Evaluation is the bottleneck (McKinsey)

McKinsey's latest technology-trends session puts words to a shift that is easy to feel and hard to budget: AI can now propose thousands of drug candidates in a short amount of time, and help evaluate them. The session's own question is the one that matters. What happens when the speed of discovery starts to outrun the systems we rely on to validate and scale?

McKinsey's caution is that in many cases AI cannot accelerate biological testing or regulatory review, so in pharmaceuticals the bottleneck slides downstream, into parts of the pipeline that were never designed for this volume of candidates. I would expect versions of the same pattern wherever options are generated. Marketing can produce a thousand variants of a campaign, legal can draft a thousand clauses, and someone still has to know which one is right.

If generation is cheap, the scarce asset becomes the capacity to judge: validation data, test environments, expert reviewers, and the governance to say no with confidence. The recommendation that follows is to make evaluation capacity an explicit budget item, owned and reviewed like any other scarce input.

Pointing AI at the evaluation problem too can help, and McKinsey notes the same technology can assist with evaluation as well as generation. The risk is correlated error: an evaluator that shares the generator's blind spots may agree with it faster rather than check it better. In regulated work the prudent design is a named accountable owner for the decision, with checks independent of the system that produced it.

None of this argues for slowing the generators down. It argues for funding the judges with the same enthusiasm, because a thousand candidates are worth nothing until someone can identify the one that works.

What is the evaluation bottleneck in your organisation, and who is funding it?

#AI #Innovation #Strategy

Sources:
- McKinsey Live, "The technology shifts every business leader should be watching" (session recap): https://lnkd.in/gSFyNg8s