Agentic AI — S1 · 5/12: Governing agents
Teams building their own agents without coordination. Processes fragmenting. Validation steps scattering.
AI is sold as a technology that breaks down silos. Without governance, it does exactly the opposite: it re-silos functions.
With an added difficulty: these partitions are hard to inventory. How many agents are actually running, who built them, which data they rely on, what happens when two of them contradict each other — few organisations have the answer today.
That is why governance has to be designed before deployment.
Four questions seem to me to come first.
Who owns the subject
An AI initiative does not progress durably without a senior voice that validates proposals, arbitrates priorities and sets the phasing. I hold that role for my own scope. What it actually costs is unpopular arbitration: saying no to a use case a team believed in, delaying a deployment because the data is not ready, or catching a use case that skipped who reviews the output. Without sponsorship, projects move forward — each in its own direction.
Who answers for what
Who designs the agents? Who validates them? Who supervises them? Who answers when something goes wrong? Without a written answer, agents become orphans: impossible to maintain, impossible to embed, and above all impossible to defend in front of an auditor.
Who needs to be in the room
Data, IT, quality, business. An agent designed without one of those four functions eventually collides with it — on data quality, technical integration, validation, or the reality of the process it is meant to serve. That is how a silo forms: not through bad intent, but because something that needed to connect was built alone.
What happens next
This is the most neglected point. An agent is not a static script: it evolves, it learns, it interacts with systems that change. In a regulated environment, that means versioning, activity logs, full traceability — and retirement criteria.
A system that evolves has to have its changes documented — what changed, when, and why.
That reflex already exists in our profession; it now has to be applied to objects that learn.
Governance is not a brake. It is what separates the organisation that takes pride in how many agents it has from the one that knows what these agents do — and why each of them is there.
How many agents are running in your organisation today — and who could tell you with certainty?