Can Biopharma Make AI Sing?

“… When I’m with her I’m confused
Out of focus and bemused
And I never know exactly where I am
Unpredictable as weather
She’s as flighty as a feather
She’s a darling, she’s a demon, she’s a lamb…

… How do you solve a problem like Maria?
How do you catch a cloud and pin it down?”

The lyrics, of course, are from the beloved Rodgers and Hammerstein musical (1959) and later film (1965), The Sound of Music, sung by the Sisters of Nonnberg Abbey as they try to make sense of the remarkable force of nature who has appeared in their midst.

Biopharma leaders grappling with AI can relate — and they’re not alone.

AI’s Productivity Paradox

As John Cassidy reviews in The New Yorker, executives across many industries are trying to square the extravagant expectations for AI — especially GenAI — and their lived experience, which from a business perspective tends to be far more muted.

Cassidy highlights a pair of recent findings: 

• A large survey conducted this summer by a team of economists at several universities and the World Bank found that nearly half of all workers reported they were “using AI tools.”
• A study from researchers associated with the MIT Media Lab found that “Despite $30-40 billion in enterprise investment into GenAI… 95% of organizations are getting zero return.”

As Cassidy notes, the contrast between activity around a new technology and its demonstrated business impact was famously observed by Nobel laureate Robert Solow, who wrote in The New York Times Book Review in 1987, “You can see the computer age everywhere but in the productivity statistics.”  (For economists, that’s a sick burn.)

Readers of this column are familiar with this “productivity paradox,” and with the gap between what AI has promised and what it has delivered (so far) to the biopharma industry.

As I just discussed, Novartis CEO Vas Narasimhan has been explicit about the gap; speaking recently before a group of Harvard MS/MBA students (disclosure: I advise the program), he emphasized the promise of AI to improve the efficiency of some discrete processes, but he didn’t seem to feel that AI was on the threshold of substantively improving the efficiency of either discovering new targets or developing original medicines.

Apparently, Narasimhan is not the only one. He described (as I recall) a recent event where biopharma leaders were asked whether they saw AI impacting either their top- or bottom-line forecasts for the next 5-10 years, and none did – though discrete opportunities for incremental impact were mentioned.

The biopharma experience aligns with both the MIT result and with comments Cassidy reports from respondents:

• “The hype on LinkedIn says everything has changed, but in our operations, nothing fundamental has shifted… We’re processing some contracts faster, but that’s all that has changed.” – COO at midsize manufacturing firm
• “We’ve seen dozens of demos this year. Maybe one or two are genuinely useful.  The rest are wrappers or science projects.” – another respondent

“Pockets of Reducibility”

Where has success been achieved? According to the MIT report, “These early results suggest that learning-capable systems, when targeted at specific processes, can deliver real value, even without major organizational restructuring.”  

This echoes Narasimhan’s point and the approach this column has championed: seek “pockets of reducibility” (to use Stephen Wolfram’s memorable phrase) — discrete opportunities where the powerful but still-emerging technology can be gainfully applied.

(I’ve also discussed the concept in the context of developing personalized approaches to health.)

Show Me The Money

For some reason, many CEOs seem genuinely shocked (not Captain Renault shocked) that the productivity gains promised by the tech companies developing AI and the consultants implementing AI have not materialized. In a recent survey of two thousand executives by Akkodis, the share of CEOs “very confident” in their companies’ AI implementation strategies fell from 82% in 2024 to 49% in 2025. 

I have seen a version of this up close: the allure of AI-enabled productivity gains, presented seductively by skilled management consultants and amplified by boards worried about falling behind, is powerful. Given a choice between (a) embarking on a grand AI-inspired productivity initiative, led by confident consultants and producing slick progress reports for the board; or (b) pursuing modest, specific opportunities where technology can be applied gainfully – without promising profound cost savings – you can guess which option most C-suites will choose.

Ultimately, the anticipated productivity gains generally don’t materialize, and cost savings are achieved the old-fashioned way: by cutting programs and reducing headcount.

Why the Long Face?

Cassidy considers several reasons why GenAI has disappointed most businesses so far. One is tool fit: the MIT study found some of the most successful AI investments tended to be highly customized, narrow tools aimed at specific processes; less successful efforts chased generic solutions or attempted to build capabilities internally. Another possibility he raises: “for many established businesses, generative AI, at least in its current incarnation, simply isn’t all it’s been cracked up to be.” 

Finally, Cassidy brings up what strikes me as the most compelling explanation, and the one I’ve often emphasized in this column: it takes a long time to figure out how to use powerful emerging technology. We systematically underestimate the time and change required for widespread, productive adoption. 

Part of this is infrastructure: you can’t scale electric-vehicle adoption without widespread charging stations; similarly, the spread of Watt’s steam engine required railways to move coal. 

Another factor is workflow: initial adoption of new technologies tends to involve the substitution of new tech into existing processes. Replacing a steam engine with an electric generator in compact factories built around a single power source didn’t boost productivity. The game-changer was radically reimagining the workflow – Ford’s assembly line, an innovation enabled by electricity but not an obvious or inevitable consequence of it.

Moreover, forcing a new technology into old processes can even reduce productivity, at least at first, before improvements (ideally) start to accrue. This pattern is called the “J-curve,” Cassidy informs us, observing that “the journey along the curve can be lengthy.”

Pull > Push

This brings up another important, very human challenge I’ve encountered firsthand. Senior management, having been sold on the putative productivity benefits of AI, often believes the technology needs to be imposed upon a benighted workforce. More often, I suspect, the lack of adoption reflects discernment more than ignorance. The right move isn’t to jam AI tools, gavage-style, into every workflow because of an abstract commitment to “do AI.” It’s to de-average implementation and focus on energized lead users who are passionate about solving a particular problem — and where an AI tool could make a real difference, especially if developed and refined as a partnership between the tool developer and the lead user.

Adoption should be pulled by palpable utility, not pushed by executive edict.

Bottom Line

At times I find myself resonating with both the optimism of evangelists, who accurately perceive technology’s potential, and the skepticism of seasoned biopharma professionals, who accurately perceive the magnitude and complexity of the challenges the technology must overcome.

I continue to believe in the extraordinary, transformative promise of AI.  But it’s not magic. The most substantial early wins will come from tactical, high-leverage applications – pulled by motivated lead users and enabled by high-EQ technology partners — rather than pushed by decree.

Top biopharma R&D talent is drawn by the prospect of creating meaningful new medicines for patients.  They may be most familiar with techniques they trained on, but, like everyone else, they adopt compelling tools (from the iPhone to ChatGPT) when those tools actually help.  If AI enables a scientist to be more effective, or a team to make better decisions, they’ll use it – especially when they see peers doing so with palpable effect. 

My two cents: an approach to AI adoption that is strongly supported by top management but fundamentally driven by lead users represents the best path forward – for companies, for technology, and for the medicines we aspire to create together, trying to hold a moonbeam in our hand.