The first week with an AI Google Ads agent has a right shape: read everything, change nothing, then make writes that cannot spend, then — only once you have watched it reason — approve changes that move money. By day 7 you will point a specialist at the account and watch it do expert work in seconds, on your command — but that trust is earned, not assumed. This playbook gives you the seven days, the prompts for each, and the checkpoint that tells you when to advance.
It happens to fit inside AdCopilot's trial too — seven days on the full Pro plan for a new workspace — but the staging is not a product limitation. Every write needs your approval from minute one. The staging is how you build a justified opinion of the agent, the same way autonomy is earned in levels rather than granted in one leap.
Day 1 — connect, then only ask questions
Add the connector URL to the client you already use — Claude, ChatGPT, Copilot, Gemini CLI, Cursor — then sign in with Google. That is the whole setup; the URL is the same for everyone, and your Google sign-in is what makes the connection yours. Then spend twenty minutes asking things you already know the answers to:
List the accounts you can see.
In account <account>, what did we spend in the last 30 days, and on which
campaigns?
Which campaign has the best cost per conversion this month, and which has
the worst?
Day 1 is a calibration exercise. You know this account; check the agent's readings against your own. When it names your top campaign correctly and its spend figures match the UI, you have established the thing every later approval rests on: it reads the account accurately.
End the day with one deliberate refusal:
Delete the worst campaign.
Watch it decline. Deletion is not gated behind approval — the capability is absent from the connector, and a REMOVED status is refused server-side. Knowing the floor exists changes how freely you experiment above it. The security page explains exactly where that floor is built.
Day 2 — run the audit
Day 2 is the structured read: wasted spend, search terms, disapprovals, structural gaps. The full half-hour script with what each answer means is your first AI Google Ads audit — run it as written. The short version, if you only have ten minutes:
In account <account>, which search terms spent money in the last 30 days
without converting? Rank by cost, top 25, with the campaign each appeared
in.
In account <account>, list disapproved or policy-limited ads, and ad groups
with no enabled RSA.
Write down the three worst findings. They are tomorrow's work — and note that the agent surfaced them in a single pass without being able to change anything. Two days in, zero approvals granted, and you already have a work list.
Day 3 — writes that cannot spend
Today the agent changes the account for the first time — through writes with no blast radius. Negative keywords are the canonical first write: pointed at the spend that buys nothing, they can only stop it, never start it.
From yesterday's wasted search terms in account <account>, add the clearly
irrelevant ones as exact match negatives to the campaigns where they
appeared. Show me the full list before the call.
Your client now shows its first approval prompt: the tool by name (the one that adds negative keywords), the account, the campaign, every term. Read it properly this once — this is the mechanism that will guard every future write, and it is worth seeing clearly on a change with zero downside. The approval loop is the whole story of why this prompt exists.
Approve it. Then check Google Ads change history in the UI: the change is there, attributed to your own Google account, because the connector acts as you — not as a service account.
Day 4 — build something paused
Second class of zero-spend write: creation. Campaigns created through the connector default to paused, and paused things cannot spend.
In account <account>, create a paused search campaign "Test - <theme>" with
a $10/day budget. Add one ad group, ten phrase match keywords for <theme>,
and draft an RSA for <URL> — show me the copy before creating the ad.
Leave everything paused.
You will grant a small chain of approvals — campaign, ad group, keywords, ad — and end with a complete, reviewable draft sitting inert in the account. Open it in the Google Ads UI and inspect it like a colleague's work. If it is good, keep it for real use; if not, tell the agent what to change and watch it revise. The full version of this workflow, brief to live, is launch a campaign with AI.
Day 5 — the first spend change
Now the agent has four days of demonstrated accurate reading and two days of clean writes. Time for the class of change that needed the trust:
In account <account>, you flagged <campaign> as budget-limited on day 2.
Show me its last-14-day cost per conversion and lost impression share, and
if the case holds, raise its daily budget by 15%. Wait for my approval.
Notice the shape — evidence, threshold, capped change, approval. That is the template for every spend change you will ever delegate: the agent argues from data it just pulled, you spend one click on judgement. If the case is weak, decline and say why; the correction teaches you both what your bar is.
Keyword-level pauses belong in the same session:
Pause any enabled keyword that spent over $25 in 30 days with zero
conversions. List them first.
A paused keyword can be re-enabled, which is what makes pausing delegable on day 5 rather than day 50.
Days 6 and 7 — turn it into a routine
The last two days convert exploration into habit. Run the 20-minute weekly routine once, end to end: weekly summary, search-term sweep, budget pacing, ad health, one improvement proposal. Time it. Then close the week with a log:
Summarise everything you changed in account <account> this week — every
write, in order, with values. I'm keeping this as my week-one record.
Cross-check that list against the approvals you remember granting and against change history in the UI. All three should agree. When they do, you have verified the property that matters more than any single optimisation: nothing happened that you did not approve.
Where you should stand by day 7
By the end of the week you should be able to answer three questions from experience, not marketing: does it read the account accurately, does it argue from evidence before proposing changes, and does every write surface for your approval first. You should have real negatives added, a reviewed draft campaign, one defensible budget change, and a routine scheduled for Monday.
What you choose to delegate next — and how much evidence you demand per approval — is now a decision you can make from a week of observed behaviour. The prompt library covers most jobs you will reach for from here.
If you have not connected an account yet, the whole playbook runs inside the free window: Start free — seven days on the full Pro plan for a new workspace, no card. Day 1 starts ten minutes after you sign in.