DROOM Get the guide

Free guide · 22 minute read

AI agents at work: what to hand over and what to keep

For owners, no tech needed. Try the best prompt from it right now, then open the whole guide free with your email.

The plan

  1. Read the one-page difference between a chatbot and an agent (section 1), so you know what you're being sold.
  2. Give every agent you try one standing rule: it drafts, you send. The prompt for that is right below.
  3. Hand over one job from section 5 for two weeks. Check its work every day, then score it with the keep-or-drop table in section 8.

Try this now: the approval rule

An agent doesn't only write. It does things, and that's both the point and the risk. Before you let one near your email, calendar or accounts, give it this rule. Paste it into the agent's saved instructions (some tools call them custom instructions, preferences or project instructions), or at the top of any task you give it.

You are helping me run [business name], a [type of business] in [city].
Standing rule for every task: you prepare, I approve.
- Draft emails, messages and replies, but never send them. Show me the draft and wait.
- Never buy, pay, book with a deposit, sign, accept terms, or share a card or account number.
- Never delete, archive or move anything without asking me first.
- Never contact a customer, supplier or staff member unless I've approved that exact message.
- If a website asks for a password or a code, stop and tell me.
- If you're unsure what I want, ask one question before doing anything.
At the end of each task, give me a short list: what you did, what you're waiting on me for, and anything that looked wrong.

Now test it. Ask the agent to "send a thank-you email to [a friend or your own second address]". It should stop and show you the draft. If it sends anyway, that tool isn't ready for your customer inbox.

1. Start here if you've never used AI

An AI chatbot is a website or app you type to, and it writes an answer back. ChatGPT (made by OpenAI), Claude (Anthropic), Gemini (Google), Copilot (Microsoft) and Meta AI all started this way. You ask, it writes, and you decide what to do with the words. Nothing leaves your screen unless you copy it somewhere.

An AI agent is a chatbot that's been given hands. You give it a goal, say "find three suppliers for this part and compare delivery times", and it works out the steps and does them. It opens websites, reads pages, fills in forms, and looks in your email or calendar if you've connected them. Then it comes back with a result. Some agents keep working after you close the app, or run on a schedule.

That one difference changes how careful you need to be. When a chatbot gets something wrong, you get a bad paragraph. When an agent gets it wrong, the bad paragraph can end up in a customer's inbox.

The one-page difference

ChatbotAgent
What you give itA question or a request for writingA goal with steps
What it gives backWords on your screenA finished task, or a task waiting for your OK
Can it open websites?Some can look things upYes, and click, scroll and fill in forms
Can it use your email or calendar?Only if you connect themUsually that's the point
Can it act without you watching?NoSome can, on a schedule or in the background
What goes wrongA wrong fact in a draftA wrong action: a sent message, a changed booking
Your jobRead before you copySet limits, approve actions, check the log

Agents are slow. One task can take several minutes while it visits pages and checks things. That's normal.

They still make things up sometimes. Everything in my other guides about checking facts applies here. The difference is you also check what it did, not only what it wrote.

Most of the agent features in this guide are on paid plans. The free plans of ChatGPT and Claude are fine for learning the chatbot side (see ChatGPT and Claude for beginners). Section 3 says which agent features I could confirm, and which you can use in Canada.

Keep private details out until you've read section 4. Connecting your inbox hands the tool your customers' names, addresses and messages. Make that call on purpose. Don't let it happen because you clicked through a sign-in screen.

Words you'll see

WordWhat it means
AgentAn AI tool that carries out steps for you as well as writing.
Agent mode / Work modeA switch inside a chatbot that turns on the agent.
Connector / connected appA link you set up so the AI can read or act in another app, like Gmail or Google Drive.
PermissionWhat you allow a connector to do: read only, or read and change.
Approval / confirmationThe moment the agent stops and asks "OK to do this?" before acting.
TaskOne job you give an agent. Some tools let you schedule tasks to repeat.
Activity logA list of everything the agent did. Read it.
Take overTaking control of the agent's browser yourself, usually to type a password.
Prompt injectionA trick where a web page or email hides instructions aimed at the agent, trying to get it to do something you didn't ask for.

Get the full guide

That was the free part. Put your email in and the other 9 sections open right here on this page, with a copy button on every prompt. The PDF is yours too.

Your email gets you this guide now and an occasional email from Joe after, each with an unsubscribe link, and your details are never sold or shared. Privacy.

More free guides

Eleven more like this one, all written for owners who are new to AI.

Last updated 2026-09-25