What is an AI agent, and when is one actually worth building?

A chatbot answers questions. An AI agent takes actions: it can look up an order, check stock, issue a refund inside your rules, escalate to a human, or write a record back into your system. The distinction matters commercially, because only the second one removes work.

The three things every useful agent has

• Grounding. It answers from your data — catalogue, order history, policies — not from general internet knowledge.
• Tools. It can call your systems, with permissions that are scoped and logged.
• Guardrails and handoff. It knows what it must not do, and it hands a conversation to a person cleanly when it should.

An agent missing any one of these tends to produce confident, unusable answers.

When an agent is worth building

Look for work that is high volume, rule-bound and currently done by a person reading from one screen and typing into another. Order status, returns eligibility, stock and delivery questions, vendor onboarding checks, and first-line qualification of sales enquiries all qualify.

When it is not

If the work is low volume, or the rules are genuinely judgement-based, or the underlying data is unreliable, an agent will amplify the mess rather than remove it. Fix the data first.

How to measure it

Pick the number before you build: deflection rate, handling time, conversion on assisted sessions, or hours returned per week. If no number moves after a month in production, the agent was pointed at the wrong problem.