ROI Uncertainty
The premise. Architecture decisions don't have known returns. Even capabilities that the team believes will produce positive impact can produce negative impact through mechanisms the team didn't anticipate.
The canonical example. use-case.abandoned-cart. High average ROI when delivered well, negative ROI when delivered too frequently or in distracting moments — users may remove the notification permission, clear the cart to silence the prompt, or unsubscribe entirely.
Practical implications for architecture.
- Measurement infrastructure must be a first-class architectural concern, not an afterthought. A platform that ships activations but cannot detect fatigue effects is incomplete.
- Reversibility matters more than peak performance. An architecture that makes it easy to turn off an experiment is more valuable than one that runs the experiment slightly faster.
- Sample-size honesty — "expert opinion" routinely overstates expected impact. Plan for n-of-1 surprises (source.cdp-recommendation-agent-md).
The agent's job. When recommending architectures, surface the measurement story alongside the activation story. "Can you measure whether this is working?" is at least as important as "Can you ship this in real time?"