How to Try an AI Agent on a Real Errand (Moving Quotes, a Dinner Booking) Without Handing Over the Keys
It booked the table. It also suggested a restaurant that closed years ago.
At a glance
First step Pick a low-stakes errand. Getting quotes, finding a table for Saturday, comparing two plumbers. Not paying bills, not selling things, not anything with your bank login.
Saves: The phone calls and forms behind small errands · Tools: An AI agent such as Meta's Muse or ChatGPT's agent mode
Example result A good first run ends with the agent showing you two or three options and waiting for you to pick, not with a confirmation email you didn't expect.
Only have five minutes? Find the setting that makes your assistant ask before it buys, sends or shares anything, and switch it on.
What to do
- Pick a low-stakes errand. Getting quotes, finding a table for Saturday, comparing two plumbers. Not paying bills, not selling things, not anything with your bank login.
- Switch on 'ask before acting'. Look in the agent's settings for anything that makes it confirm before it buys, sends messages or shares your details. Turn it on.
- Give it the least it needs. A first name and a phone number for quotes is plenty. Hold back your address, card and passwords until you've seen how it behaves.
- Ask for options, not a done deal. Tell it to come back with two or three choices and wait for your pick. Use the prompt below.
- Check what it found. Call the restaurant, look up the company. Agents can invent phone numbers and recommend places that no longer exist.
The prompt
A prompt is the instructions you give the AI. Anything in [square brackets] is a blank for you to fill in. First time? Start here
Please help me with one errand: [errand, e.g. get quotes from two or three local movers for a one-bedroom move on (date) / find a table for (number) on Saturday around 7pm near (neighborhood)]. Before you contact anyone, book, buy or share any of my details, stop and ask me. Come back with two or three options, each with its name, website, phone number and price if known, and say where you found each one. If you're not sure something is current, say so.
- Press Copy prompt.
- Open ChatGPT, Claude or Gemini.
- Paste it into the message box and fill in the [blanks].
Keep it private. Before you paste: leave out passwords, codes, and account, card, ID and policy numbers. Swap real names for "my mother" or "the company."
Share only what the errand needs. Agents act on your behalf, so a mistake goes out under your name.
What to check yourself
- Agents can make things up, including phone numbers and businesses. Check anything you'll rely on.
- 'Ask me first' settings are only as good as their testing. Try yours on something unimportant before you trust it.
- Don't give an agent access to accounts, money or messages you can't afford to have mishandled, and know how to stop it mid-task.
What happenedThe story behind it
AI "agents" don't just answer questions; they go off and do things: fill in forms, book tables, buy things. In September, Yahoo Finance's Daniel Howley put Meta's new Muse agent through ordinary errands. It put together a news digest, booked a dinner reservation, and requested moving quotes; two moving companies called him within minutes. When he asked it to buy a video game, it checked with him before charging his card.
It also got things wrong. It recommended a restaurant that had long since closed, and when pressed, admitted it had made up a phone number. NPR reported a sharper lesson the same month: a reviewer who let Muse sell items on Facebook Marketplace found it accepted an offer below his minimum and sent the buyer his home address without asking first. Meta said its review found no breach of its privacy controls.
So the useful version is small and supervised. Pick an errand where a mistake costs you a phone call, not money or privacy, make the agent ask before it acts, and check its work before you rely on it.
The verdict
A keen new assistant: fast, eager, and in need of supervision until further notice.
Sources
How sure we are: Based on published reports and official advice. We haven't tested this one ourselves. What this means