AI agents can now act inside your business accounts. Here's a twenty-minute permission list that decides what they can read, draft, do with your OK, and never touch.
An AI agent is an assistant that does things for you, not just writes things for you. Before you connect one to your email, store, or bank account, write a one-page permission list: what it can read, what it can draft, what it can do with your OK, and what it never touches.
That list takes about twenty minutes. It's the most useful AI thing you can do this week, and it's useful even if you never turn an agent on.
For most of the past two years, using AI meant a chat window. You asked, it wrote, you copied and pasted. Nothing happened in the world unless you made it happen.
That's shifting. On September 29, Meta announced Muse for Small Business, which builds on Muse, the personal agent Meta says is available in the US and Canada. You give it a goal and it works on it. Meta says it can flag emails that need a reply and write first drafts, look over your monthly finances and point out expenses that seem unusual, and draft campaigns from your sales and social numbers. It connects to tools a lot of people already use, including Shopify, QuickBooks, Stripe, Canva, Slack, Facebook Pages, Instagram professional accounts, and Meta ad accounts. Meta says it's free for most of what people need.
It isn't the only one. OpenAI's dots are always-on agents with their own computer and browser that keep working when yours is off. At the time of writing (October 2026), they're on ChatGPT's Pro levels and its top business plans, not the free plan. Google's version launched as Gemini Spark and was just renamed Tasks in Gemini, because the names keep changing and we're all just trying to keep up. It keeps working on a task while you're away, and it's on Google's paid Pro and Ultra plans, not in every country.
If you don't have a digital assistant running your inbox, you haven't missed a thing. Most people don't. This is new for everybody, the names keep changing, and we're all just trying to keep up. But the Muse announcement means a free agent that can see your Shopify orders and your ad account is now one download away for a lot of people in North America. That's worth a plan.
Here's the part that made me sit up. On October 9, Anthropic, the company behind Claude, published a report on unintended actions its own models took, mostly during internal testing. When a model hit a wall, it sometimes worked around the wall instead of stopping. In one case it submitted a practice government form to the live website. In another it accepted a data use agreement on someone's behalf. Anthropic says that, to its knowledge, none of the cases involved customer data, and it has cut all of its internal tests off from the live internet while it improves its monitoring.
I don't read that as a reason to panic. I read it as a company being straight about how these tools behave: they're persistent, and they're built to finish the job. That's great when the job is "find every invoice from August." It's less great when the job brushes up against a refund, a contract, or your ad budget.
The good news is that the companies are building in brakes. Meta's announcement says, "You're in control: nothing publishes, sends, or spends without your approval." OpenAI's dots page says a dot checks before an action affects your accounts, and either goes ahead, asks for your approval, or hands the step back to you. It also notes that stopping a task doesn't undo what's already done.
Brakes only help if you know where you want to stop. That's what the permission list is for.
Think about how you'd onboard a new assistant, a virtual assistant, or your very organized cousin who offered to help. You wouldn't hand over your bank login on day one. You'd say, "You can look at this, you can draft that, run this past me, and please never touch that." An agent deserves the same conversation, written down.
Every account you use for what you're building gets one of four levels:
Here's one rule I'd keep no matter what tool you use: anything that sends, submits, spends, or signs never goes above "with my OK." That covers email to customers, payments, refunds, ad spend, contracts, and any form with your name or your business name on it.
Here's what a list might look like for someone selling products online:
Yours will look different. A coach might let an agent read her calendar and draft session follow-ups, and put her payment links on "never." The point isn't my list. It's that you decide while you're calm, not in the middle of a launch at eleven at night.
If you do try an agent, paste the list into its instructions on day one and connect only the "can read" accounts for the first two weeks. Watch what it does. Add more access when it has earned it, the same way you would with a person.
An agent is good at the watching and sorting that eats your week in small bites:
What stays yours is anything where judgment, money, or trust is on the line. The reply to an upset customer. Whether to give a refund. Your prices. What you promise and what you won't do. I go job by job in where AI helps in your business, and the same thinking applies to people in DIY or hire help for your business.
One more thing to keep in mind. Every account you connect is one more place what you're building depends on someone else's rules. If an agent changes its plans or disappears, your permission list and your brief are still yours. That's the same idea as protecting your business from platform risk.
AI agents moved from chat windows into real accounts this fall, and at least one is free for basic use for a lot of people in North America. You don't need to rush in, and you don't need to be scared of it. You need a one-page list that says what an agent can read, what it can draft, what it can do with your OK, and what it never touches. Write it this week, keep it somewhere you own, and you'll be ready whether you try an agent next month or next year. For more practical setups, see building with AI.
An AI agent is an AI tool that can take actions for you, like sending an email, filling out a form, or making a purchase, instead of only writing text for you to copy. Most can work on a goal over time and come back to you when they need a decision.
They can be useful if you limit what they can reach. Start with read-only access, keep anything that sends, submits, spends, or signs behind your approval, and never paste passwords into a chat.
No. Most small businesses run fine on a regular AI chat tool and a few simple systems. An agent is worth trying once you know exactly which repetitive job you want off your plate.
As of October 2026, Meta's Muse for Small Business is free for basic use where Muse is available, ChatGPT's dots are on its top paid plans, and Google's Tasks in Gemini (first called Spark) is on paid plans in select countries. Check each company's own page, because access is changing quickly.
Anything that sends, submits, spends, or signs. That includes emails to customers, refunds and payments, ad budget changes, contracts, and forms in your name.
Learn AI with Belles
Start with the library, then keep building with new trainings, templates and tools, plus a monthly build and Q+A with Lily.
Keep reading
All articles
