Agentic AI in Life Sciences Manufacturing

By 2027, every life sciences shop floor will be running AI agents that nobody approved. We've made this mistake before, with ERP, and the bill this time is bigger.

I predict that when you walk a life sciences shop floor in 2027, you'll find the same thing you'd find on a handful of shop floors today: at least one AI agent nobody formally decided to buy.

Ask the engineering team if that's even possible right now, and they'll tell you no. They're not wrong about today. They're just not the ones who get to decide. AI will be embedded in almost all the applications running on that floor, and it'll likely be there in an upcoming upgrade, whether anyone asked for it or not; it's the vendor's product, not theirs, so the vendor chooses.

Every business unit is doing the sensible thing for itself in the meantime. Quality gets an AI-driven smarter deviation workflow inside its QMS. Manufacturing gets an AI-driven smarter scheduler inside its MES. Nobody's wrong at the level they're operating. The problem only shows up a level above, where nobody's looking.

We've paid for this mistake before

I've stood in enough steering committees to know how this story ends, because I've watched it end before, with major platform solutions such as ERP and MES. Multiple sites, multiple business units, each making a locally sensible decision about which instance to run and how to configure it. Nobody was reckless. Everybody optimized for their own plant, their own quarter, their own go-live date. Years later, the company was writing a multi-hundred-million-dollar cheque to consolidate instances that should never have diverged in the first place.

Agentic AI is running the identical script, just faster and quieter. ERP sprawl at least announced itself; a new instance meant a capital request, a steering committee, a line in someone's budget. An embedded AI feature announces itself with a changelog nobody reads. The sprawl this time won't appear as 5 ERP instances. It'll show up as a quality system that can't see what the maintenance system already knows, a scheduling agent that can't see what the quality system is investigating and five vendors each holding a slice of a decision nobody can reconstruct.

Why this bill is bigger than the ERP one

ERP sprawl costs money and time. This costs money, time and, because we're talking about life sciences manufacturing, not manufacturing in general, something ERP sprawl never touched: the ability to show a regulator exactly how a decision got made.

A deviation investigation today already has to pull from the batch record, the environmental monitoring system, the equipment history and the lab, four systems a human currently probably reconciles by hand. Hand that reconciliation to an agent embedded inside just one of those four systems, and it can only ever see its own slice. Either the cross-system value never gets captured, or someone quietly grants that agent reach into the other three without anyone owning the access decision, the audit trail or the answer to "who approved this."

What happens if nothing changes

Left alone, this doesn't stay static; it compounds:

• Every business unit separately audits its own vendor's AI development practices, at a fraction of the leverage, and a multiple of the cost, of doing it once, centrally.
• Redundant AI subscriptions and compute spend pile up across sites and functions, with no one positioned to see the total, let alone negotiate it down.
• The cross-system use cases, the ones actually worth the investment, never get built, because no single vendor's embedded agent can see far enough to build them.
• Eventually, either a consolidation program gets funded to unwind years of uncoordinated purchases, or a regulator asks a question about an agent-influenced decision that nobody in the room can fully answer.

The reframe

This was never an AI budget line. It's an architecture decision that happens to involve AI.

Not which agent to buy, but who owns the plumbing between them. 

Not whether the feature is smart, whether anyone can trace what it did and why. 

Not whether Quality or Manufacturing gets there first, whether either of them can get there without the other.

What to do about it now

Split the decision the way the risk actually splits, not the way the org chart currently splits.

Treat architecture, infrastructure and governance as company-wide, spanning R&D, manufacturing and quality: the data layer every agent draws from, the security and audit model every agent operates under and the supplier qualification program every AI vendor gets measured against. That's the part of a life sciences organization that behaves like ERP: get it wrong locally, and you pay centrally, later, at a multiple.

Leave the agents themselves, which use case, which workflow, which vendor from an approved list, to be owned by the business unit that runs the process. That's where the judgment actually lives, and it's not a decision worth taking away from the people closest to the work.

And decide, now, before the next vendor update quietly switches something on, whether your AI governance sits underneath the authority your Quality function already holds over GxP systems, or stands up as a second authority next to it. Get that wrong, and you won't have solved the silo problem; you'll have built a new one, with better branding.

The choice facing life sciences manufacturers right now isn't whether to adopt agentic AI. Every application vendor is making that choice for you, one upgrade at a time. The choice still in your hands is whether you find out what you've adopted before the regulator does, or after.