Ask almost any engineering leader and they will tell you their people are faster with AI. They are right. The studies hold up, the demos are real, and the adoption dashboards are green. Then the same leader looks at delivery predictability, or at the income statement, and finds that very little has actually moved.
This is the paradox sitting inside most AI programmes right now. McKinsey reports that around 80 percent of firms have adopted AI, and about the same share have seen no meaningful impact on revenue or profit. Mass adoption, real individual gains, and negligible institutional return. Something between the faster individual and the unchanged business is absorbing the difference.
That something is the operating model, and it is the part almost nobody has touched.
The faster individual is already here. The faster company is a design decision. The signature insight · Operating Model
Why faster individuals do not make a faster company
There is a precedent worth knowing. When factories first swapped steam engines for electric motors in the 1890s, productivity barely moved for thirty years. The owners had installed the new power and kept the old floor plan: the same lines, the same sequence, the same workflow built around the steam shaft. The breakthrough came a generation later, when factories were redesigned around what electricity actually made possible. As one analysis of that moment puts it, the factories that electrified first lost to the ones that redesigned the floor. The same analysis draws the blunt conclusion for today: productive individuals do not make productive firms.
Most companies are running the 1890s version with AI. They have bought the motor and kept the factory. The individual is faster, but the work still flows through ownership, decision rights, dependencies and hand-offs designed for the pace before. Only about a fifth of organisations using generative AI have redesigned any of their workflows. The rest have bolted a faster tool onto an operating model that cannot carry the speed, then wondered where the speed went.
What the pattern actually looks like
- Adoption metrics are healthy and business metrics are flat. Licences, logins and prompts climb; predictability and margin do not.
- The saving never reaches the income statement. Time freed at the task is reabsorbed by the same queues, approvals and dependencies, and never surfaces as throughput.
- AI is layered onto the old workflow, not built into a new one. The sequence of work is unchanged; only one step inside it is faster.
- The few who redesign pull away. The minority that change the operating model around the tool compound the gain. The majority bank a faster individual and a flat firm.
What to design instead
The lever was never the tool. It is the system the tool runs inside, and it has to be redesigned on purpose.
- Redesign the workflow, not just the adoption. Decide what the work should look like when a step takes minutes instead of hours, then change the ownership, sequence and decision rights around it. The tool is the easy part.
- Put people above the loop, not only in it. As agents take more of the execution, the scarce work becomes judgement, direction and oversight. Design those roles deliberately rather than leaving them to whoever is nearest.
- Measure the delta in your own data. Prove the change against your own delivery and business signals, baseline to re-baseline, so you can see whether the gain reached the firm or stopped at the desk.
The takeaway
The AI productivity paradox is not a technology problem, and a better model will not solve it. The faster individual is already here. The faster company arrives only when the operating model between them is redesigned to carry the speed. Swap the motor if you like. It changes little until you rebuild the factory.