From Demo to Daily Practice: Four Lessons on Scaling AI in Drug Discovery

Novartis’s Christian Diehl offers one of the clearest accounts yet of what it takes to move AI from the showcase into the daily work of science.

I recently came across a conversation between Christian Diehl, chief data and digital officer of Biomedical Research at Novartis, and McKinsey’s Navraj Nagra and Valentina Sartori on the firm’s Eureka! podcast. It is one of the most grounded discussions of enterprise AI in drug discovery I have heard this year. Diehl speaks from nearly a decade of building data42, the Novartis platform that brings clinical trial data together with real-world data, and he is candid about the pivots along the way.

Four of his lessons map closely to what we see across precision science organizations, and to the choices we made in building L7|ESP^®, the agentic operating system for precision science.

Cool does not scale

Diehl’s most memorable line is also his most practical: “cool does not scale.” He describes watching impressive demos in steering committee meetings, then checking back six months later to find they were never industrialized. His prescription is to make AI boring: embedded so deeply in the daily process that it becomes simply the way work gets done.

I agree, and I would add that boring is an architectural outcome. An AI capability becomes routine when it lives inside the workflow a scientist already runs, with the same samples, methods, instruments, approvals, and records. When AI sits beside the work in a separate tool, every use requires someone to move data out, interpret the result, and carry it back. That friction is what keeps promising pilots stuck at the demo stage.

Diehl also argues for discipline. Build AI around core business processes, such as finding the right hit and optimizing a molecule, and resist spreading effort across use cases that never scale. That is the right unit of design. The process comes first, and AI is applied to make it better.

Having data and having AI-ready data are two different things

Diehl draws a sharp line between holding large amounts of data and holding data that advanced models can actually use. Getting there, he explains, requires a metadata layer, context, and harmonization so that data sets can be compared. Novartis has sustained that investment for years, and he credits that stamina for its ability to use new AI capabilities as they emerge.

He shares a lesson I wish every data program had heard early. His team curated data sets it expected to be valuable and found few takers, while scientists kept asking for data that had not been curated yet. The fix was to stay close to the scientists, start from their questions, and work backward to the data product needed to answer them.

I would take the idea one step further. AI-ready data is the baseline. Agents need data that is AI-actionable: harmonized, and also connected to the process that produced it, the lineage behind each result, and the permissions that govern who, or what, can act on it. That context is easiest to capture during execution and hardest to reconstruct afterward. It is why we built L7|ESP around a shared ontology where scientific data, workflows, instruments, and people connect as the work happens.

Diehl also notes that, starting today, mature platform technology would save a great deal of time. For organizations earlier on this path, that is encouraging. The ten-year build Novartis undertook reflected the tools available when it began. Today the foundation can be put in place much faster, so more of the effort goes toward the science.

Govern agents like associates

The most forward-looking part of the conversation concerns agentic AI. Diehl suggests agents should be governed much like associates, with clear rights to do some things and not others. He raises what agents should be allowed to call outside the company through Model Context Protocol servers, how to document agent actions with an audit trail replicable enough to defend to health authorities, and where a human must act as a circuit breaker, where a human simply observes, and when a validated process can run fully automated. He compares an agent to a very smart, very fast intern whose work you would review before sending it to the CEO.

This is the conversation I aimed to start in AI Governance Starts Before Agents Can Act. Permissions, lineage, and human checkpoints work best as native properties of the environment an agent operates in, established before the agent takes its first action. In L7|ESP, people and agents work within one governed environment, with attributable identities, defined permissions, and a common audit trail. L7|SYNAPSE^™ and its chief agent framework, Seven, orchestrate specialized sub-agents inside that environment, so every agent action inherits the controls and evidence a Quality team already relies on. People decide where the circuit breakers sit, and the architecture keeps them in place.

From lab-in-the-loop to CMC-in-the-loop

Diehl closes with what excites him most as he looks toward 2030: the convergence of dry and wet labs. In silico hypotheses feed semi-automated and eventually self-driving labs, which generate the data that trains the next models and shapes the next round of hypotheses. Novartis calls this lab-in-the-loop, and its $1 billion commitment to next-generation laboratories at its San Diego site shows how seriously it takes the idea.

He is also refreshingly realistic. He does not expect AI to deliver new drugs in weeks by 2030, and he frames the near-term gain as running more loops, failing faster, and recovering faster.

I share that view, and last week I wrote about extending the loop further. Every molecule that succeeds in discovery must then move through process development, analytical methods, technology transfer, and clinical manufacturing, often across a network of external partners. The loop that matters most spans that entire journey. When recipes, methods, and quality data travel with the molecule as governed digital content, the loop keeps turning well past the bench. In our work with customers, we have seen technology transfer timelines move from 18 to 24 months toward six.

Making AI routine

What I appreciate most in Diehl’s perspective is its patience. Chase impact over novelty. Stay close to the scientists. Pivot when something does not work. Keep the patient at the center. Those principles hold whichever model leads the benchmarks this month.

For the rest of the industry, the practical lesson is about sequence. Build the operational foundation first, capture context where the work happens, govern agents inside that foundation, and connect the loop from discovery to manufacturing. That is how AI becomes “boring”, and that is when it starts to matter.

Source: McKinsey & Company, Novartis’s Christian Diehl on scaling AI beyond the demo, Eureka! podcast, September 21, 2026.

—

FAQs