AI in Drug Discovery: 10 Lessons from the Front Lines of Biopharma

White papers will tell you AI is transforming drug discovery. Press releases will tell you every major pharma company has a strategy. Neither will tell you where leaders are actually stuck, where budgets are really moving, or where the hype hits the cold wall of reality.

Over the past six months, I've spent hundreds of hours in those rooms — deep-dive sessions with Heads of Research, Heads of Digital, and VPs of Therapeutic Areas, alongside strategic conversations with AI foundation model companies and tech hyperscalers. Last week's J.P. Morgan Healthcare Conference crystallized a set of patterns that only emerge when you stop reading about the industry and start listening to the people running it.

Here's what I heard.

1. Big tech is doubling down — and they know why

AI is no longer a side project inside pharma. It's a board-level mandate. But the more interesting shift is why the world's largest tech companies — Google, Anthropic, OpenAI — have made life sciences a strategic priority.

The logic is straightforward: if the future value of AI lies in inference, drug discovery is the deepest well available. Our current understanding of biology is so limited that the gap between what we know and what is knowable represents an almost infinite opportunity. We are witnessing a global AI land grab, with biopharma as the prize.

2. The "last mile" is longer than a mile

The most important pattern I keep seeing: breadth is easy, depth is everything.

In biology, generalized solutions break at the last mile. The long tail of edge cases, context-specific biology, and program-specific nuances is where the actual value — and the actual risk — lives. If a solution doesn't go all the way into those domains, its ability to create tangible outcomes is capped.

This is why I don't believe big tech will solve this problem. And I don't believe a System of Record company throwing AI on top of an ELN will solve it either. You have to go deep. That takes time, significant investment, and a willingness to do work that doesn't look impressive in a demo.

3. Beware the era of beautiful PoCs

We are in a moment when it is dangerously easy to build a compelling demo. Agentic workflows and slick UIs can make a system look transformative in a thirty-minute meeting.

But there is a massive gap between a great demo and a production-ready tool — one that is compliant, robust, and scientifically embedded in real workflows. Moving from a Proof of Concept to something that delivers reliable value still takes 18 to 24 months.

Stop asking "Is it impressive?" Start asking: "How many times has this survived a production environment, and what is the evidence of durable value?"

4. The builder era is cooling

The classic build-vs-buy debate is shifting. The market is still roughly 50/50, but I'm watching a quiet exit from internal platform builds. Budget cuts and the pace of specialized vendors are combining to push organizations that once aimed to build everything in-house to back away.

The emerging reality is a hybrid: build your internal data lakes and highly specific tools for proprietary IP. Buy the specialized application layers and external data harmonization engines. The all-in-house platform play is losing credibility fast.

5. The "one tool to rule them all" is a fantasy

There's a lingering desire for a single system answer. Leaders want one platform that does everything. The reality is that the future discovery stack is a portfolio — and the smart leaders are starting to accept that intentional overlap is a feature, not a bug.

Two tools that do similar things can be exactly right if they provide different scientific lenses on a complex therapeutic problem. A platform optimized for structural biology and one optimized for high-throughput screening data might both offer "lead optimization" features — but their distinct perspectives are what give you the comprehensive view to actually advance a program.

Trying to collapse that into one tool means trading depth for tidiness. It's the wrong trade.

6. Data architecture is moving from integration to orchestration

How the industry brings data and tools together is evolving fast. We are moving past simple APIs toward systems that can reason across multiple sources — what's being called agentic orchestration.

The strategic question is now economic: which datasets justify the heavy lift of deep extraction and knowledge graph construction, and which are good enough via simple programmatic access? There is no universal answer. It has to be a case-by-case decision grounded in the specific therapeutic goal and honest quality evaluation of what you actually have.

7. Evaluating AI is now a core competency

With an explosion of vendors and models, the ability to rigorously evaluate AI has become a strategic capability in its own right — not a procurement function.

You need a framework that can quantify the value of a domain-specific tool versus a generic LLM, determine when your internal data is sufficient versus when you need external augmentation, and move past anecdotes to evidence-based decisions.

In my experience, this is the most important investment a pharma organization can make right now. It tells you where you actually stand — not where the vendor says you stand.

8. The data quality crisis is being papered over

Most new agentic systems are being built on low-fidelity, noisy public data. If you assume public data equals ground truth, you are building your AI strategy on sand.

Real differentiation doesn't come from the slickest agent. It comes from years of work spent cleaning, harmonizing, and extracting high-quality evidence from the messy reality of scientific literature and lab notes. No amount of orchestration fixes fundamentally broken inputs.

The painful truth: new entrants to the AI life science market have to move fast because the competition is intense. That pressure is actively preventing them from building things right. Getting the data foundation layer right can take years — even with AI. Most aren't doing it. They'll look fine in demos and fail in production.

9. Stack design is the new hard problem

Historically, a lack of specialized tools forced pharma companies to build their own imperfect solutions. Today, the ecosystem is crowded. The challenge has shifted from "find a tool" to "design a stack" — understanding how specialized platforms can work together without becoming a fragmented mess.

That's a harder problem than it sounds. It requires systems thinking, not just vendor evaluation. The organizations doing this well treat their discovery stack like an architecture decision, not a procurement exercise.

10. Most tech builders don't understand biopharma

Most entrepreneurs who come out of biopharma start drug companies — not tech companies that serve biopharma. So the people building tools for pharma are not drug hunters themselves.

The domain knowledge deficit is striking. Understanding the science is not enough. You have to understand the operational and decision-making realities of the industry — how research priorities get set, how budgets move, how a result in a Tuesday meeting becomes a program decision six months later. Most vendors don't. It shows.

The AI landscape in biopharma is moving fast and the competition is intense. But the fundamentals haven't changed — they never do. Real value comes from depth, data quality, and a relentless focus on the decisions that actually move programs forward.

The hype is loud. The signal is quieter. Learning to tell them apart is the whole game.