Most AI ad tools hand you recommendations and leave you to do the clicking. The useful question is not how clever the advice sounds. It is how much control you can safely hand over.

I did that clicking for years for small ecommerce brands and home services accounts, mining search terms at 1am while the client paid for the waste in real time. Recommendations were rarely the bottleneck. The delay between spotting a problem and changing the account was.

Reading an account and changing it are different permissions. Everything in this piece follows from that distinction.

Through groas, the ChatGPT, Claude, Cursor or Gemini window you already use can work at three levels:

  1. Ask: read the account and explain what is happening.
  2. Approve: propose a change and wait for your yes.
  3. Autopilot: act around the clock within rules you set.

I will work through those levels with one running example: an account spending $20k a month across Google and Meta, with waste hiding in search terms on one side and tired creative on the other. The aim is not to reach autopilot as quickly as possible. It is to know what has to be true before handing over the next permission.

Level 1: Ask where the money went, without changing anything

At Level 1, the AI reads the account and touches nothing. Through groas, it can examine campaign spend and conversions, Google search terms, Meta ad sets and creative, and the connected accounts in Microsoft and ChatGPT Ads.

It also gets context beyond account reports: competitor ads on Google and Meta, the landing pages and funnels behind them, customer conversations on Reddit, and SEO data from Semrush and Ahrefs. Your numbers show what happened inside the account. The surrounding data helps the AI investigate what might explain it.

For the $20k account, I would start here:

Through groas, summarize the last 14 days of Google Ads spend by campaign and flag anything with spend but no conversions. Do not make changes.

Then I would ask for the same cut on Meta, broken down by ad set and creative.

That produces a list to investigate, not a list to pause. Spend without recorded conversions might mean waste. It might also mean the conversion signal needs attention. Level 1 gives me room to separate those explanations before either one becomes an instruction.

I start here because read-only work lets me check the AI’s understanding without letting a misunderstanding alter the account. Does spend match the platform? Is it reading the right conversion? Can it explain which campaigns it included?

If it cannot name the waste accurately, it should not touch the budget.

Level 2: Approve a specific change, not a vague intention

Level 2 crosses the important boundary: the AI can now change something. From the same chat, it can adjust budgets and bids, add negative keywords, pause or enable ads, ad sets or keywords, and draft new ad copy for review.

The difference is execution. An answer that says “consider adding negative keywords” still leaves me with the work. A proposed list that I can inspect and approve closes that gap.

For the Google side of the example, I would use:

Draft the five negative keywords you would add from last month’s search terms. Show the match type and where each would apply. Wait for my OK before applying them.

The preview matters as much as the suggestion. “Remove irrelevant traffic” sounds sensible. Five exact-match negatives applied to a named campaign give me something I can actually judge.

On Meta, I would use the same sequence: identify the tired creative, inspect the evidence, then approve a specific pause. I would not bundle that decision with a budget increase and new copy just because the chat can handle all three. A small change is easier to inspect and easier to understand afterward.

This is also where I separate permission from trust. Giving an AI write access does not mean every proposed write deserves approval. I still need to see:

  • What will change.
  • Where it will change.
  • Why the evidence supports it.
  • What the current setting is.

My Level 2 rule: no blind writes. Preview the change, then say yes.

Level 3: Give autopilot one bounded job

Autopilot removes the individual approval step, not the boundaries. I set the condition, the action and the scope: one campaign, one account or one site. Then groas watches around the clock and acts within those limits.

For the Meta side of the $20k account, a draft rule might read:

Pause any Meta ad with CPA 50% above the ad set average after 3,000 impressions.

That is a candidate rule, not a universal recommendation. Before enabling it, I still need to judge whether the tracking and volume make that comparison meaningful. A precise threshold can still encode a bad decision. It just does so neatly.

The benefit is consistent coverage. Once the rule is sound, it does not need me to remember to check it on Friday. I am asleep; the rule is not.

The underlying pattern is familiar. Meta already supports automated rules with conditions and actions. With groas, I write the rule in chat with my AI and set where it applies across the connected platforms. Every change is logged with its reason. Autopilot stays off until I switch it on.

The permission is narrow by design: when this condition is met, take this action here. “Improve my account” is an ambition, not a rule.

For the full Google Ads setup, there is a separate walkthrough: PPC on Autopilot: The First-Time Setup Runbook for Google Ads That Manage Themselves.

Start with one pause rule on one problem campaign. Widen the scope only after checking what it does.

Logs and undo make control inspectable, not consequence-free

Through groas, every change lands in a log with what moved, what it was before and why it changed. Any change can be undone in one step.

That makes the difference between “the account looks different” and an explanation I can inspect. I can follow the decision from evidence to action instead of guessing why spend moved. I check logs the way I used to check change history on a shared account: newest first, reasons second, numbers third.

For our running example, the review is straightforward:

  • Did the approved negatives land in the intended campaign?
  • Did the Meta pause rule act on the intended ad?
  • Does the recorded reason match the rule I authorized?
  • Does anything need reversing?

This is not paperwork after the useful work. It is how I decide whether the next permission is earned.

I also keep approval and pacing separate. I can approve a sensible edit and still make too many changes at once to understand the result. Starting read-only and then enabling writes deliberately is the same basic pattern described in Supermetrics’ guidance on AI account permissions.

Undo reverses the setting. It does not recover money already spent. A log helps me catch a mistake; it does not make checking optional.

Move up when the decisions become predictable

I do not promote an account because a calendar reminder says it is time. I move up when the level below stops producing surprises.

At Level 1, I might spend two weeks comparing the AI’s readouts with the ad platforms. Spend should tie out. The conversion definition should be clear. I should understand why each item appears on the waste list.

That does not prove every recommendation is right. It establishes that the AI is reading the account I think it is reading. Only then do I test contained writes.

Clay figures of an advertiser approving one AI-suggested keyword change while nine others wait in a queue

At Level 2, I want approvals to become boring before I remove them. Not automatic yeses. Predictable proposals whose evidence, scope and effect I understand.

My checkpoints look like this:

  • Ask to Approve: tracking records conversions, I have 30 days of search-term and creative results to judge, and I approve five small changes without a surprise.
  • Approve to Autopilot: I approve the same fix three times in a row, can express it as a measurable condition and action, and know the undo path.
  • Stay with approvals: budgets shift weekly, the offer changes frequently, or nobody owns the log when a rule fires.

Those are working checkpoints, not a certification process. Three approvals do not rescue a poorly written rule. They give me repeated chances to find out whether the decision is genuinely repeatable.

Automate the decision I keep making, not the one I still need to debate.

Bad signals make faster decisions worse

No level of access fixes a broken conversion feed. If the account undercounts sales, a pause rule based on reported returns can stop ads that are actually profitable. The rule has obeyed. The input has failed.

That distinction matters especially when I compare the two sides of our $20k account. A Google search-term problem and a Meta reporting problem can both appear as spend without conversions. They do not deserve the same response.

Before allowing automatic pauses, I want tracking cleaned up, minimum-data conditions in the rule and no pause decisions while an ad set is still in learning. Where server-side tracking needs attention, that work comes before delegating more control.

Thin data creates a related problem. Give autopilot a campaign with four conversions a month and it can end up chasing noise: pausing apparent losers on Monday and favoring apparent winners on Thursday. A threshold does not become reliable merely because it contains a number.

Weak offers are another limit. A landing page converting at 0.4% does not become persuasive because negative keywords get added faster. That may reduce waste, but it does not fix the page.

If tracking is incomplete, volume is thin or the offer needs work, stay at Ask and fix that first. AI hands save labor. They do not create demand.

Connect once, then earn each permission

The setup starts by connecting ad accounts and the website in groas, then adding groas to ChatGPT, Claude, Cursor or Gemini over MCP.

groas Unlimited is $19 a month flat for unlimited teammates, ad accounts and websites, with every feature and all the data included. The connection gives the AI both account data and the market data groas collects, rather than leaving it to answer from general training knowledge alone.

I would put the $20k account through this sequence:

  1. Read: run the spend-with-no-conversions prompts on Google and Meta, then verify the results.
  2. Approve: choose one repeated, contained fix and inspect it before applying it.
  3. Review: check the log, including the previous setting and the reason for the change.
  4. Delegate: turn that fix into a bounded rule only after the repeated approvals make sense.

For Claude users, Connect Google Ads to Claude: Setup and the First 5 Prompts covers the connection and opening prompts.

There is no prize for completing all four steps in two weeks. The next level is earned by reliable decisions, not time served.

A new case: the dental group with missing call data

Now take a different account: a dental group spending $8k a month, half on Google search and half on Meta, with one front desk logging calls in a spreadsheet.

At Level 1, I would ask for campaign spend and recorded conversions, then compare those results with booked consultations. I would also ask for the competitor angles groas collects. But the first decision is not which ad to pause. It is whether the ad platforms receive the call outcomes at all.

If those calls never reach the platforms, plenty of spend can appear to produce no conversions. That readout is not permission to start cutting. It tells me to fix the signal before granting write access.

Once call tracking is working and the readouts make sense, I can consider a contained change: search-term negatives on Google, or a creative pause on Meta where the evidence supports it. Each gets a preview and an approval.

Autopilot waits until the same bounded decision earns repeated yeses and the log shows the changes I expected. The model holds even though the business and budget have changed: verify the read, inspect the write, then delegate the repeatable decision.

Recommendations were never the product here. The change, logged and undoable, is.