Safe AI ad management: preview, approval and guardrails

Giving an AI write access to an ad account sounds risky, and the concern is justified. A single sentence the model misreads can multiply a budget tenfold or pause the winning campaign. The answer is not withholding access but designing how it is used.

Tecrube builds that design in five layers. None is optional; they are active whether you work through the MCP connector in Claude or ChatGPT or from the /app panel.

Before a write request reaches the account it becomes a preview showing old value, new value and the objects affected: which campaign, which budget, which keyword. Add a hundred negatives and you see all hundred in a table.

The preview is where you check whether the model understood you. Saying “raise the budget twenty percent” is where you notice whether you meant daily or monthly.

After the preview an explicit approval is required. An unapproved request waits in the dashboard and nothing is written. The model cannot proceed “assuming you approved”; approval is an action only you can take.

On the Agency plan approval rights can be delegated to the client. The account lead prepares the preview, the client approves in the panel, and the “why did you change that?” question is answered in advance.

Guardrails define changes that will not be applied even if you approve them: a single budget increase above 50% is refused, a total increase above double within 24 hours is refused, and the spend guard enforces daily and monthly ceilings; removing campaigns is not in the toolset at all.

With the fifty percent cap, for example, “triple the budget” shows a closed guardrail in the preview and is refused even if approved; the message cites the limit and suggests a smaller step. Spend ceilings are set per account in the /app panel.

Every applied change is logged with timestamp, user, client (Claude, ChatGPT, Claude Code, Cursor or the panel) and old-to-new value. When performance shifts unexpectedly, this is the first place to look.

The log merges Tecrube’s own actions with Google Ads change history, so it shows all movement in the account. The internal “I did not touch it” debate ends.

The connection is established through Google OAuth; Tecrube never sees or stores your password. The permission scope is limited to the ad account; operations such as payment methods or account ownership are not defined in the connector and therefore cannot be requested.

Revoke access in your Google account and the connector stops instantly. On Enterprise, team roles define separately who can prepare previews and who can approve.

These five layers do not slow the work; reading and approving a preview takes under a minute. What slows down is the speed of mistakes. Try the flow on a small change like a search term cleanup with the 75 free credits, tune the guardrails to your account, then make it routine on Pro ($49) or Agency ($99).

Start small when you set guardrails. In the first month put a narrow cap such as twenty percent on daily budget changes, mark your top revenue campaign as untouchable and keep removals switched off entirely. As you come to trust the flow and the quality of suggestions, widen the limits to fit your account’s rhythm.

If you work as a team, separate roles from day one: the person who prepares a preview should not be the one who approves it. That simple split noticeably lowers the chance of an overlooked mistake reaching the live account and makes every later change easy to account for. Reviewing the log together once a week also builds a shared language in the team.

What if I forget to approve? Nothing happens; the request stays in the pending approvals list and the account stays unchanged. Pending previews are listed in the /app panel and you can restart the same request later.

Who can change the guardrails? The account owner by default. On Enterprise, permission to change guardrail settings can be a separate role, so a team member preparing previews cannot loosen the limits alone.

How long is the log kept? Records stay available in the panel for as long as the account is active and can be exported. Agencies can attach the log to monthly client reports.

What if the model calls a tool with a wrong parameter? That is exactly what the preview is for. If the parameter breaks a guardrail, the request is rejected before a preview; if not, it appears in the preview and you can correct it before approving.