Automating is not giving up control: it is imposing discipline on it. But automating badly is the fastest way to multiply a mistake. The Intermediate level ends with the judgement: which tool for what, how to introduce an automation without risk, what not to automate, and how to tell whether it was worth it.
Three levels of automation
| Level | What it is | What it is for | What it is not for |
|---|---|---|---|
| Automated rules (Google Ads) | "If X, then Y" over simple metrics, at a fixed frequency | Alerting and safety (pause if spend runs away) | Optimising bids or budgets (they react to noise and fight with Smart Bidding) |
| Scripts | Programs with full logic, access to everything, spreadsheets and email | Watching, cleaning, cross-referencing data, proposing, applying with rules and safeguards | Deciding strategy; making up for broken measurement |
| Generative AI (inside scripts) | Language models that read, classify and write | Semantic relevance of search terms, headlines with memory, summaries, diagnoses | Deciding about money on their own; anything with no data behind it |
The healthy hierarchy: the AI proposes, the script applies with rules and safeguards, the automated rule alerts, and you decide anything irreversible.
The four design rules of a safe automation
- Nothing irreversible without approval (a checkbox) or without a period in TEST mode first that has proven the behaviour.
- A record of every action: what, when, why, on which ID. With no record there is no audit and no rollback.
- If it fails, it touches nothing: a data, licence or server error leaves the account exactly as it was. Never "half done".
- Gradual, with waits: small changes (10-20%), with windows between changes so that Smart Bidding — and you — can see the effect.
Plus two that get forgotten: absolute limits (a maximum number of negatives per day, of budget, of pauses) as a firebreak, and business safeguards (whatever converted does not get touched; brand is protected).
How to introduce an automation
- Shadow: run it without acting, just writing down what it would do (TEST mode), for 1-2 weeks.
- Reading: compare its proposals with your own judgement. Do you agree 80% of the time? Are the disagreements adjustable thresholds?
- Calibration: tune the thresholds until the proposals are few and clear.
- Manual approval: switch it on with checkboxes; approve every action yourself for a few weeks.
- Partial autonomy: let it apply the low-risk things on its own (negatives for irrelevance, alerts) and keep the checkbox on everything else (pauses, budgets, ads).
- Monthly review: the action history, false positives, results.
An automation introduced like that never gives you a fright. One introduced "switched on from day one" always does.
What not to automate (yet, or ever)
- Strategic decisions: what to sell, at what price, in which market.
- Bidding strategy changes in one go.
- Measurement: you automate the monitoring of it, not the definition of what counts as a conversion.
- Replies to customers without human review (AI as a draft, not as the sender).
- Any action whose cost of error you cannot undo within a day.
Measuring whether the automation is paying off
Automation is judged like any other change: before/after over complete weeks, or better still, an experiment (campaigns with and without). Metrics: CPA/ROAS against the business goal, management time saved (hours a month), mistakes avoided (alerts that arrived in time), and false positives (actions rolled back). An automation that saves 10 hours a month and does not make the CPA worse has already paid for itself; one that improves the CPA by 15% is an employee who never sleeps.
The final rule
Everything deterministic — anything you can write down as a rule with data behind it — goes to a script. The unpredictable — synthesis across accounts, decisions with business context, exceptions — stays with the person (or with a supervised agent that prepares the decision). Knowing where that line sits in your account is, in the end, what an account manager's job is in 2026.
💡 Ninja trick: this module is also how the Suite itself is built: TEST mode on everything that moves money, checkboxes for anything irreversible, history by ID, "if it fails, it touches nothing", daily limits, brand and conversion safeguards, gradual changes. And the Ninja Scripts Agent is the next step up: it watches all of the above across every account, prepares the diagnosis with AI… and leaves the decision to you. The Advanced level teaches you how to use it.
What you should remember
- Rules alert; scripts apply with safeguards; AI proposes; you decide anything irreversible.
- Four rules: checkbox/test, a record, if it fails it touches nothing, gradual (+ limits and safeguards).
- Introduce it as shadow → reading → calibration → checkbox → partial autonomy → review.
- Do not automate strategy, measurement, replies to customers, or anything you cannot undo within a day.
- Measure the automation like any other change: before/after or an experiment.
That is the end of the Intermediate level. You have covered advanced Search, Smart Bidding, Quality Score, ads, Shopping and feeds, Display, Measurement II, budgets and scripts. The Advanced level goes into the black boxes — Performance Max, Demand Gen, AI Max — into ad fraud, into lead quality with real data, into GAQL and into the architectures you need in order to scale.