An AI agent for biotech CEOs
Not how to set one up. What it is actually like to have one, in the week of somebody running a company with three vendors, one partner and a board.
The pitch for agents is usually about tasks. That is the wrong frame for a chief executive, who is not short of people to do tasks. What a chief executive at a small biotech is short of is a current picture. You know roughly where the program is, roughly what the partner is waiting on, roughly what the burn looks like. The gap between roughly and exactly is where the surprises live, and closing it currently costs a day of asking people.
A week, as it actually runs
Monday starts with a short read rather than an inbox. Overnight, the passes swept what arrived, filed it against the right program or partner, moved the dates that moved, and wrote one brief: what changed, what is due this week, what is now at risk, and anything in the news worth ten seconds. The inbox is still there. It is just no longer the first thing that decides your morning.
Tuesday you have a call with the contract manufacturer and you have not looked at that thread in five weeks. The prep is already written: who you are talking to, what was agreed last time, the two things they owe you, the one thing you owe them, and the date that is quietly at risk. You did none of the preparation and you walk in prepared, which is a different feeling from walking in having skimmed an email in the car park.
Wednesday a partner writes to say a joint decision from the last committee meeting is being interpreted differently by their team. You ask what was actually decided. The answer comes back with the minute, the date, and the message where you confirmed it, cited. That is a five-minute answer instead of a two-hour reconstruction, and it is the difference between holding a position and conceding one.
Thursday you are pulling a board update together. The numbers come out of the same files the operational tracker reads, so the version that goes to the board and the version your team works from cannot disagree. There is nothing to reconcile because there was never a second copy.
The four questions it answers well
| The question | What it answers from |
|---|---|
| What changed while I was heads-down? | Everything filed since you last looked, by program |
| Does this slip break anything we promised? | The vendor's email, the timeline, and the commitment you made |
| What did we actually agree with them? | The thread, the minute, and the clause, quoted |
| What is waiting on me, and who is waiting on us? | Both directions of every open promise, with dates |
The second question is the one that pays for the whole thing. In a drug development company almost every commitment you have made externally rests on a date somebody else controls: a study readout at a contract research organization, a method transfer at an analytical vendor, a manufacturing slot. When one of those moves, the honest answer to "what does this break" requires holding four documents in your head at once. That is a lookup for a system that already holds all four, and it is a bad afternoon for a person.
What it never does. It does not send. Every reply, update and summary is drafted and waits for you. It does not assign work to your team, because it runs for one person with one mail identity. And it does not go near the science: no discovery, no assay interpretation, no patient data.
Four it will not answer: whether the program is worth continuing, which of two candidates to take forward, whether the partner is negotiating in good faith, whether the person you are worried about is going to leave. Those are the actual job, and none of them are a retrieval problem. What changes is that you arrive at them with the facts already assembled instead of spending the week assembling facts and the last twenty minutes deciding. The limit underneath that is worth knowing too: if a fact never arrives in mail, a meeting or a document, nothing catches it, and a decision made in a corridor is invisible to this exactly as it is to everyone else.
What actually changes about the job
You stop being the integration layer. That is the whole claim. The role framing behind it, and the parts of a chief of staff's job that transfer and do not, is an AI chief of staff for a biotech founder. The mechanism that makes it safe to believe the answers, the checker that re-reads a conclusion against its source before it reaches you, is the Claude chief of staff agent.
If your week is dominated by partnering rather than by programs, the function-level version is AI for business development in biotech. If your first question is whether your team should build this instead of buying it, that is internal AI tool vs vendor for a 30-person biotech. If you are wondering what happens when more than one person wants a seat, read an AI coworker in Claude for a biotech team before you plan around it, and the underlying concept is what an AI company brain is.
A week that arrives already assembled
The walkthrough shows the working app doing the Monday read, the prep for a vendor call nobody has opened in weeks, and the cited answer to what a partner actually agreed. Watch the demo and judge the gap between roughly and exactly for yourself.
See the demo