Automating Google Ads with an AI agent: how an online store's account runs for a month without a manual change

There are four ways to automate Google Ads. Google's own automated rules. Scripts. Software subscriptions. And since 2025: AI agents that read the account, evaluate it and make changes within limits.

That last one is the only one Dutch agencies have written almost nothing about. We have been running it since 2024 on real accounts, with real budget. This is how it works, in the order we built it ourselves.

Why rules and scripts are not enough

Automated rules are "if X, then Y". They see no context. A rule that raises budget on a good ROAS doesn't know that ROAS comes from one lucky large order.

Scripts can do more, but someone has to write and maintain them, and they still do exactly what they say.

An AI agent does something different: it reads the account the way a specialist does. It sees the search terms report, the budget split, the trend over the past weeks, and weighs them together. The risk is exactly that: it makes judgments. So you have to define what it's allowed to do.

The ladder: report, advise, execute

We never give an agent write access right away. Every account goes through three steps.

Step	What the agent may do	How long
1. Report	Read only. Look at the account every three days and post what stands out in Slack.	At least a month
2. Advise	Read + make concrete proposals: "raise the budget of campaign X by 15%, because…". A human approves.	Until the proposals are consistently right
3. Execute within limits	Make changes, but only within limits set in advance.	Ongoing, with a monthly report

Step 1 sounds dull, but it builds the most trust. For a month, the client sees the agent notice the same things a good specialist would, without anything being able to go wrong.

Step 1 in practice: the read-only check-in

Every three days the agent reads the account and sends a short message to the client's Slack channel. No dashboard, no 40 charts. Three to five sentences: what shifted, what stands out, what it would do if it were allowed to.

No write access, deliberately. This is the phase where you learn whether the agent understands the account.

Step 2: from checklist to supervised execution

For one client we had planned a restructuring of the account: split campaigns, redistribute search terms, clean up negative keywords. Done by hand, that is three to four hours of clicking.

We gave the agent access to eleven write actions in Google Ads (create campaign, change budget, add keyword, and so on) and let it work through the checklist while a specialist watched and approved every step. Forty minutes. Same result, no typos, and a log of every change.

Step 3: the autonomous month

The end stage is an online store whose Google Ads account runs month after month without anyone touching a button.

This is what that month looks like:

1. Playbook in. A document with the client's goals (target ROAS per campaign type, maximum budget, seasonal patterns), and the limits.
2. The agent pulls the data and evaluates the account against the playbook.
3. It adjusts, within the limits. Shifting budget between campaigns, tweaking bids, adding negative search terms.
4. Report out. An email with what it did, why, and what it did not do because it fell outside the limits.

We read that report. We don't run it. This account has been running like this for months, on real budget.

The limits that make it safe

Without these rules, you don't do this. They live in one file that every agent reads before it does anything.

• Budget ceiling: the total daily budget may never exceed X, no matter how well things are going.
• Maximum step: no single change larger than a fixed percentage at a time.
• Never pause winners: campaigns above target ROAS are not touched without a human.
• Escalation: on a deviation outside the bandwidth, the agent stops and reports it, instead of "fixing" it.
• Frozen baseline: before the agent starts, we record performance. Every intervention is logged, so afterwards you know whether the AI made the difference or the season did.

That last point is the most underrated. Without a baseline you never know whether the automation works or you simply had a good month.

What went wrong, and what we built from it

The honest part. Early in this journey I let an agent write ad copy without a review step. It made up a product claim that wasn't true. Not dangerous, but wrong, and it would have gone live if I hadn't looked.

That is where the rule that now sits above everything comes from: every agent gets limits and a review gate, always, no exceptions. Copy never goes live without a human. Budget changes only within limits. And everything the agent does can be read back.

Who this is for

This works for online stores that:

• already have a working Google Ads account (the agent doesn't create strategy, it executes it);
• spend enough per month to make fluctuations meaningful;
• have a point of contact who reads a short report.

It does not work for accounts that don't have conversion tracking yet. Then the agent has nothing to steer on. Which is exactly where our audit starts.

Frequently asked questions

Is automating Google Ads with an AI agent safe? Only with defined limits, a read-only starting phase and a log of every change. Without those three, it is unwise.

Does this replace a Google Ads specialist? It replaces the clicking and the daily checks. The strategy, the limits and reading the reports stay with a human.

Does this work with Performance Max? Yes, but Performance Max gives you fewer levers. There the agent mainly steers on budget allocation, target ROAS and the signals you feed it.

What do I need to get started? Working conversion tracking, a clear goal per campaign type and a Slack or email channel the report can go to. We build the rest.

---

Curious what an agent would notice in your account? We do a free audit of your current Google Ads setup and tell you honestly whether automating makes sense.