Writing
Notes on getting AI past the demo and into the P&L — operating-model design, causal measurement, and building the systems by hand. Proof-led, no hype.
-
Bayesian MMM: a prior is not an identification strategy
Robyn and Meridian ship with causal language, but a prior only encodes belief. What identification takes, and what an auditable MMM shows a CFO.
-
Clinic segmentation for pharma: size is not a segment
Pharma still tiers clinics by size. A census is not a clinic segmentation: segment on potential and decision structure; build the universe file first.
-
Your next-best-action is optimizing applause, not prescriptions
Agentic next-best-action engines learn from what they can see: accepted suggestions, opens, clicks. Engagement is not incrementality. Fix the reward first.
-
Governed AI clears procurement: trust is an inventory
AI governance is now a procurement gate. Buyers no longer accept an ethics charter; they want the register: inventory, data lineage, eval results, logs.
-
AI-first project management: when agents do the work
When AI agents do the work, output gets cheap and verification gets scarce. AI-first project management means redesigning the team around that shift.
-
The distance between a demo and production is an operating model
I built this site and a governed AI data-hub by hand. The gap from an AI demo to a shipped system is organizational, not technical — where pilots die.
-
Measuring AI ROI: did it move the number, or the dashboard?
Two figures on AI's return can't both be true: 74% of execs report ROI, 95% of pilots show none. Measuring AI ROI honestly means measuring the number, not the dashboard.