Agentic AI — S1 · 7/12: The blind phase
Regulatory professionals are trained in caution, traceability and compliance. Those reflexes are what make us valuable. They are also exactly what makes us circumspect about AI.
What gets read as resistance to change is often professional competence doing its job.
As agents spread into regulatory work, I observe technological anxiety, fear of obsolescence, doubt about where expertise now sits. This is not a lack of competence: it is a normal response to a technology reshuffling the cards.
Underneath it all sits one question: what is my value when AI does what I did yesterday? It touches professional identity and deserves better than a reassuring answer.
The blind phase
You often read that AI means less control and more trust. It is not more trust. It is trust made compulsory.
When preparing a deliverable goes from weeks to hours, the intermediate checkpoints do not disappear by choice: there is no longer any duration for them to sit in. One control point is left — the last one. Either accept, or start again.
This is the problem our industry solved more than twenty years ago. In medicinal products as in medical devices, we never reduce quality to a check on the finished product. We build it in by design and through in-process controls.
The same principle applies here: control what goes in, what feeds the agent; reinstate intermediate controls, even if that means changing who performs them, by setting autonomy thresholds; and qualify whoever does the final review.
Control does not decrease: it concentrates. Whoever signs off at the end now carries what several checkpoints used to carry between them.
Human value shifts toward judgment, supervision and the design of controls. New roles are emerging: AI Supervisors, AI Orchestrators, AI Quality Reviewers, AI Governance Leads.
It’s an elevation, but one that has to be prepared for. Nobody becomes a supervisor simply by ceasing to execute.
Responsible leadership, here, is not about convincing. It is about accepting that grief cannot be decreed, naming what is lost, not demanding enthusiasm, and saying what you do not yet know.
Agentic AI is not a threat to people. But it is not a gift either. It redefines work around what humans do best — thinking, judging, deciding, connecting, understanding, creating, embodying — provided they are given the means to do it.
In your organization, who designs the controls once the agent is integrated in the chain?