What happens when an AI system improves exactly the metric you gave it — but the business gets worse?
It sounds contradictory.
It isn’t.
Imagine asking an AI agent to reduce acquisition cost.
It finds the expensive campaigns and pauses them.
The dashboard improves.
The team celebrates.
But some of those campaigns were creating future demand rather than converting existing demand.
A few weeks later, the top of the funnel is weaker, the same warm audience is being recycled, and growth starts slowing down.
The AI did exactly what it was asked to do.
And that is the uncomfortable part: AI does not have to malfunction to cause damage. Sometimes it just has to succeed too literally.
It optimized a local metric inside a wider system it could not see.
Before giving an AI system control over a meaningful business decision, I would ask four questions:
1. What business outcome sits above this metric?
2. What delayed effects could the system miss?
3. Which actions must remain reversible?
4. What signal should make the agent stop and ask for help?
Autonomy should not begin with a stronger prompt.
It should begin with a better objective, clear boundaries, and a safe way back.
Have you seen a metric improve while the system around it became weaker?
I haven’t posted here in a long time. I’m changing that — not to become a “content creator,” but to share practical notes on building products, 0→1 decisions, AI product strategy, leadership, and mentoring. Less hype, more useful thinking from real work.
Друзья, кому шикарную работу к Leonid Chernyy - одному из самых классных руководителей - даже обладая большим опытом, вы сможете у него научиться много чему полезному. https://t.co/m3K8Q9TH8k