a proposal can multiply the price by ten without changing what gets delivered. Here is the distinction that actually matters.

The difference in one line

An assistant answers. An agent acts.

An assistant retrieves information and gives it to a human, who decides what to do. An agent takes steps in your systems — books the appointment, opens the ticket, submits the claim, updates the record — within rules you set, and escalates when it hits something it should not decide.

That difference is not cosmetic. An agent needs permissions, an audit trail, override authority and a defined escalation path, because it is doing things rather than suggesting them. That is most of why it costs more.

The honest test

Ask: after this system gives its output, does a human still have to do the work?

If yes, you need an assistant, and you should pay assistant prices — low thousands. If the work itself should be completed without a person touching it, you need an agent, and you should expect to pay for permissions, logging and human-in-the-loop design.

Most businesses asking for an agent need an assistant first. Not because they are wrong about the destination, but because an assistant on the same workflow tells you within two weeks whether the underlying content and process are good enough to automate — which is the thing that actually decides whether the agent will work.

What to ask a vendor selling you an agent

• Which specific actions will it take without a human approving each one?
• Who can override it, and is that named in writing?
• What does it log for each action — input, rule applied, output, reviewer?
• What happens when it is uncertain? “It escalates” is not an answer; ask to see the threshold.

If those four answers are vague, you are being sold an assistant at agent prices.