The rise of agentic in drug R&D - Owkin & AZ

TL;DR

AstraZeneca licensed Owkin's K Pro platform in a three-year deal, with competitive intelligence named as the explicit autonomous capability. I think the gap between "agentic as positioning" and "agentic as delivery" is the most important thing techbio leaders should be reading every announcement for. Agents work best today where inputs are bounded, outputs are reviewable quickly, and users can validate fast — CI clears all three. The harder problems (scientific judgement, long-horizon chaining, end-to-end regulated workflows) are solvable but need purpose-built systems, not general-purpose orchestration. The teams that pick the right unglamorous workflows first will be ready when those harder problems become tractable.

AstraZeneca licensed OWKIN's K Pro platform last week in a three-year deal. K Pro is positioned as an AI Scientist. The autonomous capability explicitly named in the announcement is competitive intelligence.

That detail is worth dwelling on. Owkin's CEO described the future of pharma as agentic. The deal itself is more specific about where the agentic part actually lands. The space between agentic as positioning and agentic as delivery is what techbio leaders should be reading every announcement for.

What I think is realistic this year

Agents work where the inputs are bounded, the outputs are reviewable in minutes, and the user can validate quickly. Competitive intelligence clears all three. Trial registries, patent filings, recruitment data, and outcome signals are structured enough to ingest, comparable enough to summarise, consequential enough that a senior analyst will read every word. I ran CI in one of my previous roles. The bottleneck was never insight. It was the 80% of the day spent collating, deduping, and cross-referencing. An agent that owns that 80% turns a five-analyst team into the output of fifteen.

What is genuinely harder right now

Agents that exercise scientific judgement on novelty or mechanism. Agents that chain reliably across long horizons in regulated contexts. Agents that operate end-to-end in target validation or regulatory decision-making. These are solvable problems. They need purpose-built systems and domain data, not general-purpose orchestration.

The useful frame for pharma and techbio leaders is workflow-by-workflow.

Most early wins will come from workflows that look unglamorous but compound quickly. The teams that pick those workflows first, deploy real systems with real data, and learn what works will be the ones ready when the harder problems become tractable. And they will, sooner than most expect.

Key takeaways

• The autonomous capability explicitly named in the AZ–Owkin deal is competitive intelligence — not broad scientific autonomy, despite the "AI Scientist" positioning.
• The space between agentic as positioning and agentic as delivery is what I think techbio leaders should be reading every announcement for.
• Agents work best today where inputs are bounded, outputs are reviewable in minutes, and users can validate quickly — CI clears all three of those bars.
• Having run CI myself, I know the bottleneck was never insight — it was the 80% of the day spent collating, deduping, and cross-referencing; an agent that owns that 80% turns a five-analyst team into the output of fifteen.
• Agents exercising scientific judgement on novelty or mechanism, chaining across long horizons in regulated contexts, or operating end-to-end in target validation are solvable problems — but they need purpose-built systems and domain data, not general-purpose orchestration.
• The useful frame for pharma and techbio leaders is workflow-by-workflow: pick the unglamorous workflows that compound quickly, deploy real systems with real data, and learn what works.