AI for life sciences: what it does outside the lab
Almost everything written under this heading is about the bench: target identification, molecule design, imaging, literature. This section is about the other half of a drug company, the half that runs on email, dates and promises, and that nobody sells software for.
A biotech under about fifty people has no operations department. The VP of business development is also the person who remembers what was promised to a partner in June. The program director is also the person who notices that a slipped tox report has moved the filing date. The founder is the archive. None of that work is scientific, all of it is load bearing, and it is nearly invisible to the tools sold into this industry, because those tools are sold against the science.
That is the gap these pages are about. The system described across this section reads the mail and the meeting transcripts you were already generating, files each one into the program or the deal it belongs to, updates the record, and tracks what you owe and what you are owed. It is single seat by design: one install, one person, one mail identity, so a five person team is served today by five installs rather than one shared brain. It is an AI knowledge management layer for a drug development team rather than a research assistant, and the distinction matters more here than in most industries.
The two halves of drug development
Ask what AI does for a drug company and you get the first column. The second column is where a small company actually bleeds time.
| The science half | The operations half |
|---|---|
| Target and molecule work, screening, modeling | Who owes which answer to the partner, and by when |
| Assay data, analysis, figures | Which date moved, and what moved downstream of it |
| Literature and prior art | What the last review decided, and why it was decided that way |
| Study design and interpretation | The notice period buried in the middle of a vendor email |
The right column has no category name, which is exactly why it has no category of software. It gets absorbed by whoever is least busy, and it is the first thing to fail when a week goes sideways. The operations half of drug development is the page that argues this at length, and it is the argument the rest of this section rests on.
What is on these pages
By function, because that is how the work actually divides. AI for business development in biotech covers the partnering desk: diligence question lists with dates on them, term sheet versions, the comp you quoted from memory, the partner who went quiet. AI for program management in drug development covers the integrated timeline, the risk log nobody revisits, the vendor deliverable that moves the critical path, and the action item from last month's review that quietly died.
Three pages sit deliberately at the edge of what this can do. What a biotech CSO actually needs separates research work from the decision record around it. AI agents for drug discovery says where they stop, and where this stops, which is much earlier. AI agents for pharma takes the pharmacovigilance question head on rather than letting a searcher arrive and find out on their own.
Three more cover regulated territory from the outside: clinical operations at a small biotech, regulatory affairs, and medical affairs. In each case the honest position is the same. A small company can use a system like this for the coordination around regulated work. It cannot use it for the regulated work itself.
Where your data actually sits. Your records are plain text files on your own machine, which is an ownership and portability property. It is not a data residency one. The system runs on Claude, so the model does not run on your laptop, and every document, thread and note it reads goes over the API to the model provider. Anything you would not put through a hosted model should be scoped out at install, and scoping it out costs nothing.
If your problem is that the science is slow, none of this helps. If your problem is that nobody can say, on a Tuesday, what the current answer is to a question a partner asked three weeks ago, keep reading. That is a different problem, and it is the one worth deciding whether to build or buy a solution for.
The half nobody sells software for
Software exists for it now, and the demo is a walkthrough of it: mail and meeting transcripts read as they arrive, filed to the program or the deal, with the owed answers, the moved dates and the notice period buried in a vendor email all still visible. Nothing in it touches the science column, so go and look at the other one.
See the demo