Case study · AI agents & Google Ads

A marketing AI team for Google Ads that changes nothing without approval

Since April 2026 we have been running a team of seven specialised AI agents for a client that advertises internationally. It analyses ad accounts and revenue data, reviews test ideas, writes reports and landing page briefs, and prepares changes to the account. A change is only carried out through a technical approval gate – and only once a human confirms it. Next, the analyses will run as fixed routines that deliver proposals for approval, reports and warnings on their own.

Armin StadlerAs of: Deutsche Version
specialised AI agents
7
documented work steps
500+
specialist agent assignments
120+
approved change packages in the ad account
37

As of September 2026, counted since the project started in April. Work steps are entries in the project’s version control, assignments are tasks handed to the specialist agents, and change packages come from the approval gate’s log (July to September 2026).

Starting point

Ad accounts spanning several markets and campaign types produce more data than anyone can analyse properly on a regular basis. The typical mistakes do not come from carelessness but from shortcuts: an analysis without a cross-check, a number from the wrong source, a change nobody can trace later.

At the same time, an AI agent with write access to an ad account is a real risk. A misunderstood instruction can burn budget or damage campaigns that work. So the goal was both: far more analytical capacity – and an architecture in which the AI changes nothing without explicit approval.

What we built

A team of specialists instead of a general-purpose assistant – and a technical gate for everything that writes to the ad account.

Specialists with clear roles

An account analyst reads the Google Ads data, an attribution analyst the revenue data. An experiment reviewer assesses test ideas before launch, a challenger looks specifically for flawed reasoning. They are joined by reviewers for landing pages and infrastructure, and by a deliberately risk-tolerant opportunity scout whose proposals are labelled as such.

Reading is open, writing only through the gate

Read queries go through an allowlist of permitted interfaces. Only one place in the code can write, and it offers exactly four permitted operations. New ad groups are always created paused, and there is no deletion path. The specialist agents themselves have no write access at all.

Every change as a checked package

A change is prepared as a plan, validated against the Google Ads API in advance and bound to exactly that content by a checksum. The plan is valid for 30 minutes, can be executed only once and needs a human to confirm it in a dialog that defaults to “Cancel”. Every attempt is recorded in a log that is only ever appended to, never overwritten.

Evidence over gut feeling

Every recommendation needs evidence from the ad account or the database. Tests are defined before launch with success and stop criteria, and findings are kept as learning notes – so knowledge does not disappear with the session.

How the team works

  1. 01

    Frame the question

    The main session defines the task and hands it to the right specialists.

  2. 02

    Gather evidence

    The account and attribution analysts deliver numbers from Google Ads and the database – by query, not from memory.

  3. 03

    Challenge

    The experiment reviewer and the challenger test the conclusion for flawed reasoning before it becomes a recommendation.

  4. 04

    Decide

    The main session brings the views together. Budgets, bids and campaign status are still changed by a human, directly in the account.

  5. 05

    Apply with approval

    Permitted changes go through the gate as a package and are only executed after confirmation in the dialog.

  6. 06

    Document

    Result, evidence and decision are stored under version control. The need for new landing pages goes to the landing page system as a brief.

What the AI may do – and what it may not

The most important boundary is not phrased as a request to the AI but drawn in code: only the gate writes to the ad account, and the gate asks a human.

What the AI takes on

  • Analysing account and revenue data, with the query as evidence
  • Preparing tests, proposing success and stop criteria
  • Writing reports and landing page briefs
  • Proposing change packages and validating them in advance

What a human decides

  • Every execution of an account change – by confirming it in the dialog
  • Budgets, bids and switching campaigns on and off
  • New campaigns and any deletion
  • Approving every landing page brief

Result

The team has been in ongoing use since April 2026. More than 500 documented work steps and over 120 specialist agent assignments show how routinely it is used. Between July and September 2026, 37 change packages went through the gate – each one confirmed individually by a human.

The biggest effect is on the way of working: analyses come with evidence, tests with criteria set in advance, changes with an approval and a log entry. We have not measured hours saved – so we do not quote any.

The architecture carries over: the same approach now also runs as a second, read-only instance for another ad account.

Next stage: routines that report in

Planned

Today a human usually kicks off the team’s work. Some routines already run unattended: a cost reconciliation runs several times a day and checks itself.

The next step is planned: the team works in fixed routines on its own and reports back with results. The approval gate stays exactly as it is – even automatically generated proposals are only carried out once a human confirms them.

  • Regular analysis of accounts and running tests, without a manual start
  • Proposals for approval that a human confirms or rejects
  • Automatic reports and warnings as soon as metrics leave their expected range

What carries over to your business

  • Specialised agents with narrowly scoped tools instead of a general-purpose assistant.
  • A write gate in code – regardless of how good the agent is.
  • Mandatory evidence and test criteria set in advance.
AI & automationGoogle Ads

Frequently asked questions

Can the AI change the ad account by itself?

Only through the approval gate, and there only with four operations: create an ad group (always paused), add an ad, revise an ad, add keywords. Deleting is not provided for. A human has to confirm every execution in the dialog, and budgets, bids and campaign status are still changed by a human directly in the account.

Which AI is behind it?

The team runs in Claude Code; every role is a separate subagent with narrowly limited tools. Google Ads is connected through our own API client – with an allowlist for read queries and the gate described above for changes.

Does this replace a performance marketing team?

No. It shifts the work: less gathering and recalculating, more deciding. Responsibility for budget and strategy stays with people – that is how the system is built.

Can this be transferred to other ad accounts?

Yes. Roles, review steps and the gate do not depend on the product. What gets adapted are the permitted accounts and campaigns, the data sources and the metric the team steers by.

Is this compatible with Google Ads?

Access runs exclusively through the official Google Ads API – the route Google provides for tools and automation – not through the user interface. Google now even offers an official read-only access point for AI agents itself. New ads and keywords go through Google’s normal review, and a human confirms every execution beforehand.

Next step

Want to build something similar?

In a first call we work out whether the approach carries over to your business – and what it would take.

Book a first callAI & automation

More case studies

All case studies

More on the blog