From AI Activity to Operating Leverage
The operating model is not the technology. It is the conditions that let AI produce value.
I keep using the phrase AI operating model, so it is time to say exactly what I mean by it and how I implement it.
An AI operating model is the working environment that turns AI capability into value. It is the set of conditions that determines where AI belongs, what it is allowed to touch, who owns the consequences, how intervention works, what evidence is captured, how learning travels, and how value is proved.
I help organisations turn AI from scattered activity into operating leverage. Not by starting with tools, pilots or generic use cases, but by finding the workflows where AI can change something the business already cares about: margin, risk, speed, cost, customer experience, revenue, compliance or decision quality.
That matters because AI does not create value simply by being available. It creates value when the surrounding work is ready to absorb it. A model can draft, classify, summarise, analyse, recommend or act, but the business still has to decide where that capability belongs, what decision it improves, what authority it carries, what evidence it leaves behind, and what would prove it should scale.
Usage means people have access. Absorption means the organisation can carry what AI changes and turn local learning into repeatable capability.
This is the bottleneck. AI rarely removes the work as cleanly as people imagine. More often, it moves the work. The constraint shifts from execution into judgement, routing, authority, exception handling and learning.
For me, the operating model is not a set of bureaucratic gates. It is the set of conditions that allow faster, safer learning.
It starts with value anchoring. AI has to be attached to a business lever before the organisation starts multiplying use cases. The question is not “where can we use AI?” The question is where the business needs better intelligence, faster interpretation, stronger evidence, better routing or more reliable execution. If AI cannot be connected to a value lever, the work may still be interesting, but it is not yet operating leverage.
Then comes work mapping. AI does not enter the process diagram. It enters the real work: the handoffs, exceptions, judgement calls, compensating behaviours and small human repairs that keep the organisation moving. This does not mean months of analysis. It means getting close enough to the work to see what the formal process misses. The AI problem is often not the AI. It is the workflow the AI is being asked to enter.
Then the organisation has to define delegation boundaries. A summariser is not the same as a system that changes a customer record. A drafter is not the same as a reviewer. A recommender is not the same as an actor. Each role carries a different level of authority, risk and evidence. The job is to decide where AI can move quickly, where it needs supervision, and where it has not yet earned more authority.
Then come authority and interruption. Once AI moves beyond assistance and starts carrying intent through a workflow, the organisation needs to know who can stop it, override it, narrow it, or own what it touches. That is what makes speed sustainable. When interruption is designed, autonomy can widen with less drama.
The same is true of evidence and controls. Serious AI-enabled work has to leave a trail where the work happens. Not a separate compliance exercise after the fact, but evidence produced by the workflow itself: what happened, why it happened, under whose authority, using which inputs, what changed, and what was escalated. That is what makes the system easier to trust, easier to govern, and easier to improve.
Finally, the work has to be measured by outcomes, not activity. The useful question is not how many prompts, pilots, licences or copilots exist. It is what changed in the work. Did risk fall? Did cycle time improve? Did decisions get better? Did revenue or margin move? Did the workflow become more reliable, more valuable, or easier to govern? If the answer is not visible in the work, the programme is probably still operating at the level of activity.
In practice, I start with a senior calibration session. The aim is to give leadership a shared lens and language for where AI is already entering the organisation, where value is possible, where ownership is unclear, and where effort is becoming duplicated or hard to govern.
Then I choose one or two workflows where value, risk and ownership are real enough to matter. Not the whole enterprise. Not a grand transformation programme. A workflow with business weight.
Then we map only what needs to be mapped to make the next move useful. Enough to understand the decisions, handoffs, exceptions, controls and evidence. Enough to define AI’s role. Enough to decide what it can do alone, what needs approval, who can intervene, and what would prove it is worth scaling.
Then we run it small enough to stop but real enough to matter.
A pilot that cannot affect the business teaches too little. A deployment that cannot be stopped risks too much. The serious work sits between those two mistakes.
The learning then becomes reusable. The organisation is not left with another isolated experiment. It has a pattern for where AI belongs, how authority should widen, what evidence matters, and what should be scaled, redesigned or stopped.
The value is not a slicker AI strategy. The value is faster learning under control. Leadership gets a way to decide where AI belongs, where it does not, what should be funded, where friction can be removed, where controls are needed, and where autonomy can widen.
The missing work is not model access. It is absorption capacity.
I wrote about this previously as the AI absorption problem.
This is the work I do: helping organisations absorb AI into workflows, controls, roles, evidence and operating decisions so they can learn faster, scale what works, and stop noise, risk or waste before it hardens.
Stuart Winter-Tear
AI as Capital Discipline