AI for drug development: the operations half nobody automates
Everything written under this heading is about the science: targets, molecules, trial design, literature. Meanwhile a drug program is also a machine made of threads, promises, decisions and obligations, and that machine is run by hand.
Read the field and you would think a drug company is a laboratory with a finance function attached. Work in one and you find something else. A program is a standing negotiation between eight or nine parties, most of whom you do not employ, conducted almost entirely in email, and the current state of that negotiation is not written down anywhere complete. It is distributed across inboxes, call memories, and a spreadsheet that was accurate on the day it was last touched.
That is the operations half. It has no category name, which is why it has no category of software, which is why nobody automates it.
What the operations half is made of
Four object types, and every one of them is currently stored in a human being.
| Object | What it is | Where it lives today |
|---|---|---|
| Threads | An open exchange with a vendor, partner, adviser or agency that is not finished | Your inbox, and only your inbox |
| Promises | Something you said you would do, or somebody said they would send you, usually on a call | Memory, until it hurts |
| Decisions | What was settled, when, and on what basis, so it is not re-argued in six weeks | Minutes nobody rereads, or nowhere |
| Obligations | A notice window, a renewal, a milestone trigger, a reporting commitment | Mid paragraph in a contract or an email |
Notice what these have in common. None of them is data in the sense the industry uses the word. They are not assay results or trial endpoints. They are state: what is currently true about a set of relationships. They change constantly, they change by email, and there is no system of record for them at a company of thirty people.
Why nobody automates it
Because it does not look like a product. The enterprise version of AI in this industry is about scientific content: taking a large corpus of documents and making it interrogable, so that a question gets an answer with a traceable source behind it. That is a genuine problem and some serious companies are selling into it. It is also a different problem. It assumes the thing you are missing is retrieval over a body of knowledge that already exists in written form.
The operations half fails earlier than that. The thing you need was never written down. Nobody minuted the decision. Nobody recorded the promise. The master service agreement renewal, and the date after which the vendor's terms can no longer be changed, both arrived mid paragraph in an email about a delivery slot, so the first of October is a deadline no document anywhere carries. Retrieval cannot save you, because there is nothing to retrieve.
The distinction that decides whether any of this is for you. When a week goes sideways and you are heads down, what falls through? If the answer is analysis you did not get to, this is the wrong tool and probably the wrong page. If the answer is threads, follow-ups and who owes what, that is the whole target.
What a system for it actually does
It reads the mail and meeting transcripts that already exist, and it writes back. That second half is the part that matters and the part most tools skip. An email arrives from a tox vendor saying the draft report moves nine days. The system files it to the program, changes the date on the record, follows the change downstream into the review slot and the document assembly week and the invoice trigger, closes the item that email resolved, opens the ones it created, and drafts the reply. You approve or you discard. Nothing outbound leaves without a human saying yes. It runs for one person at a time, so a colleague's commitment becomes your follow-up rather than their task, and it is not a validated system: nothing it writes is a controlled record or goes into a submission, and patient data stays out of scope rather than fenced.
It holds the promise from a call three weeks ago as a dated open item and brings it back on the morning it starts to cost something. It watches your sent mail, so what you committed to going the other way is tracked the same way. It answers "what did we decide about that, and why" from your own files, with the source. It flags a notice window before it comes due rather than on the day.
Each of those is unremarkable on its own. Together they are the difference between a record that is current because someone maintained it and a record that is current because nobody had to. The functional versions of this argument are on the program management page and the business development page, and the general shape is AI knowledge management for a drug development team.
None of it is discovery, and the honest limits of agents in that space are the subject of a separate page. Whether the operations half is worth building or buying, given that no vendor sells exactly this, is the question on the project management software page and across the rest of this section.
All four object types, out of human memory
Threads, promises, decisions and obligations assembled from mail and meeting transcripts, nothing typed in: the renewal date that only ever appeared mid paragraph in a note about a delivery slot is carried as a dated obligation, and the promise made on a call three weeks ago returns the morning it starts to cost something. The demo is the quickest way to judge whether that is your half of the problem.
See the demo