This entry draws heavily on Amrita Pati's Day 1 keynote at DataDrivenPharma West 2026. Amrita Executive Director of Computational Biology and Translational Analytics in Amgen's Precision Medicine department.
The keynote in brief
In her Day 1 keynote, Amrita made a case that biopharma is in the slump phase of the AI productivity paradox, where heavy investment has yet to show up in the bottom line. She also confirmed that the popular 1 of 3 AI pilots being a success figure is consistent with their experience at Amgen.
The AI productivity paradox, and the formula behind it
Amrita cited Blue Matter Consulting's April 2026 white paper, AI Is Everywhere in Biopharma R&D…Except in the Productivity Numbers, by Tara Austraat-Churik and Tim Barder. The Gartner Hype Cycle reminiscent pattern plays out in the following way: transformative technology attracts heavy investment and then disappoints for years before the delayed productivity boom finally arrives, and Amrita believes biopharma is sitting in that disappointing stretch with AI right now.
One of the key issues for pharma is the measurement delay. For example, an AI-guided decision to stop a trial early saves a tremendous amount of money, but won't show up in productivity numbers until this win is accounted for in aggregate across multiple programs in the form of reduced time or money (or both) - which takes years. Data foundations are another drag, since most companies are still working on wiring together siloed systems. Expectations are off as well: an AI that writes a good email won't automatically support a clinical decision for a diverse patient population.
Blue Matter sums it up in the following Net Productivity Formula:
N is the number of AI tools adopted.
U_i is the efficiency gain each tool delivers on its own.
C_friction is the cost of human bottleneck work at every unintegrated connection between tools: manual data entry, context switching, troubleshooting, validation.
I_level scores the maturity of the central intelligence layer that connects the tools to one another.
The point Amrita drew out is the N² term. Gains add up tool by tool, but friction grows with the square of the tool count, and it compounds with cognitive load, low trust, and human bottlenecks. Integration sits in the denominator, which helps to offset the N² term.
The Blue Matter paper makes the conclusion that a mediocre tool that's fully integrated can deliver more than a superior tool sitting in a silo. Therefore, picking the best tool for each task is not the best strategy in a holistic sense.
As a personal aside, I feel this too in the daily operations of DataDrivenPharma - while each new standalone tool may solve an important business problem, there is a cost for each additional solution in the stack. This was the case long before GenAI entered the picture, but it’s still interesting to see it laid out in an equation. I definitely get tool fatigue, and I’m seeing the value of good-enough solutions in a more comprehensive product, over near-perfect solutions in narrow products. This has an impact for our small business, and compounds exponentially for a pharma company’s large headcount and complex operations.
From the floor
The audience questions came down to trust. The first was trust in the models themselves: as models produce more plausible-sounding inferences, AI-generated noise could pollute the biological knowledge base in the way low-quality journals pollute the literature. The second was trust in the companies building them: do the frontier AI labs have the best intentions, and is trusting them with hard-earned biological data a good move?
Amrita leaned on the optimist side, saying that the problems are there, and in the discussions she’s had between pharma and AI labs, there is a lot of positive intent, and there is a genuine interest to further the field.
The Close
Amrita closed with a genuine sentiment that the limit is human imagination. Today's AI is trained on answers to yesterday's questions; and most companies can get access to the same models anyway. The edge will go to teams that use models to find biology's blind spots and generate data nobody else has.
