Test what your AI agent assumes before it acts
When an agent proposes something big, find out whether the assumption behind it holds on a small group before it reaches everyone.
See the record →Use cases · Operators running AI agents
Decide whether to let an AI agent move your ad spend
Before an agent moves your budget, test its reasoning on a small share of spend and decide on cost per paying customer.
The situation
Your marketing agent wants to move 40% of the ad budget from LinkedIn to Google because Google clicks are cheaper. Cheaper clicks don’t always mean cheaper customers. If the agent is wrong, a month of budget goes to leads that never pay.
What you find out
- Whether the agent’s proposal holds on the number that matters, not the one it optimized for.
- How far to go, and when to check again.
- Which of the agent’s proposals a person approved, and on what evidence, when someone asks.
Decision Recordworked example
- assumption
Google leads become paying customers often enough that the lower click cost wins.
- experiment
Move 15% of the budget to Google for three weeks, with every lead tagged by source.
- success criteria · locked
Cost per paying customer from Google is at or below LinkedIn’s. Locked before the budget moves.
- evidence
Google: $410 per paying customer. LinkedIn: $520.met
decision
Move forward
Move the rest of the 40% in two more steps, checking the same criteria at each step.
Marketing lead decides; agent proposes the move and tracks cost per paying customer by source.