Publication date: Friday, 30 October 2026
Author: Malik Carter, Chief Intelligence Architect
Deverout and Associates | Strategic Intelligence for Transformation Leaders
The next enterprise AI failure will often be organisational rather than technical: the agents can act, but the business cannot explain who coordinated them, who controlled them, or who owns the result.
The Signal This Week
AI systems are moving from individual assistants to networks of specialised agents. That shift changes the management problem. A single tool can be evaluated as an application. A network must be operated as a system of actors, hand-offs, permissions, controls, and decision rights.
The California Management Review’s proposed Agentic Operating Model describes four interdependent layers: Cognitive, Coordination, Control, and Governance. Its central argument is useful even where its wider examples require separate verification: individual agents may perform well in isolation while the overall system becomes unpredictable at scale. The gap is not simply a model-quality problem. It is an operating-model problem. [1]
The infrastructure question is widening at the same time. The OECD’s work on securing AI compute treats compute as a strategic capability shaped by access, skills, energy, supply chains, and cross-border co-operation. It recommends needs assessments and policy objectives before major public investment, rather than assuming that domestic hardware alone creates resilience. [2]
The common signal is this:
When intelligence becomes distributed, coordination becomes a leadership responsibility.
The Agent Is Becoming an Organisational Actor
An agent does not need to be a legal person to create an organisational accountability problem. It can have a role, receive an objective, access information, make recommendations, trigger actions, and hand work to another agent or person.
That means the old categories become less useful. “The system did it” is not an ownership model. Neither is “a human approved it” if the approver could not inspect the evidence, challenge the recommendation, or stop the action.
The practical question is not whether the agent is autonomous. It is whether the organisation has made the agent’s autonomy legible. Leaders need to know what the agent is allowed to do, which decisions it can influence, when it must escalate, and who carries the outcome when the chain fails.
Four Layers of the Operating Model
The Agentic Operating Model offers a disciplined way to examine the system. The layers should not be treated as a software stack that belongs only to technology teams. Each layer is an executive operating decision.
1. Cognitive: What intelligence is being delegated?
Specialised models and agents can be easier to evaluate and constrain than one general system with broad authority. The decision is not only which model is most capable. It is which capability is appropriate for the task, the evidence available, the risk involved, and the organisation’s ability to explain the output.
Delegating a narrow task to a narrow agent may create more components to manage, but it can also make responsibility more intelligible. Generality is not automatically maturity.
2. Coordination: How do agents work together?
The coordination layer governs routing, sequencing, negotiation, and hand-offs. This is where many organisations will discover the orchestration gap: decentralised software can move faster than the human structures responsible for setting boundaries.
A workflow may involve a retrieval agent, a reasoning agent, a compliance check, a scheduling service, and a human approver. If no one can see the complete chain, each component may appear acceptable while the combined outcome is not.
Coordination must therefore be designed. The organisation should define which agents may call one another, what information travels with the hand-off, which actions require independent confirmation, and where the chain stops.
3. Control: What prevents unacceptable action?
Static permissions are not enough when agents generate novel actions in changing environments. Controls need context. A transaction, recommendation, or external communication may be acceptable at one value or confidence level and unacceptable at another.
Useful controls include thresholds, behavioural baselines, separation of duties, explicit stop mechanisms, and exception routing. The goal is not to review every low-risk action manually. The goal is to make high-impact or uncertain actions difficult to complete invisibly.
A human-on-the-loop model can be stronger than a human-in-the-loop model when the boundaries are real. The human does not need to approve everything. The human needs enough visibility and authority to intervene when the system crosses a defined boundary.
4. Governance: Who owns the result?
Governance assigns the business owner, risk profile, decision rights, evidence requirements, and lifecycle responsibilities. Every production agent should have a named owner, a defined scope, a retirement path, and a record of material changes.
This is where the organisation answers the question that technology alone cannot answer: who is accountable when a coordinated system produces a consequential outcome?
Compute Is an Operating Choice
The OECD’s analysis is a useful corrective to infrastructure slogans. Resilience is not created by announcing a data centre. It depends on the relationship between compute access, energy, skills, supply chains, policy objectives, and actual organisational use.
The relevant question for an enterprise is not “Do we own hardware?” It is “Can we sustain the AI-enabled work that matters if access, price, geography, or supplier conditions change?”
That leads to a practical distinction between capacity and continuity. Capacity is the ability to run workloads today. Continuity is the ability to keep priority workloads running tomorrow under changed conditions. Cross-border collaboration, multiple providers, portable data, local skills, and tested fallback procedures may contribute more to continuity than ownership of a single facility.
The Orchestration Readiness Test
| Question | Weak signal | Strong signal |
|---|---|---|
| Can we see the chain? | Each team knows its own agent | One trace shows agents, people, data, and hand-offs end to end |
| Is authority explicit? | Broad access inherited from a service account | Permissions match role, task, risk, and expiry |
| Can we stop it? | “Someone would notice” | A named owner can pause, revoke, route, and recover |
| Can we explain it? | Output stored without context | Model, configuration, evidence, and decision path are traceable |
| Can we sustain it? | One supplier and one untested route | Priority workloads have tested alternatives and recovery targets |
| Can we learn? | Incidents become anecdotes | Exceptions update evaluations, controls, and operating practice |
The test should be applied to one consequential workflow before the organisation attempts to govern an entire agent ecosystem.
A 30-Day Orchestration Sprint
- Days 1–5: Select the chain. Choose one workflow that crosses teams or systems and document every agent, person, data source, hand-off, and external action.
- Days 6–10: Assign the owner. Name the business owner and record the decision rights that remain human. Define what the chain may recommend, change, or trigger.
- Days 11–15: Instrument the hand-offs. Make identity, timestamps, evidence, model versions, prompts or instructions, and exceptions visible at each transfer.
- Days 16–20: Install boundaries. Add thresholds, separation of duties, stop controls, escalation paths, and a recovery procedure for high-impact actions.
- Days 21–25: Test continuity. Simulate a model outage, a supplier change, a data-quality failure, and an unauthorised action. Record recovery time and unresolved dependencies.
- Days 26–30: Decide the operating model. Scale the workflow, narrow its authority, redesign its coordination, or stop it. Record the conditions required for responsible expansion.
The Bottom Line
AI capability is becoming distributed across agents, services, teams, and infrastructure. That distribution can create speed, but it also creates a new leadership obligation: make the chain visible enough to govern.
The organisation that wins with agents will not be the one with the most autonomy. It will be the one with the clearest coordination.
Continue the thread
Which part of your organisation’s AI chain is least visible today: intelligence, coordination, control, governance, or continuity? Continue the conversation at deuerout.com for the next issue, updated intelligence, and practical transformation frameworks.
Sources and attribution
- California Management Review, “Governing the Agentic Enterprise: A New Operating Model for Autonomous AI at Scale,” March 2026: source.
- OECD, “Securing AI Compute,” 2026: source.
Editorial note: The four-layer Agentic Operating Model is presented as a conceptual framework proposed by the California Management Review article. Company examples and performance claims in that article are not used here as independent benchmarks.
