Two advisors have made their pitch this quarter. Both are persuasive. One offers an AI-specific risk framework with controls, red-teaming, audit trails, and model cards. The other proposes an AI centre of excellence: a coordination body, capability building, and a set of standards. Neither is priced against the accountability structure that would make either of them work: persistent teams with clear ownership of outcomes and the authority to decide.
In Brief
- AI governance and product operating model governance are the same design problem — clear accountability, defined decision boundaries, empirical quality standards, persistent teams.
- Only 21% of enterprises have mature agentic AI governance in place, while 74% plan moderate agentic AI use by 2027 (Deloitte, 2026).
- Frameworks bolted onto the existing operating model do not close the accountability gap they were bought to solve — the operating model itself is the accountability structure.
- The organisations that get AI value are the ones that had already redesigned around empirical, cross-functional teams — the redesign is the governance.
- Buying another AI framework without changing how the work is structured funds the appearance of governance while the underlying gap continues to widen.
The pattern that surfaces across executive conversations on AI governance in 2026 is a category error that runs deeper than framework selection. AI governance is being treated as a separate design problem from the operating model the organisation has been trying to shift for the last decade.
Two governance problems are actually one
Deloitte’s 2026 State of AI in the Enterprise found that 21% of respondents say their organisations have a mature governance model in place for agentic AI, while 74% expect to use AI agents at least moderately by 2027 (Deloitte, 2026). The gap between those two numbers is the problem. Adoption is running four years ahead of the governance conditions the same organisations say they need.
The typical response is to initiate a governance framework: an AI-specific risk register, a set of controls, a coordination layer. What that response misses is what the governance conditions actually are. Clear accountability for outcomes. Boundaries for what a team can decide without escalation. Empirical evidence used to accept or reject work. Persistent teams that own a product long enough to be answerable for it. Every one of those conditions belongs to the product operating model. They are what the previous decade of transformation programs was funded to build for human teams.
The structural equivalence becomes visible as soon as two independent lines of work on AI governance are read alongside the operating model literature. Alvarez-Telena and Diez-Fernandez’s 2026 three-ring architecture for governing agents in on-platform organisations arrives at the same accountability layering from the opposite direction (Alvarez-Telena & Diez-Fernandez, 2026). The immutable core sits with humans: mission, values, the right to abstain. Standards sit jointly with humans and system, including metrics, governance, and audit, where the system proposes and humans approve. Protocols sit with the autonomous layer for coordination parameters, thresholds, and tuning within bounds. Their principle is unambiguous: the closer to why, the more human control; the closer to how, the more system autonomy. That principle restates the product operating model’s own logic: strategic intent held by executives, product decisions held by empowered teams, execution decisions held by the people closest to the work.
Saini’s 2026 analysis in California Management Review reaches the equivalent conclusion by naming the operating model directly. The Agentic Operating Model is the enterprise design decision that makes agentic AI governable at all: cognitive, coordination, control, and governance layers, with each agent tied to a clear business owner and human-on-the-loop supervision replacing the older approval-before-action model (Saini, 2026). His observation that agentic systems must be treated as organisational actors rather than tools demonstrates the same principle as the previous decade’s argument that products must be treated as persistent entities rather than temporary project outputs.
If the equivalence holds, the organisations that already have the operating model conditions in place should be pulling ahead on AI value. BCG’s tier analysis is exactly that test. The organisations extracting real value from AI are the ones that redesigned processes around AI adoption, while those that layered AI onto processes designed for a different operating logic remain in the lower tiers (Apotheker et al., 2025). The tier gap is an operating model gap made visible by an AI question.
Two vocabularies hide the same design problem
The reason capable executives keep initiating parallel governance frameworks is that the two problems present differently. The product operating model gets discussed in the vocabulary of teams, backlogs, funding cycles, and portfolio design. AI governance gets discussed in the vocabulary of risk, model cards, ethics boards, and controls. Those vocabularies were built in different communities, they land in different parts of the organisation, and they are pitched by different advisory categories. The structural equivalence is only visible when the two are named alongside each other: both are fundamentally about who is accountable for what work, under what evidence, with what authority to decide. Almost no organisation is set up to see them alongside each other. The AI governance question typically lands with a Chief Data Officer or a Chief Risk Officer, while the operating model question sits with a Chief Operating Officer or a Chief Transformation Officer. The result is separate owners, separate vocabularies, and separate budgets applied to what is structurally a single design problem.
The equivalence runs deeper than terminology. The empirical conditions that make an AI-augmented team’s output trustable and correctable at the speed the technology operates draw on the same discipline the product operating model requires for human teams (Hodgson, 2026): transparency of what is being done, inspection at short intervals, and adaptation on the evidence. Those conditions apply to any system that produces value when the work is uncertain and the environment moves faster than an annual plan.
The frameworks are not wrong. They are insufficient, because they prescribe accountability and cannot create it — that is the operating model’s job, and it was never done.
Buying the framework grows the accountability gap
Initiating an AI-specific governance framework without changing the operating model beneath it creates the appearance of governance while the accountability problem remains unresolved. The framework produces artefacts, a policy, a register, a committee, a training program, that satisfy the audit question without changing the conditions that produced the problem. Six months in, adoption has outpaced the framework, and the executive who initiated it is being asked why the same accountability questions remain unanswered.
The more significant consequence is that the problem does not stabilise. Each agentic use case deployed without clear accountability, defined decision boundaries, and empirical quality standards creates additional risk that no one in the organisation is accountable for. The Deloitte and BCG evidence bases are consistent: the organisations that already had the operating model conditions in place are the ones extracting AI value, and the rest fall further behind each quarter. What is worsening is the gap between what the executive committee believes the organisation is governing and what agents are actually being deployed to do — a gap that widens every time a use case deploys without clear accountability.
The test for any AI governance proposal
The question to put to any AI governance proposal is whether the operating model beneath it can deliver what the framework assumes: persistent teams, clear accountability for outcomes, and the authority to decide within defined boundaries. Most frameworks are well-designed on their own terms. If accountabilities are still diffuse, if teams are still assembled per project rather than persistent to a product, if funding still moves on annual cycles while agents deploy in weekly ones, then the framework requires accountability conditions the operating model was never designed to provide.
The organisations that make agentic AI governable in 2027 will be the ones whose operating model already established the accountability conditions the framework needs to work. The executive who treats the two governance questions as a single design problem, funded once and owned once, is buying a coherent transformation of how the organisation makes decisions. Funding them separately produces two frameworks that will both stall at the same point — when it becomes clear the organisation still has diffuse accountability, project-based teams, and annual funding cycles.
References
- Alvarez-Telena, S., & Diez-Fernandez, M. (2026). The three-ring architecture: Governing agents in the era of on-platform organisations (arXiv:2606.07119).
- Apotheker, J., Duranton, S., Lukic, V., de Bellefonds, N., Iyer, S., Bouffault, O., & de Laubier, R. (2025). From potential to profit: Closing the AI impact gap. Boston Consulting Group.
- Deloitte. (2026). State of AI in the enterprise 2026: AI agents scaling faster than governance. Deloitte Insights.
- Hodgson, M. (2026). Empiricism at machine speed: Scrum as a governance framework for agentic AI teams. Scrum.org.
- Saini, S. (2026, March 20). Governing the agentic enterprise: A new operating model for autonomous AI at scale. California Management Review.
