What is an AI agent? A 5-minute primer

Five minutes, no jargon, with a worked example. By the end you’ll know what an AI agent actually is, what the four common shapes look like, and the six red-flag phrases that mean a vendor doesn’t.

The plain-English definition

An AI agent is a language model that, given a goal, decides which tools to call, calls them, reads the results, and decides what to do next - in a loop, until it judges the goal met.

Three words do all the work in that sentence: tools, loop, and goal.

• Tools are functions the model can call: send an email, read a file, query a database, hit an API, run a calculation. Without tools, you have a chatbot.
• Loop means the model can run again on its own output - read what happened, decide what to do next. Without a loop, you have a one-shot script.
• Goalis a clear stop condition. “Send the customer a quote and confirm it landed” is a goal. “Be helpful” is not.

The diagram

The same loop, repeated. The model isn’t smarter than a chatbot - it just gets more turns and the ability to act between them.

The four common shapes

A worked example: Sue’s plumbing

Sue runs a 4-person plumbing business. Quotes are eating her evenings. Each quote takes 15 minutes - read the enquiry, check the diary for availability, look up parts, write the reply. Twelve enquiries a week is three hours of Sue’s Sunday gone.

The non-agent version (a chatbot)

Sue installs a chatbot on her website. It can answer “what areas do you cover?” but it can’t check her diary, can’t look up parts pricing, can’t actually quote. Sue still does the quote. Saves 30 seconds.

The agent version

Sue’s assistant agent has three tools: read_calendar, lookup_parts_price, and draft_quote_email. When an enquiry arrives:

Quote time drops from 15 minutes to 2. Sue still hits send - the agent never emails the customer directly because the cost of a wrong quote is high. That’s the human gate.

Six red-flag phrases

• “Fully autonomous” - for any business-critical action, autonomous means unsupervised, which means undiagnosed when it goes wrong.
• “Our agent never hallucinates” - language models always can. The right framing is “here’s how we reduce and detect it”.
• “It learns from every interaction”- usually means “we keep your data and may train on it”. Ask where the learning happens and who owns it.
• “Replaces your team” - sales-deck phrasing. Useful agents augment a small team, they don’t replace it.
• “Plug and play, no setup” - a real agent needs to know your tools, your data, and your guardrails. Setup is the job. Skipping it is the failure mode.
• “Proprietary AI”- almost always wrapping someone else’s frontier model (Claude, GPT, Gemini). Ask which one. If they won’t say, walk away.

The honest one-question test before building one

Could a competent junior assistant do this task with the same tools, given clear instructions? If yes, an agent will probably help. If no - because the task needs judgement, relationships, or context not written down anywhere - you’re trying to skip the wrong step.

Going deeper

• Agentic AI is having its moment - what’s real and what’s not
• The AI readiness self-assessment - score whether your business is set up to deploy one today.
• Prompt patterns - the prompt shapes that show up inside most working agents.