August 5, 2024
Why Gen AI isn’t enough for the life sciences
Generative AI’s expansive capabilities present opportunities to tackle a variety of tasks that previously were not addressable by software. The natural temptation that comes from this is to see everything as a solution that AI might solve. And while that might be true at a surface level, when you break down exactly what the “jobs to be done” are for software in industries like the life sciences, that’s not so true.
Using AI in the Real World
Generative AI’s expansive capabilities present opportunities to tackle a variety of tasks that previously were not addressable by software. The natural temptation that comes from this is to see everything as a solution that AI might solve. And while that might be true at a surface level, when you break down exactly what the “jobs to be done” are for software in industries like the life sciences, that’s not so true.
This post highlights what it takes to build a good Generative AI application and why.
Understanding “Traditional” Software
Traditional software operates deterministically, performing the same task consistently according to a set of predefined rules. This reliability has led to significant advancements in creating structured datasets, automation, and more. However, the deterministic nature of traditional software limits its effectiveness in dealing with the unstructured data prevalent in the real world. While structured data has immense value and applications, it only covers a portion of the data landscape, and, accordingly, a portion of work to be done.
Understanding Generative AI
In contrast to traditional software, AI is probabilistic, excelling at processing unstructured data but comes at the cost of inherent unreliability. While generative AI is theoretically capable of addressing many problems, its probabilistic nature makes it unsuitable for tasks requiring high reliability (consider those in the life sciences in particular). The more complex or lengthy the task, the greater the likelihood of unexpected outcomes. This makes generative AI a powerful tool, but certainly not a one-size-fits-all solution.
Hybrid AI Solutions
Most real-world problems involve a mix of structured and unstructured data. As such, pure AI systems or purely “traditional” software systems alone can’t address the full spectrum of tasks and needs, especially in complex, multi-step processes like those found in enterprises.
This makes two things true: 1) a pure AI system is likely to do a poorly at whatever set of tasks it's given, and 2) a pure AI system alone is unlikely to solve non-trivial and economically meaningful problem. This is where hybrid AI systems come in. These days, these two facts manifest to life sciences organizations testing AI as lots of interesting pilot projects with potential but few or no real AI use cases creating economic value (like cost savings or time savings, etc…).
Hybrid AI systems, which combine traditional software and AI, offer a more sophisticated approach that blends the reliability of traditional software and the adaptability of AI, allowing organizations to benefit from AI in a way that they can’t with standalone systems. In particular, a highly effective approach to building AI applications so far involves trying to strip out as much AI as possible from an application so that it can perform as reliably as possible.
Implications
The key implication of these observations is that deploying AI effectively requires more than just building generative AI capabilities; it involves creating comprehensive software solutions.
Developing systems that handle both structured and unstructured tasks is complex, and building reliable applications demands integrating AI with traditional software. For example, in tasks like document drafting, a collaborative AI tool that supports initial generation, redrafting, comments, and quality control steps is far more value-additive than a static and standalone drafting tool.
Overall, this means that enterprise AI becomes more difficult to build since it is much more complicated than building an “AI tool” (whatever that might be), it’s about building what is essentially an enterprise software application as most would’ve conceived of it in the pre-ChatGPT era.
Tangibly, this translates to higher cost and longer timelines associated with building AI systems than most anticipate. And for those not building AI in-house, it should make them skeptical of organizations that broadly claim AI as a feature of their offering because, at this point in time, the most high-performing enterprise AI systems are the ones that are using Generative AI for very specific, well-defined components of a workflow. This is the only way to effectively mitigate against some of the inherent issues with Generative AI tools (things like hallucinations and inconsistencies).
