Digital transformation in pharma: from AI pilots to real ROI

Walk into almost any pharma or biotech organization in 2026 and you will find artificial intelligence everywhere and agentic AI almost nowhere in particular. Enterprise LLM subscriptions are live, copilots are open in every browser tab, and the token meter is spinning. What you will rarely find is the thing the technology is now capable of: a structured project in which agents own a concrete process (a screening triage, a competitive-intelligence watch, a bioanalysis pipeline) end to end, with metrics, an owner, and validated outputs.

We call this pattern free-floating tokens, and the gap it leaves is the difference between AI that feels productive and AI that changes the P&L.

The Benchling 2026 Biotech AI Report (a November 2025 survey of ~104 biotech and pharma organizations that already use AI in R&D: these are the leaders, not the laggards) puts numbers on the pattern. AI use is heavy where the work is simple, and thin where transformation would live.

Adoption and transformation figures from Benchling (2026); pilot outcomes from MIT NANDA (2025).

Behind the headline bars: 92 percent of organizations use AI in non-scientific use cases (document authoring, literature search, software engineering) versus 81 percent in scientific ones, and regular use splits 59 to 44 the same way. Benchling’s own conclusion is that “AI use cases are concentrated in simpler, low-friction, predictable workflows.” Meanwhile 39 percent use AI only for single, task-specific jobs, and just 22 percent span multiple R&D workflows across teams: a scientist uses one model to mine papers, another for a hypothesis, a third to design an experiment, “each step existing independently.” Only 26 percent use co-scientist agents and only 21 percent have adopted AI for workflow orchestration; both top the planned-adoption list for the next two years, which tells you where the leaders think the value actually is.

MIT’s NANDA initiative put the hardest number on it in The GenAI Divide: State of AI in Business 2025 (Challapally, Pease, Raskar, and Chari): across 300+ public deployments, 52 executive interviews, and 153 surveyed leaders, about 95 percent of GenAI pilots deliver no measurable P&L impact. Two findings matter more than the headline.

The failure is organizational, not technical. Benchling’s respondents rank data quality and availability (55 percent citing it as a major factor) and IP, security, and compliance (50 percent) as the top pilot killers. Insufficient internal AI talent ranks near the bottom at 22 percent. Gartner meanwhile predicts over 40 percent of agentic AI projects will be scrapped by 2027, mostly for operationalization reasons, not model quality.

Who builds with you changes the odds. In the MIT data, externally purchased or partnered tools succeed roughly twice as often as internal builds, and pilots that blend internal specialists with external expertise reach a 67 percent success rate versus 22 percent for internal-only builds. The empirically winning modality is not buying a tool or building alone; it is building together. Hold that thought.

Bain’s How to Break Through Pharma’s Innovation Bottleneck with AI (written for clinical development, but the logic generalizes) argues that winners will place their bets carefully, embed change at the right organizational layers, and pick a strategic lane rather than scattering pilots:

Beyond the lane choice, Bain’s prescriptions map exactly onto the free-floating-token diagnosis. Orchestrate development as one AI-powered system, not standalone pilots: agents that plan and execute end-to-end workflows with escalation to humans, built on shared standards so specialized agents plug in without new silos, with testing and monitoring baked in (“reliability is not optional in regulated environments”). And redesign the operating model: human-in-the-loop is “not a single checker but a defined set of decision rights embedded in workflows.” Bain recommends a RAPID-style framework (Recommend, Agree, Perform, Input, Decide) for reviewers, model stewards, and data curators, updated SOPs with escalation paths, and one accountable AI owner per workflow. The result, in their words, is “a thinner execution layer and a stronger judgment layer.”

Steps after Bain (2026); success rates from MIT NANDA (2025).

Here is where the how matters as much as the what. Hemant Taneja, CEO of General Catalyst, has spent years arguing that the industries that matter most (healthcare above all) will not be fixed by the classic disruption playbook. His transformative principles replace it with radical collaboration: technology companies and incumbents transforming concrete operations together, from inside. General Catalyst went as far as acquiring a health system (Summa Health) to prove the model, and Taneja’s one-liner names the destination: “the Amazon of healthcare is not a $1 trillion company but a $1 trillion ecosystem.”

The MIT data quietly validates the philosophy: 67 percent success when internal experts and external partners build together, 22 percent when organizations go it alone. Free-floating tokens are the purest go-it-alone modality there is: a subscription, no partner, no process, no shared accountability. Structured projects run in radical-collaboration mode are the opposite, and they are the configuration in which pilots actually survive contact with the P&L.

PharosBio has led structured AI projects in exactly this configuration with both industrial and academic partners: augmenting an in vivo research program by interpreting preclinical findings against human clinical and omics data, and prioritizing drug-combination targets in a space of hundreds of millions of candidate pairs. Each engagement follows the principles above: one concrete process, an accountable owner on both sides, explicit human decision rights, and scientific validation of every output. Using those principles, we help companies improve ROI in the areas with the highest potential and the shortest path to value. More worked examples are in our case studies.