A small biotech rarely lacks ambition. It lacks slack.
The same people who carry the science are also preparing board materials, resolving vendor questions, tracking regulatory changes, rebuilding forecasts, searching prior decisions, and translating one function’s context for another. The resulting constraint is not simply headcount. It is the amount of expert attention the operating system consumes.
That distinction matters. A 50-person company should not imitate the org chart of a 500-person company. It should build a different mechanism for moving knowledge, decisions, and accountable work.
AI can be part of that mechanism. But only if the design begins with the operating model—not a catalog of tools.
Start with the attention bottleneck
The useful question is not, “Where can we add AI?” It is:
Where is scarce judgment being consumed by work that does not itself require scarce judgment?
That question separates four kinds of work:
- Judgment — interpreting evidence, accepting risk, choosing a path.
- Synthesis — gathering and reconciling context so judgment is possible.
- Production — drafting, formatting, comparing, routing, and reporting.
- Coordination — moving work among people, systems, deadlines, and approvals.
In consequential functions, judgment remains human. The leverage opportunity sits mainly in synthesis, production, and coordination. AI can prepare the field; the accountable expert still decides what crosses it.
Build five connected layers
Point solutions create local speed. An operating system creates institutional leverage. The difference is whether the pieces reinforce one another.
1. A governed knowledge layer
Protocols, agency correspondence, study reports, contracts, meeting decisions, and internal policies are useful only when their authority and provenance remain visible. Retrieval must preserve the source, version, owner, and access boundary—not merely return plausible text.
2. Decision-ready synthesis
Teams do not need more summaries. They need briefs that expose the decision, the evidence, the uncertainty, and the unresolved questions. A good AI-assisted brief accelerates expert review without hiding the reasoning path.
3. Workflow-level assistance
The system should enter work where people already operate: preparing a cross-reference table, comparing a new guidance document with internal procedure, assembling a first-pass site review, or surfacing changes before a portfolio meeting. Every workflow needs a named owner, review point, and exception path.
4. An adoption system
Training detached from real work fades quickly. Teams learn by reshaping a live process, observing where the system helps or fails, and turning that experience into a reusable operating pattern.
5. Continuous operation
Models, source documents, permissions, and company priorities change. A capability that is not monitored will drift. Usage, quality, exceptions, and business signals must be reviewed as operating data—not as a one-time implementation report.
Choose the first move by consequence and learnability
The loudest pain is not always the right starting point. A first workflow should combine:
- enough value that leadership cares;
- enough repetition to produce evidence;
- bounded consequences if the system is wrong;
- available source material;
- an engaged process owner; and
- a result that can be measured without inventing a heroic ROI model.
An internal regulatory-intelligence brief may be a better first move than drafting a submission section. A portfolio-meeting pack may teach more than a broad “enterprise copilot.” The first implementation should reveal how the organization will govern, adopt, and improve AI—not merely prove that a model can generate prose.
Measure the mechanism before claiming the outcome
Without a clean baseline, aggressive outcome claims are theatre. Begin with operating signals:
- elapsed cycle time;
- expert review time;
- rework and exception rate;
- source coverage and citation accuracy;
- handoff delays;
- adoption by the intended roles; and
- the share of outputs that require material correction.
Only after those signals stabilize should leadership connect them to program timelines, avoided cost, or capacity. This is slower than announcing a percentage in a slide deck. It is also how a defensible business case is built.
What the small company should keep
Leverage is not the same as imitation. A small biotech’s advantage is the short distance between the person who sees the signal and the person who can act. AI should preserve that advantage.
Do not bury a compact organization under a new central bureaucracy. Establish a small operating nucleus that owns standards, shared architecture, measurement, and vendor choices. Let functional experts own the actual workflows. Centralize the guardrails; distribute the judgment.
The operating-leverage test
A credible AI operating model should let a leadership team answer five questions:
- Which decisions are we trying to improve?
- Which human remains accountable for each one?
- What evidence is the system allowed to use?
- Where can a failure be detected and escalated?
- What signal will tell us the new path is genuinely better?
If the answers are vague, the organization has a collection of experiments. If they are explicit, it is beginning to build a capability.
The goal is not to make 50 people look like 500. It is to let 50 people remain close to the science while the operating system carries more of the friction around them.