Case studyAnonymised client
The result
Routine analysis moved to the platform; every high-risk change kept under human approval.
Building a human-supervised AI platform for Google Ads management
Atom Digital designed an advertising operations platform that uses AI to analyse account data and propose actions, while deterministic rules, permissions and human approval control what can actually change.
The outcome
Routine account analysis moved into a structured review workflow, while all higher-risk changes remained subject to explicit human approval.
Evidence
Analysis and execution, kept separate
Platform — proposed vs approved actions
The interface shows AI-proposed actions on the left and the subset a person approved on the right over one working week, so analysis and execution stay clearly separate.
What the platform does
- Reads and summarises account data
- Proposes actions for review
- Constrains them against deterministic rules
- Checks explicit permissions for what can be actioned
- Routes every high-risk change to human approval
- Logs each approved change
How it was done
The situation
The business managed a large volume of Google Ads accounts where routine analysis was time-consuming, but the changes themselves carried real commercial and compliance risk. Trusting automation blindly was not acceptable; doing everything by hand did not scale.
What was not working, and why it mattered
Analysis and execution were entangled. Any tool fast enough to help was also fast enough to make an unreviewed high-risk change. The constraint was not intelligence but control: knowing exactly what a system could and could not do on its own.
Diagnosis and decisions
Atom Digital separated analysis from execution as a first principle. AI was scoped to read, summarise and propose; a deterministic rules layer and explicit permissions decided what could be actioned, and which actions always required human approval.
What Atom Digital built
A platform where AI proposes, rules constrain and a person approves, with every high-risk change routed through review and logged. The distinction between what the system analyses and what it is permitted to execute is made obvious in the interface itself.