Only 22% of biotech organizations are reported to have AI workflows that span multiple R&D teams. The other 78% are still running AI as isolated point tools...

Recently, Benchling published its 2026 Biotech AI Report, which is based on a survey of about 100 biotech and pharma organizations that are already using AI in R&D.

Before delving into it, the caveat is that it's a view of the leading edge rather than the industry average. However, a few findings are still interesting to contemplate.

To start, the "killer apps" for AI in biotech are now pretty clear. Literature review is at 76% adoption, protein structure prediction at 71%, scientific reporting at 66%, and target identification at 58%.

The pattern is that AI works well where the data is clean, local, and easy to verify (surprise, surprise!).

In the parts of the pipeline where data is messy and spread across systems, adoption drops fast: generative design sits at 42%, biomarker analysis at 40%, ADME at 29%, IND submissions at 24%, etc.

What's interesting is that bottleneck has moved, with 55% of respondents saying data quality and availability is the main reason their AI pilots fail, and 50% point to IP and compliance friction.

Lack of internal AI talent comes in last at 14%. So the story is not that much about "we need better models" anymore.

The gap is that most organizations are trying to run modern AI on top of data systems that were built before any of this was possible, and that mismatch is showing.

Next, the talent is being grown internally rather than hired in, which is expected but still interesting to know.

Specifically, 67% of organizations source AI talent by upskilling existing scientists, while only 21% hire from tech. Apparently, it's faster to teach a bioinformatician some ML and AI wrangling than to teach an ML engineer some biology (not sure about that, though).

Anyway, another thing is that the build-vs-buy question has settled into a hybrid pattern: 60% buy commercial tools, 55% build proprietary models where their biology is unique, and 56% fine-tune third-party models. Only about a third outsource development.

Like I said, the sample is made up of biotech AI leaders, not a cross-section of the industry, so these numbers describe what the vanguard is doing, and are probably way more optimistic than the average reality.

The gap between this group and everyone else is probably the interesting story, and the report doesn't really get into it... time will tell.

Image credit: Benchling