Skip to main content

The CRO Head of IT Lens: What Regulated Systems Teach About AI Governance

· 5 min read

A method that has been doing this job for thirty years​

Long before "AI governance" was a board agenda item, there was a room where the same problem had already been solved, and paid for, and audited, and refined into doctrine. It was called computerized system validation. In pharmaceutical and medical-device companies, it is the discipline that answers one question to the satisfaction of the FDA and the EMA: how do you know this computer system does what it claims to do, and how do you prove it after the people who built it have gone home?

The answer, distilled: you never trust the claim, and you never trust the builder's confidence. You qualify in layers.

Installation. Is the system installed and configured as specified? Not "is it running." Is it configured as documented, in the environment it will actually live in.

Operation. Does every function do what its specification says, under test, one by one? Traceable: every requirement has a test, every test has a result, every result has a signature and a date.

Performance. Does the whole thing hold under real load and real conditions? Not the demo. The worst Tuesday.

And around the layers, the rule that makes the whole thing survive audits: the record is the system. If it is not documented, it did not happen. If the documentation was written after the fact and back-dated, that is a fraud finding, not a process improvement. Data integrity is not a feature. It is the floor.

Why the regulated world grew this​

I spent more than ten years inside this world, on the accountable side: the Head of IT who signs, not the consultant who advises. FDA-grade and EMA-grade scrutiny was the weather, not the project. And when you live in that weather long enough, something rewires in how you look at every computer system, everywhere.

You stop asking "is it good?" You start asking: good according to which requirement, tested against which evidence, recorded where, readable by whom in five years?

You learn that the dangerous systems are never the obviously broken ones. They are the systems everyone believes in. Belief is not a qualification protocol.

You learn the most expensive sentence in enterprise IT: "it works on my machine." The regulated world has a translation for it: unverified. Everything else, you can work with.

The migration problem nobody scheduled​

Here is the thing that makes this lens valuable far beyond pharma: almost every company inherited the problems that computerized system validation was invented to survive, without inheriting the method.

AI adoption today looks, under the hood, exactly like a regulated company adopting unvalidated systems in 1995. Budget approved before the verification model exists. Vendors making confident claims in slide decks. Shadow deployments under people's desks. Audit pressure arriving from regulators, insurers, and enterprise customers on three different clocks, none of which the organization can answer. The suitcases changed. The unpacking problem is identical.

The difference is what companies do about it. The pharma company of thirty years ago had doctrine: qualification layers, traceability matrices, change control, the data integrity ALCOA rules. The typical company today has... a chat policy and a pilot committee.

The transferable parts are not the paperwork. You do not need GAMP forms to run a mid-size enterprise. You need the skeleton:

  1. Requirements written before the system is chosen, in terms the business owns.
  2. Qualification layers: installed as specified, each function tested, whole system load-tested.
  3. Traceability: every requirement to a test to a result to a signature to a date.
  4. Change control: the system that passed the gate yesterday is only trusted today if the diff between yesterday and today is known.
  5. The record, kept like it matters, because the day it is questioned, it is the only thing that ever was.

Put that skeleton under your AI stack and the EU AI Act stops being a mountain. It becomes a checklist written by people who are about to ask the questions you already answered. Put it under NIS2 and the same. Regulations of this family are made of the same atoms: prove it, record it, keep it working.

The skeptic's corner, honestly​

Two objections deserve honest answers, because this post is not a sales pitch for turning every company into a pharma plant.

"That is overkill for us." The full ceremony is. The skeleton is not. Every organization has a system where an unverified failure costs a career, a customer, or a headline. That system deserves the skeleton. Not everything. Triage is part of the method too; regulated companies qualify by criticality class, not by enthusiasm.

"We cannot hire a validation team." Correct, and not the point. The point is that the method exists, is public, is proven under the hardest auditors on earth, and fits into an engagement when a trigger arrives. You do not need the department. You need the discipline at the moment it counts, from somebody who has carried the signature before.

That is what I do: the regulated world's hardest room, applied as a lens to the systems your board is now approving without one. The CRO in the name is not decoration. It is where the method came from, and where the scars came from, and both travel.


If your AI stack, your security posture, or your handover has no verification spine under it, that is a trigger. Describe your situation and you will hear quickly whether it is a mandate for this office.