Build vs buy software: the AI decision for a biotech

A twenty or thirty person biotech can buy the category tool, ask its one technical person to build something internally, or carry on running the company out of mail and memory. Here is the honest read on each, one category at a time.

The useful version of this question assumes the need is settled and asks only where the thing comes from: built internally, or bought from somebody. That frame is more uncomfortable than the usual build-or-buy piece, because it quietly removes the option most small biotechs actually take, which is neither.

Neither is the default for a good reason. A category tool looks like a line item you cannot justify at your headcount, and building looks like six weekends of the one person you cannot spare. So the company keeps running on threads, calendar invites and the founder's memory, and calls that a decision. It is not a decision, it is a deferral, and the bill arrives later as a missed notice window or a partner asking twice for something you already answered.

What buying a category tool actually gets you

It gets you a schema somebody else already argued about, a support contract, and an audit trail. Those are real. If your problem is a regulated document that has to survive an inspection, or a clinical data set that has to be locked, buy the system built for it. That answer does not change because AI got good, and two of the pages below reach exactly that conclusion.

What buying does not get you is a record that stays true. Every category tool in this space is a form. Somebody has to fill it in, and at your headcount that somebody is the person who least has the time. The tool then reports whatever was last typed into it, which is worse than no tool at all, because people believe it. Any honest internal AI tool versus vendor comparison has to start there: the failure mode of bought software at small scale is not cost, it is staleness.

What building it internally actually costs

Building is more achievable than it was. A capable person with a coding agent can stand up a genuinely useful tracker in a few evenings, and some companies should. The cost is not the first version, it is the second year: who maintains it after the person who built it moves on to the thing they were hired for. The five questions that settle that bet, starting with who still owns the thing in two years, are set out in internal AI tool vs vendor, and the platform question sits in Claude Code for biotech versus off-the-shelf tools.

There is a second cost that nobody scopes. An internal build normally captures state, and state is the easy half. The hard half is upkeep: reading the mail that arrived this morning, deciding which record it changes, changing it, and writing the next step down before the thread goes cold. That is the part that decides whether a system is alive in six months.

The categories, one at a time

The decision is not one decision. It is six or seven, and the answers genuinely differ.

The biotech-specific reason this matters

Take an ordinary week. A tox CRO writes to say the histopathology peer reviewer lost his slot, so the draft report moves out two weeks. That single email is not one update. It moves the nonclinical summary that feeds the module, which moves the filing date, which eats the float in front of first patient dosed, which is the date your site activation plan was built around. Nobody in a thirty person company is going to walk that chain by hand at four in the afternoon, so it gets walked in the review three weeks later, once the float is already gone.

That chain is why the answer here is not really about software categories. It is about whether the company has anything that reads what arrives and updates what it touches. That is the argument made in full on what an AI company brain is and why a biotech needs one, and it is the thing neither a bought form nor a weekend build tends to deliver.

It is also cheaper to try than either side of the build or buy argument suggests. The app installs in about thirty minutes, yourself or done with you if you would rather not do it alone, and it is in limited early access, so getting a copy starts with a conversation. That is a smaller commitment than a procurement cycle and a much smaller one than a build.

The boundary is worth stating before you read further. This is not drug discovery, it is not pharmacovigilance, it is not an electronic lab notebook, and it is not a validated system for regulated documents. It is the operations layer: the threads, the promises, the dates and the records that describe how the company is actually running.