Walk through almost any large company today and you will find people using AI.

Someone is rewriting an email with Copilot. A marketer is interrogating ChatGPT. A developer is coding with Claude. An analyst is turning a long report into a one-page summary before the coffee gets cold.

From the employee’s perspective, something significant has changed.

From the company’s perspective, surprisingly little may have changed at all.

The email still needs the same approval. The report still moves through the same departments. The customer request still waits in the same queue. The meeting still happens. Someone still copies information from one system into another.

We have become remarkably good at putting artificial intelligence inside old ways of working.

Then we wonder why the financial results are not keeping pace with the productivity gains.

McKinsey’s 2026 global AI survey captures the gap neatly. Eighty percent of respondents said AI had improved their individual productivity. Only 37 percent said it had contributed positively to their organisation’s EBIT. Among the small group McKinsey classified as AI high performers, nearly three-quarters said they were fundamentally redesigning workflows around AI. Among everyone else, only about a quarter were doing so. McKinsey & Company

Perhaps we need a better definition of adoption.

Task adoption → Workflow adoption → Organisational transformation

Task adoption

This is where most of us first encounter AI.

A task that took an hour takes twenty minutes. Research gets faster. Emails improve. Code appears in seconds. Presentations no longer start with a blank slide.

The person becomes more productive.

That matters. The productivity benefits of AI are not imaginary. The International Labour Organization’s review of the evidence finds substantial gains at task level, particularly for well-defined cognitive work and less experienced workers. International Labour Organization

But improving a task is not the same as improving a company.

A field experiment involving 7,137 knowledge workers across 66 companies illustrates the problem. Employees were given generative AI inside tools they already used for email, meetings and writing. Workers who used it saved significant time on email and worked less outside normal hours.

Yet the researchers found no meaningful change in the quantity or composition of their work. NBER

People changed, but the system around them did not.

This is the distinction that matters:

Individual productivity is not organisational productivity.

Imagine AI reduces a particular task from four hours to one.

Excellent.

But after those sixty minutes, the output waits two days for approval.

You have saved three hours of processing time inside a process that still takes days.

The employee feels the improvement immediately. The customer may barely notice.

Workflow adoption

The next stage starts with a different question.

Instead of asking:

How can AI help this person complete this task faster?

Ask:

Why does this work happen this way in the first place?

Now the unit of analysis changes.

You begin looking at handoffs. Waiting. Approvals. Duplicate data entry. Information trapped in different systems. Decisions that automatically travel upward because the organisation was designed before AI could evaluate an exception, retrieve context or perform part of the work itself.

This is where AI starts becoming more than a productivity tool.

A recent INSEAD field experiment involving 515 high-growth startups called one obstacle the “mapping problem”: companies struggle to discover where and how AI can create value across their production processes.

When firms were exposed to examples of how other companies had reorganised production around AI, they discovered 44 percent more AI use cases. They completed more tasks, were more likely to acquire paying customers and generated higher revenue than the control group. INSEAD

The point is not that every company should copy somebody else’s AI playbook.

It is that there is a big difference between giving employees AI and redesigning work around what AI now makes possible.

The ILO describes a similar problem at a much larger scale. Strong productivity gains at task level have not yet translated consistently into firm, sector or economy-wide productivity growth. Its research points to something previous technological revolutions taught us too: new technology produces its biggest gains only after organisations change around it. International Labour Organization

Electric motors eventually changed more than the machines they powered. They changed how factories were designed.

AI may require something similar for knowledge work.

Copilots made this easy to ignore. Agents will not.

A copilot helps a person perform a task.

An agent can increasingly perform parts of the workflow itself. That makes some old processes look increasingly strange.

DBS recently rolled out an agentic AI system to around 1,500 corporate bankers. Specialised agents handle more than 70 tasks involved in producing the first draft of a corporate credit memo. Bankers and risk managers then review the analysis, apply judgement and spend more time on higher-value client work. DBS Bank

Notice what DBS did not do.

It did not simply give bankers a chatbot and ask them to write credit memos faster.

It decomposed the work. That distinction will become more important as agents improve.

Once AI can move across systems, retrieve information, execute tasks and escalate exceptions, companies will have to confront questions that copilots allowed them to postpone.

Why does this task pass through three departments?

Why does this approval exist?

Why are five software applications involved?

Which decisions genuinely require a human?

Where should human judgement become more important precisely because machines are doing more of the routine work?

And eventually, an even more uncomfortable question:

What should this company look like if we designed it today?

That is where workflow adoption begins turning into organisational transformation.

Roles change. Decision rights change. Teams change. Software changes. Metrics change. Some work disappears. Other work becomes more valuable.

At that point, AI is no longer something employees use.

It has changed how the company operates.

For the last few years, leaders have understandably asked:

How many of our employees are using AI?

That question is becoming less useful.

A better one is:

What works differently here because AI exists?

If the answer is that employees write emails faster, build presentations faster and summarise documents faster, that is useful progress.

But your employees may have adopted AI. Your company has not.

Real adoption starts when tasks change. It deepens when workflows change. Transformation begins when the organisation itself changes around what the technology now makes possible.

The companies that gain the most from AI may not be the ones that use it most.

They may be the ones most willing to redesign themselves around it.