Most AI programmes stall in the same place. The technology works in a pilot, and then nothing around it changes: the workflow, the roles, the controls and the way value is counted all stay as they were. The pilot stays a pilot.
An operating model is how an organisation turns effort into value: who decides, who does the work, with what tools, under what controls, measured how. Scaling AI means changing all of those together. We look at five parts, and work through them in order.
The five parts that have to change.
01. Value ownership.
What changes. Every AI use case gets a named business owner, a baseline and a target in the P&L before anything is built. Funding follows value cases, not technology roadmaps.
Signs it hasn't. AI sits in a central team or with IT. Success is reported as pilots launched, licences bought or usage. Nobody outside the AI team can say what a use case is worth.
The first move. Pick the few use cases closest to your core workflows and give each one owner, one number and a date.
02. Workflows.
What changes. Work is redesigned end to end around what AI does well, rather than AI being bolted onto each step of the old process. Hand-offs, approvals and checks move to where they add value.
Signs it hasn't. AI speeds up one task while the queue before and after it stays the same. The process was never written down, so nobody can say where the agent should stop.
The first move. Map one high-volume workflow as it really runs today, then design the version where AI does the routine work and people handle the exceptions.
03. Roles and teams.
What changes. Roles shift from doing the task to specifying, checking and handling exceptions. Teams form around outcomes, with one senior person owning a workflow and the agents that run it.
Signs it hasn't. Headcount is cut before the work is redesigned, and the judgement and know-how that went with it has to be hired back. Training happens once, away from the work.
The first move. Redesign roles task by task for the first workflow, train people on their own live work, and be open about what is changing.
04. Platforms and data.
What changes. Build only the platform and data each use case needs, in your own environment, and design it so a model can be swapped out. Fix the data a decision depends on, then keep it fixed with checks.
Signs it hasn't. A multi-year data programme has to finish before anything starts. The architecture is tied to one model provider. Agents can't reach the systems they need.
The first move. For each first-wave use case, check whether the system can reach the data it needs and whether that data is accurate enough for the decision.
05. Governance and measurement.
What changes. Controls are matched to risk and built in from the first wave: heavy where customers, money or regulators are involved, light elsewhere. One named executive owns it. Value is tracked against the baseline in the P&L.
Signs it hasn't. A sign-off gate at the end that teams work around. Nobody knows what AI is actually in use. Nobody can show the board what has been returned.
The first move. Take an inventory of what is in use, sort it by what could go wrong, and name the executive who owns AI risk.
The order matters.
Value ownership comes first, because everything else is judged against it. Workflows and roles come next, redesigned together for the first few use cases. Platform and data work follows what those use cases actually need, which keeps it small and fast. Governance and measurement run through all of it from the first wave, not after it.
Then repeat. Each wave is chosen on the evidence of the last: what moved the number, what didn't, and what the organisation can now absorb. That compounding is where the durable advantage sits. It is set out step by step in the Lumo method.
Five common mistakes.
- Leading with the platform. Buying the platform first and looking for use cases second. The spend lands; the value doesn't.
- Bolting AI onto the old process. The model gets faster at one step of a workflow nobody redesigned, so the end-to-end result barely moves.
- A central AI team with no business owners. Capability builds up in one place and never reaches the P&L, because no business leader owns the number.
- Cutting before redesigning. Announcing reductions before the work has changed. People resist quietly, and the know-how the tools can't replace walks out.
- Governance at the end. A sign-off gate bolted on after the build slows everything down and still misses the real risks.
How to tell it is working.
- Business leaders, not the AI team, can say what each use case is worth and whether it is on target.
- AI runs inside core workflows, and the end-to-end cycle time or cost has moved, not just one step.
- People in changed roles can explain where they check the output and why.
- The board sees value against a baseline in the P&L, alongside the controls that apply.
- The next wave was chosen on evidence from the last one.
Related: how to measure the ROI of AI, how AI changes work, how to govern AI and whether your data is ready.