Why AI hasn't delivered the productivity gains we promised

---

Imagine a factory in Ohio around 1900.

The owner has just replaced his steam engine with an electric motor. No more coal. No more boiler. No more worker spending his day keeping the fire alive.

He is proud of the upgrade. His factory has entered the electrical age.

But walk inside, and almost everything looks the same.

The new motor turns the same long metal shaft near the ceiling. Leather belts still connect that shaft to the machines below. The machines still crowd around the power source, whether or not that arrangement makes sense for the work.

The factory has a new engine. It still works the old way.

And output per worker barely improves.

Across American manufacturing, the bigger gains from electricity took decades to emerge. They came when owners changed more than the source of power.

The breakthrough was a different factory

The crucial change sounds surprisingly small: give each machine its own electric motor.

Now the machines no longer need to connect to one central shaft.

That changes everything.

A machine can go where the work needs it. The next machine can sit beside it. Materials can move through the factory in a sensible sequence, instead of taking awkward journeys between machines arranged around the power supply.

Factories can spread across a single floor. Layouts can change. Individual machines can run when needed.

The benefits keep spreading. Different ways of working require different skills. Workers have more decisions to make. Managers have to rethink training, supervision and pay.

Electricity makes all this possible. But buying an electric motor does not make it happen.

The big gains come from redesigning the factory around what electricity allows.

That is the part of the story I think about when people ask me about AI.

A faster tool inside the same organisation

I hear some version of this question regularly:

“We bought the AI licences. We ran the pilots. The demos were impressive. Where are the returns?”

Sometimes, we have done what the factory owner did: installed a better engine and left everything connected to it exactly as before.

AI writes the report faster. But the report still passes through five reviewers.

AI finds a maintenance problem earlier. But the team still waits for the same meeting before anyone can act.

AI answers a customer’s question. But resolving the issue still requires three departments and a manager’s approval.

Each improvement can be useful. Yet the overall process may barely change.

To understand why, we need to ask what our organisations were built around.

Much of their structure reflects a practical fact: experienced people have limited time.

A difficult question goes to a specialist. An exception goes to a manager. Information goes to a central team, which reviews it and sends instructions back.

We built approval chains, escalation paths and entire departments around getting work to the people qualified to decide.

Those arrangements are our version of the factory’s central shaft.

They often serve a real purpose. But we can become so used to them that we stop asking which are still necessary.

What could happen closer to the work?

AI does not make human judgment unlimited or remove the need for accountability. It can, however, make analysis and advice available in places where they were previously too expensive or too slow.

A frontline worker may be able to check a fault against maintenance records immediately.

A customer service officer may be able to resolve a routine exception within clear limits.

A supervisor may be able to compare options without waiting days for a specialist’s report.

That raises a more interesting question than “Which tasks can we automate?”

If people could get reliable help at the point where the work happens, how would we organise the work differently?

Which decisions could happen sooner? Which handovers could disappear? Which cases would still need an expert?

And what would we then hire, train and reward people for?

If AI produces the first draft, a person’s value may shift towards checking it, handling exceptions and deciding what to do next. That changes the job, not just the time taken to complete it.

Saving time is only the beginning

The factory owner had a good reason to move slowly. Replacing a working building and its machinery was expensive. He might reasonably wait until they needed replacing anyway.

Organisations today face real constraints too: old systems, contracts, budgets and regulation.

But many of our barriers are easier to change than a brick factory. They are meeting schedules, approval rules, departmental boundaries and habits.

We should be careful about treating all of them as permanent.

A Bank of Korea study illustrates the problem. It estimated that generative AI saved workers roughly ninety minutes a week. But those savings did not automatically translate into more output. Whether workers could use the time productively depended on their freedom and motivation to do so.

That is easy to recognise.

If someone finishes a report an hour earlier but still has to wait until Friday’s approval meeting, the business has not gained an hour of progress.

The next decision happens at the same time.

A calmer working week is a worthwhile benefit. But if we promised faster service, more capacity or higher output, we still have work to do.

Controls need to change with the work

In regulated operations, this gets harder.

An approval may prevent a dangerous action. A second reviewer may protect against fraud. An audit record may establish who was responsible.

We cannot simply remove those controls because AI makes something faster.

We need to work out how to keep their protection as decisions move to different places.

What can the system recommend? What can it do automatically? When must a person intervene? What evidence must be recorded so a decision can be checked later?

Answering those questions takes operational knowledge, careful testing and agreement among people responsible for the risks.

It rarely makes for an exciting demo. Yet it may determine whether the demo ever becomes a better business.

The lesson from electrification is that a powerful new technology creates possibilities. Organisations still have to turn those possibilities into a different way of working.

A century ago, buying the motor was the beginning.

The bigger gains came when owners realised the factory no longer needed to revolve around the shaft.

We have bought the AI.

Now we need to decide what we are ready to rebuild.

Historical sources: Paul David, “The Dynamo and the Computer” (1990), and Warren Devine, “From Shafts to Wires” (1983). The AI time-savings research is by Suh and colleagues at the Bank of Korea.