AI agents explained: what they can do today, and where to be careful
"Agent" is one of the most used words in AI right now. Strip away the hype and the idea is simple: an AI that can take actions to finish a task, not just write an answer. Here's what that means in practice.
Chatbot vs agent
A chatbot responds to a message. You ask, it answers, and you decide what to do next. An agent is given a goal and works towards it in steps. Depending on the tool, it may search the web, open pages, click buttons, fill in forms, read and write files, or use apps you've connected, checking its progress as it goes.
Think of the difference between asking a colleague "what are some good venues for a team lunch?" and "find three venues near the office that can seat 12 next Thursday, and put them in a table with prices and links."
What agents are good at today
- Research runs: collecting the same information from several websites into one table.
- Admin with clear steps: organising files, renaming documents, drafting a batch of similar replies.
- Comparisons: checking products or services against criteria you set.
- Preparation: filling in forms up to the point of submission, so you can review and press the final button yourself.
Where they still struggle
- Vague goals. "Sort out my holiday" is too open. "Find return flights under $600 for these dates" is workable.
- Unfamiliar or messy websites. Pop-ups, logins and unusual layouts can trip them up.
- Judgement about people. Tone, relationships and sensitive decisions still need you.
- Knowing when they're wrong. Agents can confidently report success on a task they got slightly wrong.
Five rules for using agents safely
- Approve before anything irreversible. Sending, paying, booking, deleting and sharing personal details should always come back to you first.
- Give the least access needed. Don't connect your whole inbox or bank for a task that needs one folder.
- Ask for the plan first. Have the agent list its steps before it starts. It's easy to spot a misunderstanding early.
- Start small. Build trust on low-stakes tasks before handing over anything important.
- Ask for a debrief. When it's done, ask what it did, what failed and what you should double-check.
A prompt to try
I want you to [task]. Before doing anything, write out your step-by-step plan, the information or access you'll need, and any point where you'll need my approval. Stop and ask me before you send, submit, buy, delete or share anything.
Do you need an agent at all?
Often not. If a task follows the same rule every time ("when an invoice email arrives, save the attachment to this folder"), a simple automation in your email or a no-code tool is more predictable. Agents shine when steps vary a little each time and some reading or judgement is involved.
Module 4 of our free course, Future-Ready, has an exercise to sort your weekly tasks into "AI can help", "agent", "automate" and "keep doing yourself".
Want to build these skills properly?
Our course Future-Ready has seven short modules, 32 copy-paste prompts and a 30-day plan.
Get the course

