Sluggish Corporate AI Adoption Has Motivated Entrepreneurs To Pick Their Spots 

As economic historian Carlota Perez has described, there is typically a significant time lag between when the promise of novel technology begins to emerge and the productive deployment of this technology at scale; TR readers will recall the discussion here from June 2023.

Today, we are seeing this with generative AI, an emerging technology that everyone is still trying to get their arms around (see this TR discussion).  The inevitable uncertainty associated with these early days has hardly discouraged large consulting firms, who have all developed “AI playbooks” and are busy persuading potential corporate clients that they are lagging their peers in AI and risk impending extinction. 

A Cautionary Tale From Big Pharma

Nevertheless, meaningful (vs performative) adoption of generative AI with most large enterprises has been predictably sluggish (see this TR discussion from August 2024).

Ziv Bar-Joseph, the former VP-Head of R&D Data and Computational Sciences at Sanofi, candidly and generously shared on LinkedIn lessons learned from his recent experience at the large French pharma. These takeaways include (emphasis added):

– Pharma moves slowly. It takes over 10 years to develop a drug. So its not surprising that planning and decision making at big pharma can look very long and frustrate potential partners (and our own employees). Some of it is just bureaucracy. But much of it is intentional. These are often decisions that will have long lasting impact and its important to get them right.

– Adoption remains a major challenge. Not because of any principled objection to AI or new technology. But rather because it is very hard to change the way people work. Other issues affecting adoption of new technology are changing needs, people leaving and change in priorities and focus.

–  Sanofi is not an AI company. While it continues to develop cutting edge AI tools, Sanofi would prefer to purchase or partner rather than build AI products internally.  This makes economic and business sense. But it can be frustrating for our internal teams, especially when the decision comes after internal work already started on a similar product (usually because at the time we started the external solution was not available).

These findings are neither unique to Sanofi nor likely to shock regular TR readers. Yet Bar-Joseph’s insights emphasize a real challenge faced by the field: how to most effectively leverage AI when it’s brutally difficult to meaningfully implement AI in large corporations at a non-glacial time scale.

Don’t Make The Tool – Implement It

One answer that seems to be emerging – at least among impassioned AI investors – is to identify focused and more manageable opportunities (vs transforming a giant pharma corporation) and drive the AI mediated change yourself.

Cass Mao, a Silicon Valley tech entrepreneur, wrote recently on LinkedIn:

I know 10 different people leaving venture investing right now to do a PE [private equity] play with AI.
The basic thesis being: more opportunity than ever right now to drive value in SMB [small and medium sized businesses] using technology. But adoption is slow, super fragmented market, a ton of competition with 100s of new tools launching every week.
“Easier” to buy a small company, drive adoption, and reap rewards via direct ownership of the bottom line – a few million in personal upside within a few years.
vs. be deploying capital in the fast moving froth, high vals, high churn, high competition. so many tools funded and fighting the distribution war, competing with similar tools to seek adoption.
you’re seeing tools that cost $20/month eliminate **tens of thousands of dollars of cost**.
better to be the SMB with 50 different ways to save $30,000, than one of 50 SaaS cos charging $240/year.

In short, she’s suggesting that:

     (a) Cheap AI tools can save a significant amount of money if deployed effectively and in the right context;

     (b) If you can identify the right opportunity (basically a job amenable to your tool, and a small organization where you can actually implement the technology), you can do better implementing the AI yourself and pocketing the revenue from the efficiency gains vs trying to sell a particular AI tool in a very crowded market.

The question, of course, is how (or whether) this can be applied to biopharma.

Consider the following approach, suitably anonymized, but inspired by a real techbio example (there are probably a number of startups trying something similar).

Let’s say you’ve identified a specific, valuable aspect of drug development where you believe your technology (AI or something else) gives you an economically valuable competitive advantage. You might believe, for example, you have a more efficient way of running clinical trials. Such a company might then seek to in-license clinical-stage assets from pharmas, execute efficient clinical development past a value inflection point, and then out-license the further de-risked assets back to pharma at a higher valuation. 

The basic strategy itself isn’t particularly original, and startups, with conviction around one asset or another, in-license molecules all the time. The difference in this case is that because you believe your technology allows you to prosecute assets more efficiently, you focus on leveraging this perceived advantage, ideally by raising sufficient money (from entranced tech VCs, say) to fund enough shots on goal. 

Through your superior efficiencies of clinical development, you hope, you have a high chance of clinical, and hence financial success. (You, and especially your tech investors, may also hope your technology allows you to identify more promising assets than skilled drug developers alone, though I’m still very skeptical about this part.)

Beware of Pyrrhic Technological Victories

The success of this type of approach will depend, of course, both on how well the technology works and how much of competitive advantage it actually delivers – does it really move the needle for the tech-enabled company? 

Consider this example from genetics: metabolism of the blood thinner warfarin is strongly influenced by two genes, CYP2C9 and VKORC1. Genetic testing can determine whether you are likely to be a “fast” or “slow” metabolizer. Yet, utilizing genetic testing turns out to offer at best minimal advantages to the traditional clinical approach of “going low and going slow,” to arrive empirically at a therapeutic dose. Thus, while it might seem in theory that genotyping technology could improve care, in practice, most doctors haven’t embraced it because the impact seems less than the aggravation.

Similarly, techbio companies need to be sure both that their technology really works and that it imparts a meaningful advantage. It’s a tough ask, but one that’s clearly attracted the interest of both investors and entrepreneurs who are convinced about the promise of AI in biopharma and are intensively pursuing the highest-leverage place in which to deploy it.