An AI agent is an AI that you give a goal to, rather than a question. It decides the steps, uses your tools (email, calendar, spreadsheets, a browser) and comes back with finished work or a draft for you to check. Most small businesses don't need one yet, and the ones that do need it for one repeatable job that eats a chunk of their week.
What is an AI agent, in plain English?
An agent is a chatbot with a toolbelt and a to-do list. You say "find three venues in Perth, put them in a table with sources and draft an enquiry for me to review". It goes off, opens a browser, reads pages, builds the table and writes the draft.
Four parts make that work: a model (the brain), tools it's allowed to use, somewhere to run, and your instructions. The tools and permissions decide how useful it is and how risky it is. A tool that can read your inbox is a very different thing from one that can send from it.
If you want the longer version of how those parts fit together, I broke it down in ChatGPT Dots vs Grok Bot vs Meta Muse vs Hermes.
Chatbot, automation or agent: what's the difference?
A chatbot answers. An automation follows a fixed recipe. An agent works out the recipe as it goes.
| Chatbot | Automation | Agent | |
|---|---|---|---|
| You give it | A question | A trigger and fixed steps | A goal |
| It decides | What to say | Nothing, it follows the rules | Which steps, which tools, what order |
| Example | Ask Claude to rewrite an email | A new booking sends your intake form | Research each new client and drop a one-pager in Notion |
| Breaks when | You ask something vague | One step changes | It misreads the goal |
The same tool can be all three on different days. What matters is who decides the next step: you, the rules, or the AI.
I wrote up the automation side with real examples in 4 business automations for non-technical founders, including the one place I do let an agent decide things.
Does my small business actually need an AI agent?
Probably not today. You need one when a job repeats, needs a little judgement, and eats time you can name.
Here is how I'd sort it:
- Rules are enough. A new enquiry gets a confirmation email. That's an automation, and usually the right first thing to automate. An agent adds cost and surprises for no benefit.
- You need a draft. Rewrite an email, summarise a call. A chatbot or a Claude Project does it.
- It needs judgement across several steps. Read the intake form, research the company, summarise what matters. That's agent territory.
The test is boring on purpose: can you write down the job in five lines, and could a new hire do it from those five lines? If yes, an agent can probably do a first pass. If you can't describe it, no agent will rescue you.
Most women I work with get further in a month from one good automation and a well-set-up chatbot than from a fleet of agents. The agents come after.
What is a safe first job to give an agent?
A read-only, recurring job where the output is a draft you review. Think weekly briefing, not "run my inbox".
My starting suggestion: every Monday, read your calendar and your sent emails from the last week and write a one-page summary of who you owe a follow-up. Claude Cowork can run tasks like this on a schedule, using the same connectors as your normal chats, and Anthropic's help page lists daily briefings and weekly reports as examples (schedule recurring tasks in Cowork, checked 23/10/2026). If you're new to it, start with my Claude Cowork tutorial.
Why this job works:
- It only reads. Nothing leaves your business.
- It's weekly, so you spot a bad output fast.
- You already know what a good summary looks like, so you can judge it.
What should an AI agent never do unsupervised?
Anything you can't undo, anything that talks to a client and anything that touches money. Draft those. Don't send them.
- Send emails or DMs to clients. It can write the reply. You press send.
- Spend or move money. No payments, refunds, invoices or ad budgets.
- Delete or overwrite files. Give it a copy, not your only version.
- Post publicly. A wrong post is live before you've noticed.
- Make promises. Prices and deadlines are yours to set, and so is what's in scope.
Do the 10-minute version of this today. Open the connectors on whichever AI tool you use and switch off every action that sends, deletes or posts. Leave read and draft switched on.
The defaults are not your friend here. When I connected Gmail to Grok Bot, mine was sitting at 31 of 31 actions enabled, send and trash included. Writing "don't send emails" into your instructions is tempting, but an instruction can be misread. A switched-off send button can't be. The screenshots are in my Grok Bot setup guide.
Quick answers
Is an AI agent the same as ChatGPT?
Not quite. A normal ChatGPT chat is a chatbot: you ask, it answers. ChatGPT also has agent-style features, like ChatGPT work, where you brief it on an outcome and it carries out the steps across your apps.
Do I need to code to use an AI agent?
No. The managed ones are set up by connecting your apps in a few clicks. Coding only comes in when you want to choose every part yourself, which most small businesses don't.
Will an agent replace my VA?
It won't replace a person who handles judgement calls and client relationships. It can take over a narrow, repeatable task, like a weekly summary or first-pass research.
What's the risk of getting this wrong?
An agent with too many permissions can send, post or delete before you've looked. That's why the first job should be read-only, with you as the last step.
Get help setting up your first one
A member posted: "I'm a solo consultant. As of today I have AI staff. 🌊 Two actual agents I built for my own client delivery: one that onboards a new client, one that offboards them". That's one person's build, not a typical result, but it's the kind of thing we work on.
Inside Wright Mode you can watch me build in the monthly live Build, bring your own job to the fortnightly Co-Work and Q&A, and ask someone when the permissions screen stops making sense.
