Human-in-the-loop AI for Google Ads: a change-management model that keeps operators accountable
Large language models are very good at two things in a paid media account: summarising what happened, and drafting what to do next. They are not good at owning the outcome of a change. That gap — between suggesting a mutation and being accountable for it — is where most "AI PPC" products quietly break down.
The honest version of AI in ad accounts is not autonomy. It is a faster, better-documented version of the workflow a senior account manager already trusts: propose, review, approve, log. This post lays out the change-management model we built Adsynth around, and why each part matters if you manage spend for clients or a board.
The problem with "autonomous" optimisation
A demo where an agent rewrites campaigns unattended looks impressive for ninety seconds. In a real account it creates three problems:
- No accountability. When a bid change tanks ROAS on a Friday, "the AI did it" is not an answer a client or a CFO accepts.
- No policy awareness. Models do not know your contractual constraints, brand-safety rules, or the campaign your client explicitly asked you never to touch.
- No reversibility. Unlogged changes are hard to find and harder to undo. The blast radius of a mistake grows with the model's confidence.
The fix is not a smarter model. It is a workflow that puts a human decision between the suggestion and the API call.
The four-step model
We treat every mutating action as a small, reviewable proposal rather than an instant change.
| Step | What happens | Who owns it |
|---|---|---|
| 1. Propose | The agent drafts the exact change (entity, field, old → new value) with a stated hypothesis | Agent |
| 2. Review | The operator sees the diff, the reasoning, and the expected impact in plain language | Human |
| 3. Approve | The operator approves, edits, or rejects. Nothing hits the Google Ads API before this | Human |
| 4. Log | The approved change is written to an audit trail with who, what, when, and why | System |
The key property: no change ships without a logged human decision. That single rule is the difference between a tool built for agencies and one built for demos. If you are evaluating AI PPC software, ask the vendor directly whether changes can be pushed without an approval — the answer tells you who the product is really for.
Label Google's recommendations as Google's
A subtle but important detail: when you surface optimisation ideas that come from Google's own RecommendationService, label them clearly as Google Ads Recommendations — distinct from your AI assistant's narrative. End users (and compliance reviewers) should always be able to tell which suggestions originate from Google's API objects and which are third-party analysis. It protects trust, and it keeps you aligned with Google Ads API programme expectations.
What a good audit log actually contains
"We have logging" is not enough. A useful audit trail answers an investigation in under a minute:
- The actor (which user approved, not just "the system").
- The before and after values, not a vague label like "optimised bids".
- The hypothesis — why the change was proposed — so you can evaluate the decision later, not just the outcome.
- A timestamp and account so weekly reviews can correlate changes with performance shifts.
When you log the reasoning, your weekly review stops being a post-mortem of numbers and becomes a feedback loop on decisions. That is the foundation for the review rhythm we describe in running weekly PPC reviews around decisions, not dashboards.
Where this leaves AI
None of this makes AI less useful — it makes it trustworthy. The model still does the heavy lifting: pulling the data, spotting the anomaly, drafting the negative-keyword list, writing the RSA variations. The human spends their attention on the handful of decisions that carry risk, with full context, instead of clicking through dashboards. That is a better job for both parties.
FAQ
Does human-in-the-loop slow teams down?
For low-risk, high-volume work (negatives, obvious budget pacing) approvals are near-instant. The model is about putting human attention where the risk is, not adding friction everywhere.
Can I approve changes in bulk?
Yes — batching low-risk proposals is fine. The point is that a person makes the call and it is logged, not that every change needs a separate meeting.
How is this different from Google's auto-apply recommendations?
Auto-apply hands control to Google's heuristics. An approval-gated model keeps the decision — and the accountability — with your team, while still letting you adopt Google's recommendations when they make sense.
Related reading
- Run weekly PPC reviews around decisions, not dashboards
- Cross-channel budget allocation: when a bandit beats your spreadsheet
Adsynth queues every mutating Google Ads change for explicit human approval and logs the decision. See how the platform works or start a free trial.
