Note to Biopharma: Agentic Digitization Is Not Scale

Executives are spending eight figures per plant on agents that don’t replicate. The problem isn’t the technology. It’s the diagnosis

A major biopharmaceutical company paid a major software and AI firm in the high eight figures to build an agent that automated one manufacturing plant. The solution worked. It also did not port. When the customer asked to extend it to other sites, the vendor quoted similar sums for each. The customer burned capital without reshaping its network. The vendor forfeited the margin, leverage, and credibility a genuinely scalable product generates. Both sides lost.

This is not an isolated failure. It is the predictable result of a misdiagnosis that runs through biopharma’s agentic AI programs.

The Misdiagnosis

Biopharma leaders treat agentic AI as a digitization problem. It is not. Digitization is not scale. Executives prize new capabilities but rarely apply first-principles thinking to what scale actually requires, assuming that a new digital capability — an agent — will automatically solve the problem. It won’t.

Two habits deepen the error. First, biopharma companies seldom document their processes in detail, and when they do, they treat the resulting maps as immutable world models. Rube Goldberg workflows get digitized wholesale rather than redesigned. Agents are bolted onto bad processes; the capability is real, but so is the fragility it inherits. Second, executives default to the wrong success criterion. If the problem is “we haven’t digitized enough,” the answer is more agents. If the problem is “we lack speed and scale,” the answer looks entirely different. Most programs are built against the first diagnosis and should have been built against the second.

The implication: agentification and scale are independent. Non-technological solutions that deliver speed and scale are wins, but biopharma ignores them because they lack the glitter of “new capability.”

Five Hallmarks of Scalable Software

Test any agentic-AI program against the five properties software engineers design in when something needs to scale.

Abstraction. A scalable system exposes a common interface over messy specifics. In semiconductors, this is the Process Design Kit (PDK). A fabless designer writes to a PDK; any compatible foundry can manufacture the design. The PDK hides physical differences. Biopharma has no PDK equivalent. Every plant is a snowflake, so an agent built for Plant A is dead on arrival at Plant B.

Modularity. A scalable agent is assembled from reusable pieces — scheduler, deviation investigator, compliance checker — that can be swapped or upgraded without rewriting the whole. In a Rube Goldberg process, any change detonates the system.

Statelessness. Each task carries the context it needs, does the work, and finishes. No hidden memory across runs. This matters under GxP, where each run must be validateable as a unit, and for scale, since stateless tasks run in parallel without fighting for shared state.

Horizontal scale. Capacity grows by adding instances, not by rebuilding. Cost-to-add should approach zero. When a vendor quotes another eight figures for the second plant, the price itself tells you the system does not scale horizontally. The vendor is replicating labor, not software.

Observability. Every agent action is captured, timestamped, and traceable. This is already mandatory under 21 CFR Part 11. At scale, it becomes existential: a system you cannot see cannot be improved, validated, or recovered — and certainly cannot run across ten plants.

The eight-figure single-plant agent failed on all five. The vendor’s quote for plant two is the tell. With these properties in place, a second plant is cents on the dollar. Without them, it is another eight-figure invoice.

The Fix

The fix is not more agents. It is what semiconductors did in the 1980s when fabless manufacturing became possible: build an abstraction layer that makes every site look the same to whatever runs on top. Once that layer exists, an agent written once runs everywhere. Without it, every plant remains a custom engagement.

The pattern is structural, not plant-specific. Medical writing is another domain where biopharma is pursuing agentification without scale. The “plants” are document types — clinical study reports, protocols, investigator’s brochures, regulatory submissions, labeling — multiplied by therapeutic area. The abstraction question becomes: what is the PDK for medical writing? The failure mode looks different too: instead of eight figures per plant, it is an agent trained on oncology regulatory documents that doesn’t transfer to cardiovascular ones, or a stack of per-template, per-region, per-sponsor customizations that never consolidate.

Until executives ask explicitly what their agents will cost to replicate, and test the answer against all five hallmarks, the bills will keep coming — one plant, one document type, one therapeutic area at a time.