A few years ago, if you asked a pharma executive what "AI in life sciences" meant, the answer was usually some variation of AlphaFold. Which was fine, because AlphaFold deserved the attention. It also wasn't really the answer, because AlphaFold is one model that does one thing well, and the actual landscape of AI in life sciences had already become something much wider and stranger.

By 2026, a 2025 review counted more than 200 foundation models in drug discovery alone, and that's before you get to the clinical, regulatory, medical affairs, and commercial systems being built on top of them. There's GPT-Rosalind from OpenAI in research preview. There's Evo2 trying to model entire genomes. There's Recursion sitting on 50 petabytes of phenomics data. There's MedGemma reading chest X-rays, TxGemma benchmarked across 66 therapeutics tasks, BenevolentAI 's R2E grounding answers in retrieved evidence, and Insilico Medicine's Chemistry42 running a small-molecule discovery pipeline that has actually produced clinical candidates.

So I want to do something useful here: walk through what's actually in this landscape, where the real value lives, and the one thing I think the industry keeps getting wrong about it.

Stop treating "AI in life sciences" as one thing

It is not one thing. A protein structure model and a regulatory drafting assistant share roughly as much DNA as a centrifuge and a CRM. They solve different problems, run on different data, fail in different ways, and require different kinds of governance to use safely.

The cleanest way I've found to think about it is five model families.

The first family is biological and chemical models. AlphaFold 2 and 3, RoseTTAFold, ESM3, RFdiffusion, Evo2, Geneformer, AtomNet, Chemistry42, Schrödinger. These work over the molecular foundations of life - proteins, ligands, genomes, cells. They predict structures, generate molecules, design proteins, and rank targets. This is the part of the field with the most spectacular progress, partly because the problems have clean validation loops. You predict a structure, you test it. You generate a molecule, you assay it. AlphaFold won a Nobel Prize for a reason.

The second family is clinical and patient intelligence models. Med-PaLM 2, MedLM, MedGemma, GatorTron, TxGemma, AMIE, Recursion OS. These reason over the messy stuff: EHRs, trial protocols, claims data, imaging, adverse event reports. The opportunity is enormous (the clinical phase is where most drug spending goes and most drugs fail), and so is the difficulty. The data is sensitive, the ground truth is debatable, and the stakes touch patients directly. Google 's own MedGemma model card states that outputs are not intended for direct diagnosis or treatment decisions without validation. That sentence is doing a lot of work.

The third family is knowledge and evidence models. BioGPT, GPT-Rosalind, BenevolentAI's R2E, and the increasing number of enterprise RAG systems built over proprietary medical, regulatory, and commercial content. These don't generate molecules or read X-rays. They synthesize the literature, draft the response document, compare the label, find the precedent. This is where medical affairs and regulatory teams live, and where the bottleneck isn't writing - it's traceable writing that can survive MLR review.

The fourth family is commercial and market intelligence models. There's no AlphaFold for launch planning. At least not yet. Most commercial AI in pharma right now is governed LLM/RAG systems over payer policies, HEOR evidence, competitor intel, and field activity. The model itself is rarely the differentiator; the integration with the workflow is.

The fifth family is workflow and agentic systems. This is the layer that's started to matter most in the last year. NVIDIA BioNeMo, Schrödinger LiveDesign, Chemistry42, GPT-Rosalind in tool-use mode, and a growing set of platforms built around the model rather than as the model. These coordinate other models, call tools, route data, and try to fit AI into how scientific or commercial work actually happens. (Full disclosure: this is the part of the stack I work on at LogicFlo AI, which is why I think about it a lot.)

Where the value actually lives

Now here's the part the press releases rarely tell you straight. AI value is not evenly distributed across the value chain.

The strongest public evidence - the prospective papers, the clinical candidates, the benchmarks that mean something - is concentrated in early R&D. Target identification, hit finding, virtual screening, structure prediction, and lead optimization. Atomwise reported a 318-project prospective study suggesting AI can replace high-throughput screening as a first step in many discovery programs. Insilico has taken AI-discovered molecules into Phase 2. Schrödinger's physics-plus-ML platform has become the operating system for a meaningful slice of medicinal chemistry. RFdiffusion can design protein binders that work in the lab.

As you move downstream - clinical development, regulatory affairs, manufacturing, commercialization - the model evidence gets thinner. Not because AI doesn't matter there. It matters enormously. But because the value increasingly depends on things that are not the model: retrieval quality, document traceability, audit trails, integration with Veeva Systems or QMS or MES, the ability to defend an output in front of a regulator or an MLR reviewer.

This is why TxGemma can be benchmarked across 66 therapeutics tasks and still not be plug-and-play for clinical decision-making. It's why Med-PaLM 2 specialists preferred its answers to generalist physician answers 65% of the time in a research pilot, but you still wouldn't deploy it as a standalone clinical tool. It's why GPT-Rosalind is in research preview through a trusted-access process, not a self-serve API. The model can be brilliant. The model is also one component in a stack that has to be defensible end to end.

The FDA's 2025 draft guidance on AI in regulatory decision-making, reinforced by 2026 guidance on Good AI Practice in drug development, made this explicit: the regulatory unit is not "AlphaFold" or "GPT-Rosalind" in the abstract. It is an AI system used for a defined context of use, with versioned data provenance, validation plans, failure-mode analysis, drift monitoring, and change control. The model is a feature. The context-of-use and governance wrapper is the product.

The mistake I keep seeing

The most expensive AI mistake I see pharma companies make right now is a procurement mistake disguised as a technology mistake. It looks like this: an internal team gets enthusiastic about a model, runs a benchmark on a sample of their data, the model performs well, and the team writes a memo recommending adoption. Eighteen months later there is a polished dashboard, a few well-attended demos, and approximately zero impact on any pipeline decision. The model worked. The model also never actually entered the workflow where decisions get made.

The question that keeps coming up in conversations is "Which model do you use?" It's a fair question. It's also slightly the wrong one.

The more useful question is: where is the workflow breaking down, and what kind of system fixes it? Because the model is rarely the bottleneck. The bottleneck is usually some combination of trusted data, retrieval, role-based permissions, version control, audit trails, and human review - and an actual integration into the place where the work happens. A regulatory team that has to copy a model's output into a Word doc and then manually check every citation is not getting much acceleration. A medical affairs team whose AI-generated response has no source attribution is just generating more MLR work.

The companies getting real value out of AI right now have figured out that the model is one layer in a stack, and most of the work - and most of the value - is in the other layers. The data layer. The retrieval and orchestration layer. The governance layer. The human-review layer. Benchling and Dotmatics have built businesses on what amounts to "the data backbone that makes the model actually useful," and they're doing fine. NVIDIA 's BioNeMo strategy is explicitly to be the platform layer rather than the headline model. The pattern is consistent.

For regulated industries, this distinction is everything. If an AI output could affect a scientific, clinical, regulatory, manufacturing, or promotional decision, the governance wrapper is part of the product. Regulators and litigators do not separate "model selection" from "quality system design," which means buyers shouldn't either.

What to actually do with this

If you're a pharma leader trying to make sense of where to put AI investment in the next eighteen months, three things might help.

First, when you evaluate an AI capability, separate the model question from the system question. The model question is "does this perform well on tasks like the ones we have?" The system question is "can we actually deploy this in our workflow with the data, controls, and audit posture we need?" Both have to be yes. Most procurement processes only seriously evaluate the first one.

Second, calibrate your expectations by stage. AI is most ready in early discovery, where validation loops are clean and the ROI on a faster, better hit list is measurable. It is genuinely promising but governance-heavy in clinical development. In regulatory, medical affairs, manufacturing, and commercialization, the model is rarely the differentiator - the workflow integration is. None of these are bad places to invest, but they require different mental models. The mistake is buying a clinical AI tool with the playbook you used for a discovery platform.

Third, ask vendors a different question. Not "what does your model do?" but "what does your model do inside my workflow, with my data, under my governance constraints, reviewed by my team?" The vendors who can answer that question concretely are the ones worth shortlisting. The ones who can only show you benchmarks are selling you a demo.

The next phase of AI in life sciences won't be defined by which company releases the smartest model. It will be defined by which companies can turn model capability into organizational capability - moving from information to action with speed, trust, and accountability. That's the boring-sounding sentence I keep landing on, and the more I work in this space, the more I think it's the only one that matters.