A century of industrial engineering already mapped the road we're on. The productivity gains from AI are real. They're just stuck at the wrong scale.
I'm old enough to have lived several lives. I started my career as an Industrial Engineer way back in the 1980’s, back when that meant time-and-motion studies (I was the guy with stopwatch and pad in hand on the factory floor), process maps drawn by hand, and a genuine reverence for two men who had never met but between them shaped how the modern world makes things: Frederick Taylor and W. Edwards Deming.
I've been watching the rollout of AI across Australian business with that training running quietly in the background. And I keep arriving at the same conclusion. We are not in a new story. We are in a very old one, at a very specific chapter.
Right now, AI is Taylorism.
My colleague David Hanus put the modern version of the problem sharply last week (link). AI is delivering real productivity gains - 10x, 20x on individual tasks - but it isn't flowing through to business results. Because a business doesn't improve when one step gets faster. It improves when work flows smoothly from beginning to end. The gain races ahead until it meets a constraint - a governance gate, an approval, a legacy system - and then it queues. He's right. And the history of industrial engineering explains why.
Taylor made the parts fast
In the early 1900s, Frederick Taylor gave the industrial world its first significant productivity methodology that made a lasting impact. Scientific management. Break every job into its component steps, study each one, standardise it, make it repeatable and faster. It was a genuine leap. Individual tasks became far more efficient than anyone had thought possible at the time.
But the factories didn't improve in proportion. A faster station simply produced more half-finished work, which piled up in front of the next bottleneck. Taylor optimised the parts. The whole production line barely moved because of the lack of end to end co-ordination and integration.
Deming made the system flow
The correction came from an American statistician, and it came out of the wreckage of a world war.
W. Edwards Deming arrived in occupied Japan in 1947 to help rebuild the country's shattered census. He stayed to teach its engineers something far more valuable. Invited back by the Union of Japanese Scientists and Engineers in 1950, Deming taught a generation of Japanese industry to apply statistics to production - to measure and reduce the variation in their processes and components they were manufacturing, to treat the business as one connected system rather than a line of separate stations, and to strip out the stockpiles of inventory that sat between steps hiding every problem.
Japanese manufacturers took it seriously in a way the West did not. Toyota took it furthest, turning it into a complete system of flow that the rest of the world is still studying. That thinking has shaped global industry ever since the war.
That is the part worth sitting with. The leap from Taylor to Deming - from optimising steps to designing flow - was real, and it was enormous. But it took the aftermath of a world war to force it. A country flattened, with no slack and no choice, was the environment that finally made the systems view non-negotiable.
I would like to think we don't need another one to take the next step.
A good person cannot beat a bad system
There is one more piece of Deming's thinking that matters here, and it is the piece most people miss. Deming was adamant that when work goes wrong, the fault almost never lies with the individual doing it. "A bad system will beat a good person every time," he said. He reckoned the vast majority of problems - by his later estimate, as high as 94% - sit with the system, not the individuals. You can hire the best people, train them, motivate them and lean on them, and it will not matter. If the system around them is broken, the system wins.
This was a rebuke to a century of management instinct, which had always reached for someone to blame or someone to reward. Deming moved the responsibility upward. The people are already doing their best inside the constraints they have been handed. Fix the process and flow between steps - and the outputs improve on their own. Leave the system broken, and no amount of individual effort will save it.
Hold onto that, because it is about to describe our exact situation with AI.
They were never rivals. They were stages.
Taylor and Deming are often set against each other. They shouldn't be. You cannot design good flow across steps that were never made competent in the first place. Taylor made the steps capable. Deming made the system flow. It is a sequence of maturity - capability first, then capacity - and pretty much every organisation that has ever “grown up” has walked it in that order.
AI is walking that same road again, but we are too close to the action to be able to observe the pattern.
Today, we are deep in the Taylor phase
We are taking individual knowledge-work steps - drafting, analysis, research, code - and making them dramatically faster. This is real and it is valuable. But it is step optimisation. And like Taylor's factories, a faster step is landing us with more work queued in front of the next step wether its an approval, the next hand-off or the next decision nobody has thought thro in a way that pulls everything together as an end to end co-ordinated system.
There is a twist that makes the analogy sharper still. AI does not only speed steps up. It introduces variation. Outputs that need checking. Quality that swings from one run to the next. Deming spent his life teaching that variation is the enemy of flow. A faster step that produces less predictable work can lower speed and throughput. We have, in a sense, recreated the exact problem he was sent to solve - only this time it’s in the knowledge factory.
And here Deming's sharpest lesson comes back to bite. The productivity debate right now is really an argument about the worker: is the AI good enough, is the model capable enough, is it accurate enough? Deming answered the shape of that question seventy years ago. A good person cannot beat a bad system. Neither can a good AI. Drop the most capable model in the world in to perform discreet steps in a business whose approvals, hand-offs and data are broken, and the system will beat it every time. The outcome is not decided by how clever the tool is. It is decided by the system and context the tool runs inside. Which means the productivity question was never really a technology question. It is a leadership one.
The next phase of AI is a Deming phase
It will not be a better model. It will look like this. Stop optimising isolated steps. Map the end-to-end process and find the context vacuums and constraints that actually govern how value flows. Reduce the variation in AI's outcomes so the work downstream can trust it. Remove the roadblocks and the queues that sit between steps. Then redesign the system so the speed we have unlocked has somewhere to go.
That is not a technology project. It is an operating-model one. It is the same shift Deming forced seventy years ago, translated from the factory floor to the office.
The choice the history leaves us
Last time, it took a catastrophe (WW2) to make the Japanese economy adopt the systems view, we then started observing how Toyota and other Japanese companies were dramatically improving manufacturing. We can wait for our own version of that - seems plenty of world events that could act as the trigger at the moment (I really hope we don’t go there), or we can make the leap deliberately, while we still have the room to.
Taylor made the parts fast. Deming made the system flow. AI has handed us the parts. The harder, and far more valuable, half of the work is still ahead of us.
The question I'd leave you with is the one Deming would have asked. Not "how do we make this step faster?" We already know how. The question is: where does the work actually get stuck - and what would change if it could flow?
