An AI agent in production made a decision about a customer last week. It approved something, or declined something, or routed something to a person who is now dealing with the consequence. Inside your product operating model, every other capability that makes a decision of that kind has a named product manager. There is a person whose job is to be accountable for what the capability does and what the customer ends up with. The agent does not have one.
That is the gap most agentic AI governance frameworks were written to close, and it is the gap they keep missing. The question executives are being asked to answer is what is our agentic AI governance framework? The right question is the sharper one underneath it: who is the product manager for what this agent decides?
In Brief
- Agentic AI is a new product capability and most deployments have shipped it with no product manager named.
- The decisions an agent makes about a customer are product decisions — but they sit outside the product operating model.
- A governance framework that lives next to the product operating model leaves accountability for outcomes orphaned.
- The product operating model already holds the decision rights, owner, and governance cadence the agent needs. Nobody extended it.
- The hard part is cadence: an agent decides faster than a product manager can inspect, and the operating model must adapt.
The agent is a product capability nobody owns
Agentic systems are not a category of IT infrastructure. They are a capability that has been added to a product or service, and they make customer-affecting or citizen-affecting decisions inside it. That is what makes them governable through a product operating model and not through the existing IT control stack. McKinsey’s framing of the shift is exact: agency is not a feature — it is a transfer of decision rights, and clear accountability for agent behaviour — covering business outcomes, risk, and policy — is the condition for scale, with business domains owning day-to-day governance of agent-enabled workflows (McKinsey & Company, 2026a).
That framing matters because it changes who the answer belongs to. In a working product operating model, every product capability has a named product manager who is accountable for the value it delivers and the decisions it makes inside its scope. When something behaves outside that scope, the named product manager is the human who answers for it. Apply the same test to the agentic systems being deployed today and the pattern is consistent: the agent has been added to the product, it is making decisions in production, and no product manager has been named to own what those decisions produce. The governance framework written alongside it sits in the risk or technology layer of the organisation, not in the product ownership layer where scope, escalation, and outcome decisions actually get made.
The consequence is visible in the safety data. McKinsey’s research into agentic deployment finds that 80% of organisations have already encountered risky behaviour from their AI agents — including improper data exposure and unauthorised access — with the underlying issue being that the agent did what it was optimised to do, in the absence of an owner who had defined what it should and should not do (McKinsey & Company, 2025a). That is not a model accuracy failure. It is a product ownership failure. Nobody had decided what the agent’s scope was at the level of detail a product manager would, because nobody was the product manager.
The agent has been added to the product, it is making decisions, and no product manager has been named for what those decisions produce.
Agentic AI entered through IT, not the product operating model
The reason capable executives reach for a separate governance framework rather than extending the product operating model is structural, not careless. Agentic AI arrived inside organisations through the technology function — through procurement of a platform, integration with an existing system, or a pilot run by the data and AI team. The product operating model, where it exists, governs the capabilities the business already understood it was building. An agent that surfaces through the tech stack lands outside the product manager’s working scope by default. By the time it is making decisions in production, the question of who owns its decisions has either been deferred to a governance committee or absorbed back into the IT control layer that procured the platform.
That deferral is what produces the framework. The instinct is to specify the controls, the escalation paths, and the audit mechanisms in a separate document, ratified by a separate group, because the operating model that should hold them is not built to receive them. RAND’s analysis of why AI projects fail is consistent with this: more than 80% fail, and the dominant cause is not the model itself — it is leadership, and a misidentification of the problem the AI is being asked to solve (Ryseff et al., 2024). When the problem is identified as a governance problem, the response is to write a governance framework. When the problem is identified as a product ownership problem, the response is to name the product manager. Most organisations are responding at the first level.
A separate body of work supports the same shift. The BCG and MIT Sloan global executive survey found 69% of expert respondents agreed agentic AI represents a paradigm shift requiring reimagined management — the agent treated as a new kind of team member inside the existing management framework, not as an exotic risk requiring a parallel structure (Renieris et al., 2025). The agent has joined the product team. It needs the same named ownership every other capability on that team already has.
The product operating model already holds the machinery agentic AI governance needs
The contrarian read on agentic AI governance is that the executive does not need a bolt-on framework to manage it. The product operating model already holds the machinery the agent needs. There is a named owner who is accountable for the value the capability delivers. There are decision rights about what the capability is in scope to do. There is an escalation path when it behaves outside scope, and there is a governance cadence in which outcomes get inspected against intent. The failure is not that any of this is missing. It is that nobody has extended it to treat the agent as a governed part of the product.
That extension is mostly conceptual rather than infrastructural. ZXM’s Product Operating Model work inside a national security agency surfaced the same pattern: the highest-leverage redesign was not a new governance body but the deliberate inclusion of a previously-unowned capability inside the existing product ownership model. Once a capability sits inside the product manager’s scope, the questions of scope, escalation, and outcome accountability get answered by mechanisms that already exist and that the organisation already trusts. The same logic applies to an agent. Treating it as a product capability with a named product manager — one who has been told what the agent is permitted to decide, what conditions trigger an escalation, and how often the decisions get inspected — uses the operating model the organisation already runs. A governance framework that lives next to that operating model leaves the agent’s decisions structurally orphaned, no matter how comprehensive the framework looks on paper.
The implication compounds with each new agent. A single pilot agent with an orphan accountability can be absorbed at the executive level. Five, fifteen, fifty agents — which is the trajectory the market is currently on — multiplies the gap by the number of agents. The product operating model is the only structure that scales with that growth, because every new capability gets owned the same way. A framework that sits next to the operating model does not scale; it accumulates exceptions.
Cadence is where agentic AI governance gets genuinely hard
The honest limit on this reframe is cadence. A product manager normally inspects the capability they own at a cadence appropriate to the rate at which decisions accumulate — weekly, fortnightly, monthly, with quarterly reviews against value. An agent in production makes thousands of decisions in the gap between two inspection cycles, and many of those decisions are individually consequential. The cadence the product operating model was built around does not match the cadence the agent operates at. That is the design problem the executive now holds: extending the product operating model to cover the agent requires adapting the cadence inside it, not just re-pointing the existing one.
This is the thread ZXM’s paper on AI governance design picks up. What the product operating model has to become when the capabilities it governs make decisions faster than the people who own them can review them is not a small redesign. It is the question that determines whether the operating model continues to govern in any meaningful sense, or whether the agent quietly takes over the parts of the decision the model used to hold.
For now, the executive question to hold is the smaller one and the harder one: not what is our agentic AI governance framework? but who is the product manager for what this agent is deciding, today, in production? The framework follows from that answer. It does not produce it.
References
- McKinsey & Company. (2025a, October). Deploying agentic AI with safety and security: A playbook for technology leaders. McKinsey & Company. https://www.mckinsey.com/capabilities/risk-and-resilience/our-insights/deploying-agentic-ai-with-safety-and-security-a-playbook-for-technology-leaders
- McKinsey & Company. (2026a). The agentic organization: A new operating model for AI. McKinsey & Company. https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/the-agentic-organization-contours-of-the-next-paradigm-for-the-ai-era
- Renieris, E. M., Kiron, D., Mills, S., & Kleppe, A. (2025, September 16). Agentic AI at scale: Redefining management for a superhuman workforce. MIT Sloan Management Review. https://sloanreview.mit.edu/article/agentic-ai-at-scale-redefining-management-for-a-superhuman-workforce/
- Ryseff, J., De Bruhl, B., & Newberry, S. J. (2024). The root causes of failure for artificial intelligence projects and how they can succeed: Avoiding the anti-patterns of AI (RR-A2680-1). RAND Corporation. https://www.rand.org/pubs/research_reports/RRA2680-1.html
