A recommendation cannot spend your money. An AI with permission to change your ad account can. That is the dividing line I care about when someone sells me “AI ad optimisation”: not how clever the suggestion sounds, but what the system can touch, what stops it, and what record it leaves.
For years, much of optimisation meant a Monday report: wasted clicks, new negatives, a bid cut here, a budget shift there. Then came the reply: “Looks good, go ahead.” The analysis and the action were separate jobs. A faster report only improved the first one.
Once your assistant can make the change itself through a connection like groas, the second job changes too. Optimisation becomes a question of permissions, limits and an audit trail. I’ll build that model around an illustrative account spending $20k a month with a $40 target CPA. The numbers give us something concrete to inspect instead of another promise about “AI-powered growth.”
Three different jobs get sold as optimisation
Before I judge a tool, I separate three jobs that vendors routinely bundle under the same label:
- Platform automation. Google or Meta handles decisions inside the advertising platform. Smart Bidding, for example, uses conversion history to set auction bids against your target. The platform is already acting, but within its particular job.
- Recommendations. A tool reads the account and proposes changes. A keyword costs $68 per conversion against our $40 target. A search term has spent $412 without a sale. The tool flags them, but nothing changes until someone applies the recommendation.
- Actions. The assistant reads the account, proposes or selects an edit within its permissions, and makes it. The negative gets added. The manual bid changes. The budget moves.
The same observation can sit behind all three. What differs is who makes the decision and who executes it.
A recommendation is not an account change. That sounds painfully obvious until you read a tool comparison that gives both jobs the same tick in the “optimisation” column. Before paying for intelligence, I want to know whether I am buying analysis, execution or both.
First establish whether the assistant can write
A connection to an ad account does not automatically mean permission to edit it. Google’s official Google Ads MCP server, for example, can list accounts and run GAQL reports but cannot modify bids, pause campaigns or create assets. It can help an assistant inspect the account without giving it the ability to change it.
That distinction matters more than the quality of the chat interface. Beautiful recommendations still leave someone doing the edits.
I separate access into three practical states:
- Read-only: inspect the account and explain what is happening.
- Approval required: propose a specific edit, then wait for permission to apply it.
- Autopilot: act without approval for each individual edit, but only within the rules I set.
These are different levels of authority, not different levels of intelligence. An assistant can make a useful observation in read-only mode. It can also make a bad edit very efficiently with write access.
My starting point is simple: prove the analysis before widening the permission. “Connected” is not a sufficient answer to “What can this thing change?”
Turn the $40 CPA target into a controlled instruction
Through groas, I can connect Google Ads, Microsoft Ads, Meta Ads and the website, then add groas to ChatGPT, Claude or another agent. Plain English becomes the interface for working across those connections.
For our $20k account, I would start with this:
Through groas, optimise this campaign for CPA under $40: propose changes, wait for approval, then apply.
The important words are not “optimise” or even “$40.” They are “wait for approval.” The target describes what I want. The permission describes what the assistant may do about it.
That instruction also separates two questions I want answered independently:
- Does the proposed change make sense?
- Did the assistant apply exactly the change I approved?
A persuasive explanation answers neither on its own. I need to inspect the proposal, then check the actual edit. Only after that do I have a reason to consider more autonomy.
The $40 target is an objective, not a promise. Giving an assistant permission to act does not give it permission to guarantee the result.
Define the edits before handing over the account
“Manage my ads” is too broad for a useful first instruction. I would rather start with a named campaign and a narrow class of changes. Here is what execution can mean in our running example:
- Search terms and negatives. Inspect the query report, flag the term with $412 of spend and zero sales, and propose an exact negative. After approval, add it. This is the daily grind I used to do at 1am, not a new species of strategy.
- Bids and targets. Propose lowering a manual bid from $3.20 to $2.40 where CPA sits at $68, or easing a target CPA from $32 to $38. These are different controls; the assistant needs to identify which one it intends to change.
- Budgets. Propose moving $50 a day from a prospecting campaign at $92 CPA to a brand campaign at $21. That difference deserves inspection, not an automatic transfer just because one number is smaller.
- Ads and assets. Flag a responsive search ad at 1.8% CTR while the other two hold 4.2%, then propose a pause or a headline test. A flag starts the review. It does not settle it.
- Audiences and structure. Propose excluding past purchasers from Meta acquisition, splitting an ad set or pausing a placement. I keep these changes on approval while I inspect the reasoning.
The same propose-then-apply habit appears in Hand Meta Ads Clicking to Your AI.
Start with a defined edit, not a vague mandate. “Add this approved negative to this campaign” gives me something I can verify. “Make performance better” gives me something I can argue about later.

Write rules that constrain actions, not just ambitions
Once the assistant can execute an approved edit, the next question is whether it may execute similar edits without asking every time.
Through groas, autopilot stays off until I switch it on, and it operates under rules I set. That makes the rules part of the account setup, not a footnote to it.
For our illustrative account, I would begin with constraints like these:
- Never raise a daily budget more than 20% in one day, and only propose a rise when CPA is under $40 for the prior 7 days.
- Never pause a campaign, ad set or keyword with a conversion in the last 7 days.
- Never lower a target CPA by more than 15% at once, and leave 14 days between target changes.
- Never add more than 10 negatives at once without approval, and keep brand terms untouched.
- Anything above $100 of new daily spend waits for a yes in plain English.
These are starting rules for this example, not universal settings. Their purpose is to make the boundary explicit: what may change, by how much, over what period, and when the assistant must stop and ask.
A rule written in a prompt is not, by itself, proof of enforcement. I want to know where the limit is checked and what happens when a proposed edit breaks it.
One implementation describes bid ceilings, explicit update masks and a preflight health check, including rejection of a malformed change without applying it. That is the distinction worth carrying forward. A system should not merely explain the boundary after crossing it.
The practical test: give the limit a precise shape, then check that an out-of-bounds proposal stops before execution.
Make the audit trail part of the product
A permission tells me what may happen. A log tells me what did happen. I need both.
Through groas, every change is logged with its reason, and any change can be undone in one step. That is central to the model, not a housekeeping feature added after the clever bit.
For our manual-bid example, I want a record that lets me inspect:
- What changed: the named keyword’s bid.
- Before and after: $3.20 to $2.40.
- Why: CPA was $68 against a $40 target over the period reviewed.
- Authority: my approval, or the rule permitting the action.
- Reversal: the change I can undo.
Those details describe the evidence I look for when reviewing execution. “Campaign optimised” is not evidence. It is a label.
I also want write actions easy to find, rather than buried among reports and account reads. The broader security guidance on approvals and tool-call logging makes the same distinction between reading data and changing it.
One limit matters here: undoing an account edit does not turn back the clock on spend. Reversibility helps recover control; it does not replace permission limits. I want the guardrail before the action and the receipt after it.
Judge the whole workflow with four questions
At this point, the buying test is more useful than a feature count. I use four questions, in order.
1. What can it read?
Through groas, the assistant can work with Google Ads, Microsoft Ads, Meta Ads and the site, alongside competitor ads and funnels, Reddit conversations and how AI answers describe the business.
Read scope determines what the assistant can consider. I want to know whether a recommendation comes from one campaign report or a wider view of the account and website. More context is useful only if the proposal shows how it matters.
2. What can it actually change?
Recommendations do not count as execution. Optmyzr’s Rule Engine uses IF/THEN rules to change bids, pause keywords and control budgets on schedule. That acts, but it still leaves me writing and maintaining the rules.
For this workflow, I prefer groas: my assistant can propose the edit in plain English, wait for my approval and apply it through the same connection. I am not looking for another place to maintain the clicking instructions by hand.
3. Who authorises the change?
Approval and autopilot must be distinguishable. I want read-only access while evaluating the analysis, explicit approval for the first edits, and autonomy only for the actions I have deliberately put inside a rule.
A target CPA does not answer the permissions question. Neither does a general instruction to “improve performance.”
4. What proof remains?
I want the edit, its reason and a route back. If I cannot inspect what happened, I cannot sensibly decide whether to widen access.
These questions also explain why Google handling auction bids still leaves work for an outside tool. Auction bidding, account-level decisions and the authority to execute those decisions are not the same job.
Respect the limits that permissions cannot fix
The model has a boundary of its own: better execution does not manufacture better evidence.
Platform automation needs conversion data, and repeated changes can disrupt learning. Meta, for example, typically uses about 50 optimisation events in a week as its learning benchmark. That is a reason to inspect data volume and editing frequency, not a reason to invent a universal conversion threshold for every account.
Conversion lag adds another constraint. A B2B form submitted today may not close for 60 days. The assistant can apply an edit immediately; it cannot make the eventual sale arrive immediately so the report looks tidy.
That is why a 30-day guarantee deserves scrutiny. The dashboard can show activity long before it shows the business outcome that matters.
I judge an optimisation tool partly by what it refuses to promise. Does it account for delayed conversions? Does it warn about disruptive changes? Does it explain when the evidence is thin? If it promises a $40 CPA in 30 days on 12 conversions a month, I walk away.
A fast click is useful. A fast conclusion from insufficient data is not.
Apply the model to an $8k home services account
Now change the case. A home services business spends $8k a month, records 22 calls in 30 days and wants booked jobs under $110.
The first thing I notice is not the budget. It is the measurement: a call is not the same outcome as a booked job. The account’s call count does not tell me whether it meets the owner’s target. I would not hand over full autopilot on that basis.
For the first week, I would keep groas read-only, run the propose-first instruction on one campaign and inspect the reasoning. Suppose it flags $640 across three queries with zero calls and proposes exact negatives. That gives me a specific claim and a specific edit to review.
If it also flags a keyword at $68 per call, I would not call that a failure against a $110 booked-job target. Those numbers measure different things. I need the booked-job outcome before treating the comparison as a reason to cut the bid.
Next comes a narrow execution test: one campaign, approval required, log checked daily. I approve only the edits whose reasoning I can follow, then verify that the applied changes match.
After five clean approvals, I would review whether to widen permission. Five is a checkpoint, not a certificate of competence. I would add the relevant caps, consider autopilot for a narrow search-term task, and leave budgets, ads and structure on propose-first until their decisions are equally clear.
The model now works without borrowing certainty from the larger account. Platform automation handles auction bids. My rules define the assistant’s authority. The assistant handles the approved clicking, and the log keeps the receipt.
Optimisation is no longer just the list. It is what the AI may touch, the limits it works inside and the record that lets me undo the change.

