How much should the AI do on its own, and when does it start spending my money?

Why this question keeps coming up

Five of the 73 sales calls in my backlog asked some version of this, and it came up again last week on a long call about moving an ad account over: "is the AI making the decisions?", "is it going to start spending my money? does it tell me?", "do I even need post approval?", "how do I know when it's good to just let it self-optimize?" One owner asked the sharpest version: "could it send the booking link, then turn itself off?"

Underneath every one of those is the same thing. These owners have paid an agency they couldn't reach, bought software that ran a workflow they didn't understand, and finished a course they never used. They are not asking whether AI can run their marketing. They are asking how to stay the owner when something runs without them in the room.

The three ways to run it

Full autopilot. The AI writes, posts, follows up, adjusts budgets and reports after the fact. Fast, and the first week it does something you'd never say or spends on an ad you'd have killed, you turn the whole thing off. Most owners never turn it back on.

Approve everything. Every caption, every text to a lead, every budget change waits for you. Safe, and it quietly dies. A lead who texts at 8pm waits until you finish your last patient. Eight of those 73 calls were about who calls leads back while the owner is treating all day. An approval queue just moves that bottleneck into an app.

Guardrails. You decide, before launch, which actions the AI may take alone, which it may take inside a limit, and which it must ask about. Then you widen the limits as it proves itself. This is how I run it in my own accounts and in clients', and it's where nearly every clinic lands once they've seen the other two fail.

The guardrail that matters most is a number. In the Campfire Benchmark, 15 cash-clinic Meta accounts from October 2025 to September 2026, cost per lead was $39.85 pooled across every account, $61.34 counting only patient leads with the lead-magnet downloads removed, and cost per booked appointment was $131.96 pooled across the three accounts that track bookings. Those are the kinds of numbers a cap is written against: "pause any ad set above this cost per lead for three days running" is a rule the AI can follow and you can check. "Don't waste money" isn't.

Where the line sits, job by job

Content. It drafts on its own, from your calls and your newsletters, and sends you a phone notification. Nothing posts until you approve. After a few months of approving drafts you trust, routine posts can go out on their own and anything mentioning price or a specific condition still waits for you.

Follow-up. This is where autonomy pays. A new lead gets a text within a minute, two qualifying questions and a booking link, at 2am if that's when they filled out the form. The AI qualifies and hands off: before any health detail comes up, the lead goes to your front desk or into the compliant CRM workflow. On a call last week that handoff was the answer to "can an AI setter work the DMs without a HIPAA problem?" The owner's words: "that's a good way to do it."

Ads. The AI reads the account every morning, pauses what's over the cap, and tells you what it did and what it saw. It never raises a budget, launches a new campaign or changes the offer on its own. Those are the three decisions that spend your money, and they stay with you for the life of the account. The daily note is the dashboard, which is why that page says the first job of any report is answering "did the ads book patients this week?"

How the AI earns more rope

The order, which is the same order I'd use with a new hire:

• Week one: it drafts and reports. You approve everything. You're learning its voice as much as it's learning yours.
• Weeks two to four: follow-up texts and routine scheduling go out on their own inside the script you approved. Posts still wait.
• Month two: routine posts publish without you. Ads self-pause under the cap. Budget, offer and new campaigns still ask.
• Ongoing: a mid-month check on AI usage and the cap, because usage rises with every role that runs daily.

And yes, it can send the booking link and turn itself off. A follow-up sequence has an end: once the lead books, or after two weeks of silence, it stops and tells you. That's a rule, not a hope, and you set it before launch.

What to do this week

Write three numbers down before anything runs: the most you'll pay per lead, the daily budget it may never exceed, and the time of day you want the report. Then list the actions that always need your yes: spend changes, new offers, anything that mentions a price or a condition. Hand that page to whoever installs the system, along with partner access, never a password. If they can't put those limits in before launch, they're asking you to trust them instead of the setup, and that's the thing you've already been burned by.