A working demo is the cheapest part of any AI project. You can stand one up in an afternoon now — a model, a prompt, a plausible answer on a slide. The expensive part is everything the demo skips: who acts on the output, at what cadence, with what budget authority, under whose sign-off. That gap is not a modeling problem. It is an operating-model problem, and it is where most pilots quietly die.

I built this site with AI — design-to-code, working in Claude Code, not briefing an agency. I also build the systems I ask teams to ship: a self-built MCP data-hub that serves governed commercial datasets straight into AI assistants. Doing the building myself, not by proxy, is the fastest way I know to see where a promising idea stops being deployable.

The model is the easy part

In fourteen years across pharma commercial functions, I have never seen an initiative fail because the model was not accurate enough. They fail because nobody rewired the work around the model’s output. A Next-Best-Action engine that scores every interaction on expected incremental lift is worthless if the field force still runs last quarter’s static call plan. The score has to change what someone does on Monday, or it is a very expensive dashboard.

So the real design question is not “how good is the model” but “what has to move.” Signal, cadence, budget authority, the sign-off chain. Get those wrong and the best model on the market still produces a pilot that never scales; get them right, and a modest model pays.

The score has to change what someone does on Monday, or it is a very expensive dashboard.

Measurement is the other easy thing to skip

The second place demos cut corners is proof. Attribution dashboards flatter everything — they hand credit to whatever touched the customer last. Incrementality is harder and less generous, which is exactly why it is the one worth having.

On a Next-Best-Action and marketing-mix program I led at a top-5 pharma in the CIS, we held the discipline: +7% incremental Rx, measured visited-versus-not-visited, indexed 100 → 107. Not “up and to the right.” A number with a control behind it, a scope, and a role — mine: AI lead and builder.

RECEIPT

Scope: Next-Best-Action + marketing-mix, top-5 CIS pharma
Baseline: indexed 100, visited-vs-not-visited
Role: AI lead & builder
Result: +7% incremental Rx (100 → 107)

That is the difference between an EBIT line you can defend in a budget meeting and a slide that evaporates under the first serious question.

Building it yourself is a governance instinct, not a hobby

There is a fashionable idea that senior people should stop building. I think the opposite, and not out of nostalgia. When you architect the thing yourself, you feel the failure modes before they reach a customer: the ungoverned data path, the silent fallback, the step no human ever actually reviews. Fourteen years in regulated pharma turned that into reflex — model risk, data governance, human oversight are not a compliance layer bolted on at the end, they are how the thing gets built in the first place.

That instinct is portable. An EU AI Act high-risk mindset — documentation, validation, a human in the loop — transfers cleanly to fintech, insurance, healthtech: any domain where an AI decision has a consequence someone can be held to.

What actually ships

If there is one thing this build reinforced, it is that shipping AI is a design job long before it is a modeling job. You are designing the loop the model lives inside — who receives the signal, how fast they can act, what evidence justifies the spend. The model is a component. The system is the product.

That is the work I care about, and the reason I still keep my hands on the keys.