Claude Code for biotech vs off-the-shelf AI tools

Most of what is sold as an AI tool cannot see your mail, cannot write to a record, and forgets you between sessions. The question is whether you need something that can, and what you give up if you do not.

Off the shelf covers two quite different things. One is a chat assistant: you paste in context, you get an answer, and nothing about your company changes. The other is a category product with AI features bolted on, which knows its own database and nothing else. Both are useful. Neither of them keeps a record current, because neither has permission to read what arrives or to write what changed.

The same engine, packaged three ways

Anthropic ships Claude in three shapes and says plainly that two of them share an engine. Its documentation describes Cowork as using the same agentic architecture that powers Claude Code, available inside the desktop app without opening a terminal. So the difference between them is not raw capability. It is containment, and it is what each one is allowed to reach.

The two differences that decide it

The first is containment. Cowork is limited to the folders you hand it, which is genuinely attractive if the thing stopping you is nervousness about what an agent can reach. Code runs with full access to the machine, which is more power and more responsibility, and it is the reason a system built on it can pull from mail, calendar and a document folder in the same pass.

The second is smaller in the documentation and larger in practice. Hooks, the mechanism that lets you enforce a rule on every single run rather than hoping the model remembers, are listed for Code and not for Cowork. Everything that makes an agent safe to leave running unsupervised, the gates, the checks that fire before an action, the refusal to state something as settled without evidence, is built out of that mechanism. There is also a practical wrinkle worth knowing before you plan a migration: Cowork loads the skills and plugins enabled for your account and does not read the Claude Code directory on your machine, so anything installed there has to be added again through its own customize surface.

What that means for a decision is fairly clean. If you want an assistant that finishes documents from a folder you point it at, the contained option is simpler and safer. If you want a system that reads what arrives, updates records without being asked, and enforces its own rules every time it runs, that needs the enforcement surface. The version described here ships as an app, so you get that shape without assembling it in a terminal. One seat on one machine, and drafting rather than sending, are the price of the shape rather than a limitation of the platform.

Why this lands differently in a biotech

The work that goes wrong in a small biotech is not a document that needs writing. It is a chain that needs walking. A vendor writes that a peer read slipped, and the consequence runs from the draft report to the nonclinical summary to the filing date to the float in front of first patient dosed. Nobody paid to think about the science should be doing that arithmetic by hand at five o'clock, and an assistant that only sees what you paste into it will never see the email that started it.

The second reason is the one every founder raises. Sponsor material under confidentiality, unpublished results and study designs that belong to somebody else are in that mailbox. A design where the files stay on your own machine, in plain text you own, is easier to reason about than one where a vendor holds your working record. It is worth being exact rather than reassuring: the files stay local, and the model does not. What the system reads is sent to the model provider over the interface, and there is no local model option here. Anyone who tells you otherwise is selling you comfort. For the same reason patient data and anything that has to sit in a validated system are scoped out at install rather than carefully fenced.

When off the shelf is the right answer

When you want answers rather than upkeep. When nothing about your week depends on a record being true tomorrow. When your company has an information security review that will not approve a program reading a mailbox, in which case the review is the constraint and no amount of architecture argument moves it. Those are real, and the full set of trade-offs, including who owns the thing in two years, is on internal AI tool vs vendor for a thirty person biotech.

What tips it for most people is that the packaged version is not a project. It is an app, roughly thirty minutes of real work to install, and it can be done with you rather than by you. It is in limited early access, so getting a copy starts with a conversation. Building the same behavior yourself is the part that takes months, and then keeps taking them.

Setting the same platform up as a working assistant is covered in Claude as an AI chief of staff in a biotech, the platform-level comparison is Claude for biotech, and the concept underneath all of it is an AI company brain. The category-by-category decision map is build vs buy software for a biotech.