Decomposition, judgment and verification: how an AI-built account gets run
When implementation gets cheap, value moves to the steps on either side of it. In Google Ads that means breaking a business goal into things the account can measure (decomposition), choosing which of those to act on (judgment), and testing that what got built does what was asked (verification). This article explains how Fire Pixel runs those three steps and where the AI sits.
Implementation got cheap
Anything in a Google Ads stack that can be fully described can now be built by a language model in a fraction of the time it took. The n8n workflow that reads a CRM webhook, scores the lead and posts it back to Google. The BigQuery query that joins spend to pipeline. The Apps Script that emails you when a target moves. The Cloudflare Worker that catches a call button click on the search results page. All of it is a specification and a test away.
That is a large change for an industry whose fees were mostly implementation hours. It also changes what a client should be paying for.
Decomposition
A client goal arrives as "more good leads" or "profitable growth". Neither is a number Google can bid to. Decomposition turns it into things that can be measured and fed in.
Take "more good leads" for a lead-gen business. What makes a lead good is a CRM stage: qualified, quoted, closed. How that reaches Google is an offline conversion upload keyed on the click ID. Where the missing leads are is the phone calls from the search results page that never hit the website. What each stage is worth comes from the unit economics: average order value, close rate, margin. Each of those is now a small, testable build with a clear definition of done.
The fixed-fee diagnostic we sell below £5k a month is this step, written down.
Judgment
Decomposition produces a longer list than any account needs. Judgment is deciding what to build first and what to leave alone.
Some examples of how that goes at Fire Pixel. For lead-gen sites, Microsoft Clarity plus n8n gives the same answers as server-side tagging at a tenth of the setup, so that is the default. A brand list inside Google's own settings blocks brand traffic more reliably than a negative keyword list that someone has to maintain. A CSS is worth doing for any Shopping account because the saving is mechanical. A data warehouse is worth doing once CRM data and spend both exist and nobody can see them side by side.
Judgment also means knowing what the model does not know. Every account has a constraint that is nowhere in the data: a product line the client is exiting, a region they cannot serve, a discount code that makes the margin negative. Those come from the conversation with the client, and they change what gets built.
Verification
A language model builds what you said. If what you said differs from what you meant, the build will be wrong and it will look finished. So nothing goes live untested.
Verification at Fire Pixel means firing a test conversion and following it: into the CRM, through n8n, into the warehouse, into Google with the right value attached. It means reading the SQL rather than trusting the chart. It means checking that the Apps Script fires on the schedule it claims and that its alert email arrives. Once live, it means the weekly loop: spend against target, Google's conversion count against the CRM's, search terms against the brand list, settings against what was agreed.
Google's 17 August change to bid targets is a verification problem. The target used to be a setting. It is now something to check.
When this approach works
The approach works when four things are true. The outcome is understood well enough to write down. The work happens on computers. The environment can be inspected, so a wrong build can be traced. And success can be tested. A Google Ads account meets all four. A brand launch, a rebrand, or an account whose success depends on something nobody has measured does not, and those need a different kind of help.
Related: Where the work in Google Ads moved to · The weekly loop · Standard · Data warehouse
FAQ
Is the AI making decisions on my account?
The AI builds. A person decides what to build, checks it, and owns the result. Bid decisions inside the auction are made by Google's algorithm, which is what it was before AI got involved in the build.
What gets checked every week?
Spend against budget and target, Google's conversions against your CRM, search terms against the brand list, any change to settings or targets. Anything that moved gets a written note.
How can one person offer all of this?
The build cost dropped. The stack is n8n, BigQuery, a small amount of Apps Script and Cloudflare Workers, and a language model that writes most of it. What one person adds is deciding, checking and owning the outcome.