Building an In-House AI Team: A 90-Day Blueprint for Biotech Ops Leaders

Most biotechs that decide to "build an AI team" spend their first year hiring people who cannot do the work, because the leaders making the hires do not yet know what the work is. This is a survivable but expensive mistake. The following is a 90-day plan that front-loads the decisions that matter and defers the ones that don't.

Days 1–30: hire the translator, not the modeler

Your first hire is not a machine learning engineer. It is a computational scientist with domain fluency — someone who has sat in an SAR review, or a DoE meeting, or a protocol design session, and can translate between the language of the science and the language of the model. Title varies (Principal Scientist, Head of Computational X). What matters is that they have shipped, in industry, under regulatory scrutiny.

This person's first job is to write a one-page inventory of the top ten decisions in your organization where AI could change the outcome — ranked by frequency, dollar impact, and data availability. Everything else follows from this document.

Days 31–60: infrastructure before models

The second hire is a data / MLOps engineer. Not glamorous. Absolutely essential. If your compounds live in three ELNs, your assay results in a fourth system, and your process data in a fifth, no model will save you. The engineer's job is to make the top-three data sources queryable in one place with lineage.

This is also the moment to make the buy-vs-build call on foundation models. In our view, the answer is buy for the great majority of biotechs — few organizations can out-train Isomorphic, Genesis, or the hyperscalers. Your edge is your proprietary data and your workflows, not your model weights.

Days 61–90: the applied hire, and the first shipped decision

Only now do you hire an applied ML scientist — and you hire them against a specific decision from the day-30 inventory, not against a generic job description. Their success criterion for the first 90 days is not "publish" or "train a model." It is: one decision in the organization is now measurably faster, cheaper, or better because of something they built.

What to stop pretending you can do

• Foundation model training. You cannot. Do not try.
• Regulatory-grade validation without dedicated QA/CS resource. Budget for this from day one or descope.
• Retaining ML talent on biotech comp bands. Either pay competitively or partner externally. The middle path — underpaying and hoping — ends in a 12-month churn cycle.

The team at 12 months

If the 90-day plan works, at the one-year mark you have: one translator/lead, one MLOps engineer, two applied scientists, and a defined partnership with one external vendor for a capability you have chosen not to build. That is the entire team. Anything bigger, this early, is a hiring problem, not a capability plan.

Sources

1. AI Risk Management Framework (AI RMF 1.0) — National Institute of Standards and Technology.
2. The FAIR Guiding Principles for scientific data management and stewardship — Wilkinson et al., Scientific Data, 2016.