Your New AI Agent Came With a Rules Panel. Who's Filling It Out?

What should an AI agent be allowed to do without asking you?

An AI agent's rules panel is really a decision-rights document. Before you pick what it can do without asking, list the decisions your business makes every week and sort them. Reversible and internal, it acts. Customer-facing but editable, it drafts and asks. Money, promises, or access, it hands the decision back to you.

Here's what we'll cover:

1. Why a brand-new agent can turn your inbox into a confirmation queue
2. What the dots and Muse permission settings actually say
3. The Rules Panel Sort: four buckets for your recurring decisions
4. How Lisa fills out her panel in one sitting
5. Who in your business should write the rules

It's Tuesday, 10:12am. You turned on your new agent last week so you could finally get out of the weeds. Your phone buzzes.

"May I move the kickoff notes into the client folder?"

Sure. Allow.

"May I let the team know the notes are there?"

Allow.

"May I rename the folder from Kickoff to Kickoff Notes?"

You're on a client call. You tap Allow without looking. By the time you're in the school pickup line, it has asked to send a reminder, asked to file an invoice PDF, and asked to fix a typo in a document only you will ever read. It's the most polite crew member you've ever had. It's also playing the longest game of Mother May I in history, and guess who got cast as Mother. (Nobody mentioned that Mother doesn't get a lunch break.)

You didn't get your time back. You got a permission slip with a login.

Early users are saying it almost word for word. One wrote, "I've traded doing the work for supervising every step." That's Lisa on a Tuesday. She finally handed work to AI, and now the work waits for her in a nicer-looking queue.

The agent isn't broken. The panel is empty.

Why a new AI agent can turn your inbox into a confirmation queue

A well-built AI agent asks when it doesn't know who decides. If nobody in your business has written down which decisions are routine and which ones belong to you, the safest thing the agent can do is ask you about everything. The confirmation queue is a missing-rules problem, not a tool problem.

These agents are built to be careful, and that's a good thing. Meta says that with Muse for Small Business, "nothing publishes, sends, or spends without your approval." Careful is what you want from something that can send email in your name. But careful plus no rules means everything routes to you. A lot of owners already live that pattern with their human team. The agent just learned it faster.

A client with a team of eight had one unwritten rule for years: "Run it by me." Her team followed it. Her project manager followed it. When she set up an agent, it followed it too, because her instructions were full of "check with me first." Within a week she had the same bottleneck with a new face on it. Once we sorted her recurring decisions into buckets, most of the pings had a clear home, and the ones that still came to her were the ones that should.

Strategy First. AI Second. The tool can only follow the rules you've decided. It can't decide them for you.

What the dots and Muse settings actually say

OpenAI's dots let you set custom rules with four behaviors: Take action without asking, Take action when you say so, Ask before taking action, and Hand off to you. Meta's Muse lets you choose "Ask for some actions" or "Always ask," then answer each request with options that include Allow once and Deny.

Here's what's in each panel, straight from the vendor docs.

OpenAI dots. After setup, you open Settings, then Personalization, and choose Custom rules under Permissions. You describe an action and pick how your dot should handle it. The four choices are Take action without asking, Take action when you say so, Ask before taking action, and Hand off to you. OpenAI's own examples: Ask before taking action for sending messages to customers, and Hand off to you for deleting shared project files. You don't need a rule for every approval. OpenAI says clear instructions in your conversation come first, and custom rules are for "specific ongoing boundaries." Dots are rolling out across Pro, Business Premium, and Enterprise plans, so not every ChatGPT account has them yet. (OpenAI, Introducing dots and Control your dot)

Meta Muse for Small Business. Under Settings, then Permissions, you choose "Ask for some actions" or "Always ask." Ask for some actions means Muse asks before write actions (like sending a message) and important read actions. Always ask means it asks before any action. When Muse asks, your choices include Allow once and Deny. Muse is available in the US and Canada. (Meta, Muse for Small Business and Meta Help Center, how Muse works with your guidance and approval)

Dots gives you a rule per action. Muse gives you a dial plus a decision every time it asks. Either way, the clicking is the easy part. Knowing which button goes with which decision is the real work.

Introducing The Rules Panel Sort

The Rules Panel Sort is a one-page way to fill out any AI agent's permission settings. List your recurring decisions, ask two questions about each one, and drop it into one of four buckets that match the dots rules. Then use the same sort to answer Muse's Allow once or Deny prompts the same way every time.

One page. Four buckets. Your rules panel stops being a guessing game.

Step 1: List the decisions, not the tasks

For one week, jot down the small calls that keep landing on you. Not "email" or "scheduling." The actual decision. Can we move this client call? Do we send the reminder today? Is this invoice ready to go? Can this contractor see the shared drive?

The same handful will come back again and again. Those repeats belong on the panel.

Step 2: Ask two questions about each one

First question: if this goes a little wrong, can it be undone without anyone outside the business noticing? Second question: does it make a promise, move money, or open access to something?

Those two answers do most of the sorting for you.

Step 3: Drop each decision into a bucket

Here's the sort. Copy it, swap in your own decisions, and keep it next to the settings screen while you fill it out.

• Internal and easy to undo. Example: filing call notes into the right client folder, tidying task names, sending the team an internal reminder. Dots rule: Take action without asking. When Muse asks: Allow once, quickly.
• Routine, but you want to pull the trigger. Example: sending the weekly team recap, pulling last week's numbers into a report. Dots rule: Take action when you say so. When Muse asks: Ask for it directly, then Allow once.
• Customer-facing but editable. Example: drafting a client follow-up, drafting a reschedule note, drafting a social post or review reply. Dots rule: Ask before taking action. When Muse asks: Read the draft, edit if needed, then Allow once.
• Money, promises, or access. Example: refunds, discounts, sending a quote, agreeing to a new deadline, changing scope, sharing files outside the team, passwords and logins, deleting shared files. Dots rule: Hand off to you. When Muse asks: Deny, and do it yourself (or hand it to the person who owns that decision).

The last bucket isn't "AI can't help here." The agent can still gather the facts. It just doesn't make the call.

Step 4: Run the Friday test

At the end of the week, scroll back through every approval request. For each one, ask: did this actually need me?

If the answer keeps being no for the same kind of request, move that decision one bucket to the left. If something went out that made you wince, move it one bucket to the right. One decision category at a time.

How Lisa fills out her panel in one sitting

Lisa doesn't start with the settings screen. She starts with a list of the decisions that interrupted her last week, sorts them with two questions, and then fills out the panel to match. The settings take minutes once the decisions are clear.

Picture her Thursday afternoon. Team of eleven, new agent, pings since Monday. She pulls up the approval history and writes down what it asked about. Lots of requests. Only a handful of kinds of decisions.

Filing and folder cleanup goes in Acts on its own. The weekly team recap goes in Acts when you say so, because she likes to glance at it before Monday. Client follow-ups and reschedule notes go in Drafts and asks, and she decides her client success lead should approve those, not her. Refunds, quotes, and anything that changes a deadline go in Hands it to you.

Then she opens the panel. It takes less time than the list did. The next Tuesday, the agent still asks her things. They're just the right things now.

Who in your business should write the agent's rules?

The person who owns the decision should write the rule for it. That's not always the owner, and it's rarely the most tech-savvy person on the team. If nobody owns a decision yet, that's the first thing to fix, before any rule goes in the panel.

Here's what I'm seeing this month. Owners hand the setup to whoever is "good with tech," and that person fills in the panel with their best guess. Their guess is usually "ask Lisa," because that's the safe answer when nobody wrote down who decides.

The rules panel is a decision-rights document wearing a settings screen. If your client success lead owns reschedules, she writes that rule. If you own refunds, you write that one. If you're not sure who owns something, you've found the real work.

The good news? You already make these calls every day. You're just writing them down. You stay the brain. AI becomes the hands.

Frequently Asked Questions

What should an AI agent do without asking me?

Let an AI agent act on its own for decisions that are internal and easy to undo, like filing notes, tidying task names, or sending internal reminders. Anything a client will see should be drafted for review first. Anything that moves money, makes a promise, or opens access should come back to you or the person who owns that decision.

What's the difference between "Ask before taking action" and "Hand off to you" in ChatGPT dots?

With Ask before taking action, your dot prepares the action and waits for your approval before doing it. With Hand off to you, it asks you to take the action yourself instead. Use Ask before for customer-facing work you'd edit and approve. Use Hand off for money, promises, and access, where the decision itself should stay with a person.

How do Muse for Small Business permissions work?

In Muse, you go to Settings, then Permissions, and choose "Ask for some actions" or "Always ask." Ask for some actions has Muse ask before write actions and important read actions. Always ask has it ask before any action. When it asks, your options include Allow once and Deny. Muse is available in the US and Canada.

Who in my business should write the AI agent's rules?

The person who owns each decision should write the rule for that decision. If your client success lead owns rescheduling, she writes that rule. If you own refunds, you write that one. If nobody clearly owns a decision, decide that first. The agent can only follow rules that someone has actually decided.

Why does my AI agent keep asking me for approval?

Usually because it doesn't have clear rules about which decisions it can make. When rules are missing, a well-built agent defaults to asking. Sort your recurring decisions into buckets, set the panel to match, and check each Friday which requests didn't actually need you. Move those one bucket toward more freedom.

Map Your Decisions With Other Owners

Reading The Rules Panel Sort is step one. Filling it out next to other service-business owners, and seeing how they sorted the same decisions, is where it clicks.

Ops Yacht Club is the free community where owners map who decides what, for their team and their AI crew. Strategy First. AI Second. No gatekeeping.

→ Join the free Ops Yacht Club

Mother can take her lunch break now.

Keep reading

• What Does AI Cost You to Use? (The Review Tax Nobody Budgets) →
• "Just Looping You In" Is Quiet Approval Dependency →
• You Trained Them to Ask You Everything (That Is Why You Are Still the Bottleneck) →