AI in performance marketing: what works and what stays a demo

The useful question: what does a wrong step cost before someone notices?

How clever a model is does little to help your decision. Commercially, what counts is what happens when it gets things wrong. That is why I sort AI in the ad account by the consequences of its mistakes rather than by its features, and three levels remain.

The levels look similar because the same model can work in all three. Their risk still differs considerably. An assistant that analyses a campaign wrongly has not yet spent a single euro. The same error with write access keeps running all weekend.

The damage at the execution level can be calculated

The formula for assessing an automation:

Damage = spend per day × days until the mistake is noticed × share of spend that earns no money

Suppose a shop spends €1,000 a day, and an automated change steers 20% of that into traffic that earns nothing. The numbers are invented, the mechanics are not.

The cheapest lever is the middle term. You do not want to cut daily spend, and you do not know the error share in advance. The time until a mistake is noticed is up to you: a fixed daily look at deviations limits the same error to a thirtieth of what a monthly report would let through. Wrong bid targets show up later because of delayed conversions, so the change history belongs in that daily look.

Budget errors have a ceiling of their own, which is less reassuring than it sounds. For most campaigns, the daily spending limit is twice the average daily budget, and the monthly limit is 30.4 times it. Both depend on the value that is set. If an agent enters €10,000 instead of €1,000 in the budget field, the limit for a single day moves to €20,000. Whether demand supports that is another question, but the brake is no longer where you thought it was.

The formula has two blind spots. It only sees money that is burnt. A wrong exclusion saves spend and costs revenue, so the calculation shows zero while you lose contribution margin every day. Steps that cut off traffic therefore need a second term: contribution margin forgone per day times days. With negative keyword lists I consider this the more treacherous mistake, because it never appears as a cost in any report. The formula also misses consequences outside the budget, such as a policy violation because of incorrect claims. There is no calculation for that, and such steps stay with a person.

It also works in the other direction: at €20 of daily spend, the same scenario adds up to €56 after 14 days. That is less than the working time an elaborate approval would cost. Very small accounts can therefore afford to automate more loosely.

Reading already saves time, and a mistake has no immediate consequence

This is the level where I see a clear benefit. Google offers a dedicated route for it: according to Google, the official MCP server for the Google Ads API is read-only and cannot change bids, pause campaigns or create assets. That connects a language model to the account without it being able to touch anything. What this looks like technically is covered in the article on the Google Ads MCP server.

How much time it saves is poorly measured. A vendor of such tools estimates that reviewing search terms shrinks from about an hour to about ten minutes, while the decision on exclusions stays with a person. I have not found independent, controlled measurements of time saved or error rates in Google Ads accounts, and anyone who gives you a precise percentage should also provide the method.

My assessment of where reading pays off: sorting search terms, finding anomalies across many campaigns, breaking down differences between two periods, drafting a report. What these tasks have in common is that the result is quick to check and collecting the information takes longer than checking it. If that ratio is reversed, the AI saves nothing.

The limit lies in what is not in the account. The same vendor article notes that context outside the account, such as a price promotion, stays invisible to the AI. This is how a harmless reading error turns into an expensive action. Suppose ROAS drops sharply in one week because the shop is offering 20% off. The analysis finds a plausible cause in the account, because that is the only place it looks. Before you act, check two things: does the figure match the interface, and does the explanation know everything that is currently happening in the business? The process for this is described under checking AI analyses of account data.

An approval is only as good as the approver's knowledge

Preparing sounds safe because a person sits in between. Google itself takes this route: according to Google, the account assistant, introduced as Ads Advisor, can suggest actions for Performance Max and apply them directly after review and approval. The API also has validate_only, a check in which the request is validated but not executed. That is a good dry run for the technical side. It says nothing about the business logic behind the change.

The weak point is the click on approve. The platform sees ROAS, you earn contribution margin. A worked example with invented numbers:

A suggestion to move money from B to A looks reasonable in every account view and, with averages unchanged, loses 45 cents on every euro moved. Why this is not an exception is explained in the article on why ROAS 8 can be worse than ROAS 4. An approval without margin data is just a click.

An approval only holds if it can answer these questions:

• What does the step do in euros of contribution margin, not just in ROAS?
• Does the suggestion know what is currently going on in the business: promotions, stock levels, price changes?
• What does the state before the change look like, and how do you get back to it?
• How will you recognise that something is going wrong, and after how many days?

On top of that there is a limit on volume. Nobody really reads a package of 300 individual changes. I would cut packages small enough that reviewing them fits the time available, and would rather have fewer suggestions. Be careful too with changes that look like maintenance. According to Google, Advisor can fix policy issues, for example by changing an ad's URL, and submits this for approval. If you wave it through without opening the new page, then in the event of a mistake you have sent every click on that ad to the wrong page.

The execution level may have been running in your account for a long time

Discussions of autonomous agents often overlook the older case. With the setting for auto-applied recommendations, suggestions are implemented regularly without individual approval. The types that can be applied include adjusting target ROAS values and adding broad match keywords. These are not side issues: both have a direct effect on what your money is spent on.

Google names one safeguard: the budget is not increased by this automation. That protects the total, not how it is used. Within the same budget, the target and keyword coverage can still shift, and exactly that share is what the damage formula contains. At least you can check what has happened: applied recommendations appear in the history and the change history, and scheduled ones in a queue where you can dismiss them in advance. Both are only useful if someone looks.

Nobody knows exactly how widespread this is. One vendor reports from its own sample that 32 out of 100 accounts had the automation switched on. That is not representative, but it is reason enough to check for yourself. If the account was set up years ago or several people have had access, the answer is not obvious.

For new automation, a simple rule does the job. A step may run without individual approval if three conditions are met at the same time:

1. It is reversible: the previous state can be restored in one step.
2. The damage according to the formula, at a realistic time until it is noticed, is below an amount you have set in advance.
3. A deviation triggers an alert that someone actually reads.

Warnings, labels and reports almost always pass this test. In my assessment, budgets, target ROAS, match types, final URLs and conversion settings do not, conversion settings if only because they also change the data the bid strategy learns from. Permissions, dry runs and limits for such cases are covered in the article on securing AI agents in the account.

What remains a demo for now

A demo shows the run that succeeded; in the account, the one that failed is what counts. In my view the agent that manages an account for profit, unsupervised, for weeks on end still belongs in this category. I know of no independent measurement that supports the opposite.

General agent tests also counsel caution, even if they do not apply directly to ad accounts. In the TheAgentCompany benchmark, a simulated small software company, the best agent tested completed 30% of the tasks autonomously. Later models may do better.

For me this puts the burden of proof on the vendor. If someone claims their tool manages your account autonomously, ask about the consequences of a mistake: what happens if it is wrong, who notices and when, and who pays for it.

What is not in the data stays with a person

The line runs between calculating and deciding. The AI can calculate, sort and draft. You have to decide:

• which goal applies: revenue, contribution margin or new customers, and at what price
• how much damage you will accept for a test
• what the business is currently planning and what is in stock
• what margin a campaign really has, if that figure is not passed on cleanly anywhere

Responsibility for the spend also stays with you, no matter who or what implemented the step.

Execution comes last

The order follows the consequences of mistakes, not the appeal of the tool.

Check the setting first. On the Recommendations page, look under the Manage and History tabs to see whether auto-apply is active and for which types. Switch off anything you do not want deliberately. This needs no new tool and is quick.

Then read only. Give a model read access or an export and ask one fixed question a week that you would otherwise answer by hand. Check the first answers against the interface. Clarify beforehand which data may go into which tool.

Then let it prepare. Once the answers are correct, the AI should deliver a short list of changes with reasons and the state before each change. Put a margin table per campaign or product group next to it. If you do not have one, that is the next task, not the next tool.

Execute last and narrowly. Only for steps that meet the three conditions above, and with a set time until mistakes are noticed. In my assessment, most of the benefit lies in the first three steps.

If you only spend a low double-digit amount per day, you can keep the preparation level leaner, and if you have no clean conversion data at all, solve that first.

Features are as of November 2026, and the tools change quickly. The calculation stays the same: whenever a vendor promises a new level, put its feature into the formula.