When factories first gained access to electrical power in the 1880s, most of them did something predictable. They replaced the steam engine with an electric motor and left everything else exactly as it was. The layout stayed. The line shafts stayed. The belts and pulleys stayed.

The result was modest. Energy costs dropped. Output per worker edged up. But the large productivity gains electricity was supposed to deliver didn’t come. Not for about thirty years.

Paul David, an economic historian, found that the productivity gains from electrification showed up mainly in the 1920s, forty years after factories first had access to the technology. The lag wasn’t about the technology. It was about whether businesses were willing to redesign around what the technology made possible.

Redesign, not re-equipment

Steam-era factory layouts had one fundamental constraint: the line shaft. Every machine had to sit within belt range of a central overhead shaft driven by the steam engine. You couldn’t group machines by process sequence. You couldn’t stop one without affecting others. The layout was dictated entirely by where the power was.

Electric motors changed that. A motor could sit directly on a machine. The machine could sit anywhere there was wiring. Which meant the question of factory layout could now be answered differently: in what sequence does the work actually flow?

Ford’s moving assembly line was physically dependent on decentralised electric drives. You couldn’t have built it around a line shaft. Plant-level labour productivity gains from this kind of reorganisation ran 20-30% relative to comparable steam-era plants.

The factories that got 20-30% labour productivity gains from electrification didn’t do it by buying better motors. They did it by asking what the motors made possible that hadn’t been possible before, and then building the operation around the answer.

You’re starting with the wrong question

Most AI implementations start the same way. Pick a workflow, apply a tool, measure whether it performs as described. It’s a reasonable starting point: low risk, builds confidence, produces some efficiency gain.

The issue is that it almost never touches the constraint.

When AI is applied to a step inside an unchanged workflow, the constraint hasn’t moved. The AI is running alongside the operation, not through it. MIT’s NANDA Initiative analysed 300 public AI implementations: 95% delivered no measurable P&L impact.

Adoption is high. Impact is still low. McKinsey’s 2025 data shows 78% of organisations using AI in at least one function. Most haven’t seen bottom-line impact yet.

The question that changes this isn’t where can we use AI? That produces reasonable pilots and modest gains. It’s: where does our operation most need to improve, and where can AI make that possible in a way it wasn’t before?

You identify the constraint first. You find where improvement creates competitive value. Then you ask whether AI is the right tool to address it. When it is, the improvement is measurable, because you designed the engagement around the problem rather than the tool. Gartner reports 30% of generative AI projects will be abandoned after proof of concept by end of 2025. The line shaft was still there.

Not every problem is an AI problem

Not every improvement opportunity is an AI opportunity.

If a constraint can be addressed by a clear process redesign and a person following a different procedure, that’s a process improvement problem. AI may accelerate it, but it’s unlikely to change the fundamental nature of what’s possible.

The places where AI creates genuinely large differences share a characteristic: decisions that need to happen faster or at a scale that no manual process can match.

Predictive maintenance is the clearest manufacturing example. A machine that’s about to fail doesn’t announce itself consistently, and an unplanned failure costs significantly more than a planned maintenance window. Reading sensor data from dozens of machines simultaneously, at a granularity no maintenance team can match manually, and surfacing specific interventions before the failure: that’s something AI changes materially. The process redesign that captures the value isn’t complex: maintenance scheduling moves from a calendar to what the model recommends. Without that redesign, the model sits in a dashboard that nobody acts on.

The capex analysis of which machines to replace can probably be done well on a spreadsheet. Applying AI to it produces a marginal improvement. That’s not the engagement worth pursuing.

A note on the analogy’s limits

Electricity is a commodity. AI isn’t. A kilowatt-hour is a kilowatt-hour; AI is a family of models, data pipelines, and applications with widely varying capabilities. You can’t just “plug it in.” The configuration, the data preparation, the integration with existing systems all require more judgment than electrification did. If electricity required process redesign to unlock its value, and electricity could literally just be plugged in, AI requires even more. Organisational design, not procurement, is where the work actually is.

Electricity also changed physical processes. AI changes information processing, forecasting, and decision-making. The “factory floor” it’s reorganising is the logic of how your organisation makes decisions and allocates attention. Economists Brynjolfsson, Rock, and Syverson argue we’re currently in the low part of the J-curve for general-purpose technologies: high investment, low measured productivity, before a step change arrives as firms build the complementary capabilities the technology requires. Those who have done the harder work of redesigning are the ones who will be at the front of that step change.

Three diagnostics

If you’re considering AI at scale, start with the improvement that matters most.

  1. Look at the AI pilot that produced the best-looking results. Is it touching the actual throughput, margin, or error rate of the operation? Or improving something that runs alongside the core process without changing what it produces?

  2. Look at the AI initiative that stalled. Was the technology the problem, or was it that the process around the technology hadn’t changed? Most stalled AI projects have functioning technology. What they lack is a redesigned workflow that makes the technology’s output part of how decisions actually get made.

  3. Ask your team: if the AI went away tomorrow, what would change? If the answer is “we’d go back to doing it the way we did before,” the AI is running alongside the process, not through it. That’s the business that replaced the engine without redesigning the layout.

AI has moved quickly from interesting possibility to practical tool. What hasn’t moved as quickly is how you are supposed to implement it without letting your operations and processes grind to a halt. The competitive advantage in the next few years won’t go to the operations with the most AI. It will go to the ones who looked at the whole system and found the largest gains first.

Manufacturers and distributors who figure out how to use AI for competitive advantage in the next few years will have built something their competitors will struggle to close quickly. Everyone has access to the same tools. What no competitor can replicate by purchasing them is the process redesign behind them.

If you’re working through where AI creates that kind of leverage in your operation, that’s a conversation we’d love to join.