AI Is Not Ready to Be the Decision-Maker — And Pretending Otherwise Is the Real Risk
August 3, 2026
No Comments
Read More »
Close
The industry’s AI rhetoric has outrun its AI maturity by several years, and the gap is being covered up with confidence nobody has earned.
At a BIO 2026 panel on AI in commercial forecasting, someone said the quiet thing about where this industry actually stands: current AI-assisted forecasting is “like a rudimentary GPS.” Not autopilot, not even cruise control: a device that gives you a plausible direction and expects you to keep your hands on the wheel.
That’s the honest state of the art. So why does so much of the industry talk about AI forecasting like it’s already driving?
Here’s an anecdote that should be more famous than it is. At Coherus Oncology, SVP Ashutosh Joshi asked Copilot for a forecast. In minutes, it checked epidemiology data and ran multiple methodologies. “A lot of power,” he said. And it is. What used to take an analyst days of manual cross-referencing now takes minutes. Nobody sane wants to go back.
But notice what that anecdote actually demonstrates: AI compressing the time to produce an output, not AI producing a trustworthy output. Those are different claims, and the industry keeps collapsing them into one.
Rigel Pharmaceuticals is doing something more instructive with the same tools. EVP Joseph Lasaga described using ChatGPT-type tools for scenario planning: testing assumptions, generating directional information for target product profiles, stress-testing diligence questions. Genuinely useful, and notably not, in his own description, a driver of actual decisions. Rigel still does primary research to get a direct perspective before committing capital. That’s not timidity; it’s the correct level of trust for a tool that hasn’t earned more.
Contrast that with the industry’s dominant posture, which is closer to: we have a Copilot license, therefore we are AI-forward, therefore our forecasts are more rigorous now.
Here’s the part that should actually keep executives up at night. It isn’t the companies moving too slowly on AI. Slow adoption is a visible, fixable problem: you can see it, name it, and budget against it. The dangerous failure mode is invisible: a leadership team quietly starts treating a directional tool as a decision-making one, and nobody flags it, because flagging it looks like being anti-AI in a room full of people who want to look AI-forward.
Katie Sinaikin at Revolution Medicines named the real issue with more precision than almost anyone else on that stage. The hard part isn’t the technology. It’s knowing where you have enough confidence to bolt AI onto an existing workflow, and admitting that most companies need outside help just figuring out where that line is. She called it a confidence gap, and she’s right that it’s real, especially for a smaller company facing a pivotal, bet-the-company decision, where the cost of misplaced trust in a model isn’t an awkward board meeting but the company itself.
But here’s the provocative part, and the part worth sitting with: that confidence gap isn’t a bug; it’s the single healthiest instinct in the room right now.
The ones who should actually worry executives aren’t the people asking “how do we know we can trust this?” That question is the immune system working correctly. It’s the people who’ve stopped asking it: the ones who now present an AI-generated forecast with the same unqualified certainty they’d use for an audited financial statement, because the output arrived in a clean chart with decimal points, and decimal points look like rigor.
That’s the mechanism underneath the real risk. AI output doesn’t announce its own uncertainty. A human forecaster hedges out loud: “directionally, I think,” “low confidence given the comp set,” “this assumes the epi data holds.” A model just outputs a number. And a number, formatted cleanly, reads as more authoritative than the hedge-filled sentence a skilled analyst would have given you instead. Novartis’s Hassan Ahmed made a related point on the same panel: for large organizations, automation is only as good as the underlying process it’s automating. Bolt a confident-looking tool onto a shaky process and you don’t get a better answer. You get a shakier process wearing a lab coat.
There’s also a political dimension nobody wants to name directly, though the panel came close. Forecasting inside a biopharma company has never been a purely technical exercise; it’s a negotiation among commercial, medical, and finance, each with incentives to land on a number that helps their case. AI doesn’t remove that political layer. It gives it better camouflage. It’s much harder to challenge a forecast that arrived via “the model said so” than one that arrived via “Dave in commercial thinks so,” even when Dave and the model are drawing on comparably thin evidence.
So here’s the standard worth adopting, out loud, in every forecasting review from here forward:
The industry doesn’t have an AI adoption problem. It has an AI humility problem, and right now, humility is losing to the pressure to look transformed.
At Scimitar, Akira Robinson serves as Partner, Commercialization. He operates at the intersection of commercial strategy, launch execution, and market access, advising biopharma executive teams at critical moments where launch readiness directly determines asset value and time to market. With 20 years of experience across life sciences, diagnostics, biologics, and digital medicine, his work tackles complex commercial challenges for teams across the US and globally. His expertise spans the full commercial value chain—including launch planning, licensing, market access, pricing, analytics, marketing, sales distribution, and patient services across therapy areas, including: oncology, radiopharmaceuticals (RLT), rare disease, CNS, cardiology, and gastroenterology.
