# Do not get out of bed for an AI project worth less than $100 million.

> At the AI in Clinical Development Summit we hosted in New York, one line defined the room: don't get out of bed for an AI project worth less than $100 million. I walked away with four ideas about where pharma's AI transformation is actually landing — and a conviction that regulatory submissions are heading toward agent-to-agent transactions built on machine-native data standards.

URL: https://www.ch-healthtech.com/insights/do-not-get-out-bed-ai-project-worth-less-than-100-million
Markdown: https://www.ch-healthtech.com/insights/do-not-get-out-bed-ai-project-worth-less-than-100-million.md
Published: 2026-05-28
Author: Christian Hein
Tags: technology/agentic-ai, technology/artificial-intelligence, technology/generative-ai, industry/large-pharma, function/clinical-development

---


## TL;DR

At the AI in Clinical Development Summit we hosted in New York, senior leaders from AstraZeneca, Sanofi, Bristol Myers Squibb, Amgen, and Pfizer converged around a single discipline: concentrate AI bets, govern them like pipeline capital, and measure value in development time converted to revenue. Four ideas stood out — from generative AI unlocking early ROI in discovery, to real-time trial signals now validated by the FDA. The deeper thread is that 60 years of human-written, human-read documents are giving way to a future where regulatory submissions look more like transactions between agents, demanding data standards built for machines, not digitised paper.

Do not get out of bed for an AI project worth less than $100 million.

That was the line of the day at the AI in Clinical Development Summit we hosted in New York. In the room: Chief Digital Officers, Chiefs of AI for Science, R&D SVPs and heads of regulatory policy from AstraZeneca, Sanofi, Bristol Myers Squibb, Amgen and Pfizer.

Four ideas stayed with me.

- Discovery is where the early returns are landing. One team showed how generative AI now scales the creation of entirely new biomarkers for antibody-drug conjugates, compressing a discovery step that normally takes far longer. The first hard ROI is in the science itself.

- Your data foundation decides everything stacked on top of it. One major pharma spent two years pulling clinical operations data out of scattered spreadsheets into a single structured workbench. 300 trials now run on it, and the cost of the data operation fell by roughly 30 percent.

- Treat AI capital like pipeline capital. That $100 million "get out of bed" bar is really about concentration: a few big bets governed with the CFO, five KPIs rather than fifty, and value measured as development time converted into revenue before loss of exclusivity.

- Real-time trials are becoming a reality, even though still pilot stage. The FDA has now validated early signal capture during a live trial. It means seeing a meaningful signal while the study is still running. That said, the live raw-data firehose some people imagined after the announcement is the wrong picture, and the agency is not built for it.

The thread tying it together: for 60 years clinical development has produced a truckload of documents, written by humans to be read by humans. The formatted submission is an artifact of that constraint.

We are now nearing a future where a regulatory submission is closer to a transaction between agents, one side preparing the evidence, the other reviewing it. That demands genuinely new data standards, built for machines rather than digitised paper. Each leader in that room, in their own corner, is quietly rebuilding for it.

## Key takeaways

- I came away from the summit with a clear signal: the leaders in that room are concentrating AI investment into a few large bets, not spreading it thin across dozens of small pilots.
- Discovery is where the first hard ROI is landing — generative AI is now scaling the creation of entirely new biomarkers for antibody-drug conjugates, compressing steps that normally take far longer.
- Your data foundation decides everything stacked on top of it — one major pharma spent two years consolidating clinical operations data into a single structured workbench, and the cost of the data operation fell by roughly 30 percent.
- Treating AI capital like pipeline capital means governing it with the CFO, tracking five KPIs rather than fifty, and measuring value as development time converted into revenue before loss of exclusivity.
- Real-time trial signal capture has now been validated by the FDA, but the live raw-data firehose some people imagined is the wrong picture — the agency is not built for it.
- For 60 years clinical development has produced documents written by humans to be read by humans; we are now nearing a future where a regulatory submission is closer to a transaction between agents, demanding data standards built for machines rather than digitised paper.