Two hits, two partials, one miss

Two hits, two partials, one miss. That is how my five AI predictions for 2026 are holding up at halftime.

I published them in January. Time to mark my own homework.

1. Ambient scribes become table stakes: pretty much true

The conversation moved from "Should we pilot this?" to "Which product, how do we integrate it, how do we govern it?" Ambient documentation is being embedded into core workflows, and European adoption is finally past isolated pilots.

What I underestimated: distribution and workflow integration now matter more than marginal differences in model quality.

2. The first real "AI-discovered drug" headline: partially true Takeda's zasocitinib delivered positive Phase 3 results, outperforming deucravacitinib. Meaningful clinical validation.

But I would be more careful with the label today. Zasocitinib was computationally enabled. Calling it unambiguously "AI-discovered" stretches the evidence.

3. Self-driving labs move from demo to production, staying narrow: I think I nailed this Closed-loop platforms are moving into real R&D environments, pharma (Pfizer, Roche, Novartis) is hiring for them, and the big AI labs and hyperscalers are just getting started here.

These systems still shine at bounded tasks: reaction optimization, formulation, catalysis, specific automated workflows.

4. Trial-risk prediction becomes the default: probably too early to say AI is increasingly used in protocol design, site selection, enrollment forecasting and go/no-go calls.

But "default" was too aggressive. Unfortunately. I hope we see more acceleration in the next six months.

5. The TechBio Series B shakeout accelerates: not so sure so far Capital is clearly concentrating. BioPharma Dive counted roughly 70 venture-backed biotechs raising over $9.1 billion in H1, most already in the clinic.

Selection is harsher. But I cannot yet point to the cohort-level shakeout I predicted. The wave of quiet shutdowns remains unproven.