Hybrid AI in pharma: what goes private, what goes managed

Pharma and biotech teams now use AI every day, whether or not it is approved. Scientists paste compound tables into chatbots, writers ask for help with a CSR paragraph, and safety teams try assistants on case narratives. The instinct of many IT and QA teams is to pick one answer for all of it: either everything goes to a managed model under an enterprise contract, or nothing leaves the building.

Both answers leave value on the table. A hybrid approach routes each task to the model that fits it, based on the data the task touches.

Start with the data, not the model

The routing question is simple to state: does this request touch confidential or regulated data? In pharma, that covers more than patient data. It includes:

• Patient-level trial data, listings and investigator notes.
• Adverse event reports and the medical history inside them.
• Compound structures, assay results and structure-activity data.
• Unpublished study results, draft submissions and agency correspondence.
• Invention disclosures and draft patent claims.

Anything on that list belongs on private models running inside your own cloud account. The reasons differ by item: data protection and patient confidentiality for trial and safety data, trade secrets for compounds, and patentability for disclosures, where an uncontrolled disclosure can create problems of its own.

What can safely use a managed model

A large share of scientific work involves only public material. Summarising newly published papers, reading a draft guideline from a health authority, tracking competitor pipeline news or conference abstracts: none of this needs your data. Here, the strongest managed models add real value, and paying per token for this share of the work is reasonable.

The useful pattern is to split a task in two. A managed model summarises the public text; a private model maps that summary against your own protocols or SOPs. The literature summary benefits from frontier capability, and the mapping never leaves your account.

The policy check is the product

Routing only works if it happens before any model sees the request, and if the default is private. In practice that means a policy check that classifies each request, a rule that uncertain cases stay private, and a per-task switch your organisation controls. Some teams switch managed AI off entirely for certain programmes; CROs often need a different setting for each sponsor.

When in doubt, it stays private. That single rule does more for trust than any contract clause.

Right-sized models inside the private route

Private does not mean one large model for everything. Most routine work, such as case intake checks or document classification, can be handled by small language models that are fast and cheap to run. Larger open models are kept for complex drafting and analysis. That split is what makes private AI affordable at pharma volumes.

Keep the expert in the loop

Whatever the route, every output is a draft for a qualified expert. The answer should separate observation from interpretation, state its confidence and cite the SOP or source it relied on, so the reviewer can check it quickly. The approval, the sources and the route are logged, which is what QA and inspectors will ask to see.

A practical starting point

List your ten most common AI requests. Mark each one private or managed based on the data it touches, and note where a task can be split. That list becomes your first routing policy, and usually shows that most of the valuable work sits on the private side.