Outcome accountability is a luxury good. It works when you can complete an outcome cycle inside a single decision cycle. AI products break that assumption, so you need two measurement layers, on purpose. Most teams run only one.
Here's why one cadence isn't enough. AI features iterate ten to twenty times a week. The outcome cycle for those features still runs four to twelve weeks. By the time an outcome attributes back, you've shipped forty to eighty more changes. So an outcome number, on its own, is a lagging measurement that can't drive the decision you have to make today. You're steering with a signal that describes a version of the product you already replaced.
So you add a second layer next to outcomes, one that moves as fast as the product does. That's the whole move.
On the fast clock (weekly, and in practice every diff) you measure direction, with seven leading indicators:
- Eval pass rate. The share of your eval suite the current build passes.
- Agent quality. How good the outputs are against your rubric.
- Iteration count. How many times you shipped this week. Velocity is itself a signal.
- Design coherence. Whether the experience is holding together as it changes.
- Escalation rate. How often the system hands off to a human.
- Dispute rate. How often users push back on or correct the output.
- Latency. Speed, which is a quality dimension for anything interactive.
These tell you which way quality is heading before the slow outcome number can. When eval pass rate slides three weeks running, you don't wait for the quarterly outcome to confirm it.
Outcomes still matter. They just live on the slow clock, monthly and quarterly, where the cycle time actually fits. Review them the way outcome-driven planning always intended, and stop asking them to drive daily decisions they're too slow to inform.
One honest note on where this sits. It's a respectful argument with the canon, not a rejection of it. Outcome accountability was right for the products it was written about, where an outcome cycle fit inside a decision cycle. AI-native products are different. Run both layers: direction on the fast clock, outcomes on the slow one. If you only have one, add the other this week. Start with eval pass rate and escalation rate. They're the two that move first.