How to Design an AI Operating Model

Key takeaways

An AI operating model defines how an organisation turns AI capabilities into repeatable business outcomes. It establishes how work flows between people and AI, who owns decisions and risks, how AI initiatives are governed, and how agents are monitored throughout their lifecycle.

• Humans are the compass; AI is the engine. The operating model exists to allocate judgement, execution, accountability and oversight deliberately between them.
• Design starts with flow, not roles. Map end-to-end value flows, then resource them from a shared pool of human and digital skills.
• Build a hub and spoke, not a single centre. A central AI Centre of Excellence sets standards and governs; federated spokes own outcomes and deliver value.
• Every AI-enabled workflow runs in one of two collaboration modes: AI recommends and a human decides, or AI executes and a human monitor.
• Governance is a design problem, not an afterthought. Agent lifecycle ownership, commissioning, monitoring and retirement must be built into the model from day one.
  
  

Start with flow, not roles

Most operating models start with the wrong question: "What roles do we need?" The better question is: "What flows of value are we trying to deliver?"

The redesign of the operating layer, the architecture of a business operating with an agentic workforce, is where this shift happens.

Here is the design principle: map the end-to-end flow of value first, then decide which parts are performed by humans, which by AI agents, and how they hand off to each other. Do not start with the org chart. Start with the customer outcome, then work backwards through the workflows that produce it.

This is not abstract. A claims process, a customer onboarding journey, a regulatory reporting cycle: each is a flow. Each can be redesigned so that AI handles the pattern recognition, data retrieval and first-pass decisioning, while humans handle the judgement, exception handling and final accountability. The AI operating model is the structure that makes this division of labour sustainable at scale.

Build a hub and spoke, not a single centre

Once you have mapped the flows, you need a structure to govern and resource them. The Whitepaper The Operating Layer describes a hub-and-spoke architecture as the structural property that holds the model together.

The hub is the central function, an AI Centre of Excellence that owns governance, standards, tooling, capability development and risk oversight. It does not deliver value directly. It enables the spokes to deliver value safely and consistently.

The spokes are the federated business units, product teams or capability clusters that own specific outcomes. They run the workflows, own the P&L, and are accountable for the value the AI delivers. They draw on the hub for standards, training and governance, but they make their own decisions about how to apply AI to their specific context.

Why does this matter? A single centralised AI function becomes a bottleneck. It cannot know every business context deeply enough to make good decisions for all of them. And a fully decentralised model, with every team doing its own thing, produces the agent sprawl that drives cancellation rates above 40%, according to Gartner's June 2025 prediction.

The hub-and-spoke model solves both problems. Standards come from the centre. Execution lives at the edge. AND Digital's Agentic Operations practice works with leadership teams to design exactly this balance: the right level of central governance without losing speed or business alignment.

Decide how humans and agents collaborate

Every AI-enabled workflow runs in one of two collaboration modes. These are not abstract concepts. They are design choices you make for every workflow, every agent, every decision point.

Mode one: AI recommends, a human decides

The AI surfaces options, analysis, or recommendations. The human reviews, exercises judgement, and makes the final call. This mode is appropriate where:

• The stakes are high and errors are costly
• Regulatory or ethical judgement is required
• The human has context the AI cannot access
• You are building trust in a new capability
  

Mode two: AI executes, a human monitors

The AI takes action within set boundaries. The human monitors for exceptions, reviews outcomes periodically, and intervenes when the AI operates outside its guardrails. This mode is appropriate where:

• The workflow is well-understood and repeatable
• Speed and scale matter more than bespoke judgement
• You have clear success metrics and error thresholds
• The human can meaningfully oversee volume, not just individual decisions

Both modes keep a human accountable for the outcome. The difference is in how that accountability is exercised: before the fact through recommendation and decision, or after the fact through monitoring and intervention.

This is the concrete expression of human AI collaboration and AI and human collaboration, not a vague principle but a deliberate design choice applied workflow by workflow.

The two collaboration modes: a summary

Govern the agent lifecycle and build trust

Governance is not a policy document. It is a set of design decisions about how agents are commissioned, owned, monitored and retired. Gartner's forecast, more than 40% of agentic AI projects cancelled by the end of 2027 with governance failure the dominant cause, makes this the single most important design problem you will solve.

Here is what agent lifecycle governance looks like in practice:

Commissioning. Every agent should have a clear business case, an accountable owner, and a defined scope before it is built. No agent should be deployed without answering: What problem does it solve? Who owns its performance? What data does it access? What are its success criteria?

Ownership. Every agent needs a named human accountable for its outcomes. This is not the person who built it, or the team that maintains it. This is the person who owns what the agent produces. If the agent makes a decision, someone is accountable for that decision. If the agent fails, someone owns the failure.

Monitoring. Agents need to be observed in production. Not just for technical performance, such as latency, uptime and error rates, but for business outcomes. Is the agent delivering the value it was commissioned to deliver? Is it operating within its guardrails? Is it creating unintended consequences?

Retirement. Agents should have a planned end of life. When a workflow changes, when a better approach emerges, when the agent is no longer delivering value, it should be retired cleanly, with its responsibilities transferred and its data archived.

AND Digital's AI services cover this entire lifecycle, from strategy and transformation through to solution design, build and enablement. The governance framework is not an add-on. It is built into the ai operating model from the start.

Trust is the output of good governance, not a separate activity. When agents are commissioned with clear purpose, owned by accountable humans, monitored for outcomes and retired when appropriate, trust follows. When governance is absent, the trust deficit kills adoption, and then the project.

Frequently asked questions

How do you design an AI operating model?

Start with the organising idea: humans are the compass (strategic direction, ethical judgement, final accountability) and AI is the engine (speed, breadth, tireless execution). Then design the structure that allocates judgement, execution, accountability and oversight between them. Map flows first, build a hub-and-spoke structure, decide collaboration modes per workflow, and govern the agent lifecycle from commissioning to retirement.

What are the six dimensions of an AI operating model?

The whitepaper The Operating Layer defines six dimensions: flow and skills architecture; hybrid structural design; agent lifecycle governance; trust, safety and adoption; culture and change readiness; and flow-oriented performance. These are held together by three structural properties: hub and spoke, outcome governance, and a platform-agnostic, governance-first approach.

What is an AI Centre of Excellence, and where does it sit in the structure?

An AI Centre of Excellence is the central hub in a hub-and-spoke operating model. It owns governance, standards, tooling, capability development and risk oversight. It does not deliver value directly. It enables the federated spokes (business units, product teams) to deliver value safely and consistently.

What is hub-and-spoke structure, and why do both parts matter?

The hub provides central standards and governance. The spokes own specific outcomes and deliver value. Both parts matter because a fully centralised model becomes a bottleneck, and a fully decentralised model produces agent sprawl. The hub-and-spoke model balances consistency with speed and business alignment.

How should humans and AI agents collaborate within an operating model?

Every AI-enabled workflow runs in one of two modes: AI recommends and a human decides, or AI executes and a human monitors. Both keep a human accountable for the outcome. The choice depends on the stakes, the regulatory context, the maturity of the capability and the need for speed versus judgement.

How do you govern the AI agent lifecycle within an operating model?

Govern the lifecycle through four stages: commissioning (clear business case, accountable owner, defined scope), ownership (named human accountable for outcomes), monitoring (track business outcomes, not just technical performance), and retirement (planned end of life with clean transfer of responsibilities).

What to do next

The question how to design an AI operating model has a clear answer: start with flow, build a hub and spoke, decide collaboration modes per workflow, and govern the agent lifecycle from commissioning to retirement.

The design is not abstract. It is buildable. And it is the precondition for capturing AI's gains at scale.

Download The Operating Layer for the full six-dimensional operating-model framework, readiness diagnostic and governance structure.

If your organisation is actively designing its AI operating model and you recognise the need for a structure that balances central governance with federated delivery, AND Digital's Agentic Operations practice works with leadership teams to design, build and embed the operating models that let AI scale safely and productively.

The structure exists. The question is whether you design it or retrofit it. One of those is much more expensive.