AI Is a Production System. Pharma Already Knows How to Run One.

Pharma keeps treating AI like a science experiment when it should treat it like a factory.

That single confusion explains most of the disappointment I see. A company runs a pilot, the demo lands, leadership applauds, but nothing scales. A year later, another pilot. I call it the Pilot Trap, and pharma is more stuck in it than almost any sector. The frustrating part is that the solution is something pharma already does best: running production systems.

A pharma plant is engineered to produce consistent output at volume to a defined standard daily. It has validated processes and quality control built into the line. It has defined ownership, maintenance schedules, throughput targets, and batch records. If something drifts, the system catches it before shipping. One good batch is not success. Success is ten thousand reliable batches while designers sleep.

Now look at how the same company runs AI. A small team builds a model that works once, in a controlled setting, with a data scientist present. That is a lab prototype, not a product on a line. Yet it is presented as if the hard part is over. The hard part has not started.

The questions change when you treat AI as production

This is the shift I push all pharma leaders to make. Stop treating AI as an experiment and start treating it as a production system to engineer, run, and maintain. The questions become better and more immediate.

Not “does the model work” but “what is its uptime.” Not “is it accurate in the demo” but “what is its yield in production, on messy data, unsupervised.” Not “who built it” but “who owns, maintains, and is accountable when it drifts.” These are manufacturing questions that plant managers ask daily, questions most AI programmes ignore.

Pharma already has the vocabulary for this: GMP for manufacturing, and a culture of standards and controls that make output reliable and repeatable, not heroic and occasional. AI needs the same instinct. A model in production needs monitoring, just as a line needs in-process controls. It requires a maintenance schedule, since models decay as the world changes. It needs defined, quality inputs, because a production system is only as good as its raw materials, and in AI, the raw material is data. Feed a line bad inputs, and you get bad batches. Feed a model fragmented, ungoverned data, and you will get the unreliable output that made pharma nervous about scaling. The caution is not wrong, just aimed at the wrong stage. The answer to unreliability is not to keep it in the lab forever, but to industrialise it.

You do not build a new factory for every product

Here is where the metaphor becomes a decision. You do not build a separate plant for every product; instead, you create shared production infrastructure and run many products through it. AI is the same. Most pharma companies do the opposite, letting a dozen brand teams each create their own versions of the same data pipeline, governance, and deployment, none of which are production-grade. That is a dozen prototype shops, not a plant. It is enormously expensive and never scales, because nothing is shared or compounded.

The effective architecture is hub-and-spoke. The hub, the plant, owns the data foundation, reusable components, monitoring, governance, deployment standards, and the hard engineering that no single team can justify alone. The spokes, brands and field organisations, run their use cases through this shared capability. The hub turns experiments into a production system. Without it, you are not industrialising AI; you are hand-crafting it, one pilot at a time, forever.

Where pharma actually is

On the maturity curve I use with clients, most pharma companies sit at stage 1 or 2 of 4. Stage one is experimentation. Stage two is coordination, with a centre of excellence and a tidy list of use cases. Both stages feel like progress, but neither produces durable value, as neither is a production system. Stage three is where value compounds, when AI is wired into core commercial and medical workflows on shared infrastructure and runs reliably enough for the business to genuinely depend on it. That is the manufacturing mindset applied to AI, a stage almost nobody in pharma has reached, despite years of spending. Meanwhile, the wider services economy, worth more than four trillion dollars, is being rebuilt on exactly this logic, and pharma is behind.

Scale is a different discipline, not a bigger pilot

I learned what it takes to industrialise something the hard way. At Hogarth, we went from nothing to more than 8,000 people across 40 countries. Scale is not just a bigger prototype; it requires a different discipline. What enables scaling is not the brilliant first execution but the boring infrastructure, standards, shared systems, and relentless focus on consistency that make the thousandth execution as good as the first. The same is true of AI. Winning companies will not be those with the most impressive pilots, but those that build the plant.

So the practical move for a pharma leader this quarter is clear. Stop counting use cases. Start counting production systems. Ask yourself, how many AI applications are running every day, owned, monitored, maintained, and delivering measurable impact. For most companies, the honest answer is close to zero. Focus on that number. It is not the length of your use-case list that is the real measure of AI maturity.

Pharma’s path forward is clear: Treat AI not as another round of experiments, but as the next production system to master. The capability is there, now it is time to apply it. If pharma brings the same operational rigour to AI as it does to manufacturing, the industry can set the standard for scalable, reliable AI use worldwide. The transformation starts when you make this mindset shift.