Why AI pilots stall (and what electrified factories teach us)

AI pilots stall because nobody in the room understands both the work and the tools. Your leadership team knows where the hours go: the report that eats every Thursday, the approval that waits in someone's inbox, the customer emails answered by copy and paste. They don't have time to learn AI. Vendors and consultants know AI, but they don't know where your hours go, and nobody pays them to stay until it ships.

So companies do the reasonable-looking thing. They buy licenses, run a pilot, applaud a great demo, and very little reaches production. We call it pilot purgatory.

In a report dated July 2025, MIT's NANDA initiative found that only about 5% of enterprise generative AI pilots were producing fast revenue gains. The rest delivered little to no measurable impact on the P&L. (The preliminary study drew on 52 interviews, 153 survey responses from senior leaders and a review of more than 300 public AI initiatives.)

Factories got electricity decades before they got the productivity, because they plugged the new motor into the old floor plan. The economist Paul David told this story in 1990, in a paper called "The Dynamo and the Computer." His line: in 1900, you could have seen dynamos everywhere but in the productivity statistics.

A steam-era factory was built around one engine turning long shafts, with belts running down to every machine. Long shafts waste power, so factories went up, not out: multistory buildings, machines placed where the shaft was rather than where the work flowed. The first electric factories replaced the engine and kept everything else.

The gains came when people redesigned the building around the motor. Lighter construction, one floor, machines arranged in the order the product moved, a line you could rearrange without shutting the whole plant down.

In most companies, nobody. Follow the money actor by actor and every one of them is behaving rationally, which is exactly why the pilot stays a pilot.

Nothing about this is new. In 1901, F. B. Crocker, whose Crocker-Wheeler Electric Company sold the motors, wrote: "There were many factories which introduced electric power because we engaged to save from 20 to 60 percent of their coal bill." The bigger payoff was what he called "indirect savings" (David lists them: less labor oiling belts, smoother work flow, better machine control). Vendors sold the coal bill; the prize was the floor plan.

Today's version of the coal bill is the license pitch: every seat does the same old work a little faster. It's real. It's also the small number. The big number sits in work nobody would design the same way if they started this year.

You put one person who understands both the work and the tools inside leadership, and you give them a deadline. Factories hit the same bottleneck. David points out that the 1920s gains depended on building up a cadre of factory architects and electrical engineers who knew the new approach, and that took years: factory builders were small, scattered firms with high turnover, so the know-how spread slowly.

1. Inside leadership. In the meetings where the hours get spent, with enough standing to change the floor plan.
2. Understands both. Someone who can map where the hours go and then build a working prototype on your real data the next month, with no loyalty to any vendor's product.
3. A deadline. A hard end date forces the question every pilot avoids: does this go to production or not?

That's the job we built the 90-Day Placement to do. One senior AI operator joins your leadership team. Days 1–30, they shadow leadership and map what to automate, ranked, each with an estimate of hours saved. Days 31–60, they turn the top opportunities into working prototypes on your real data, and you decide what goes to production. Days 61–90, they ship, document, and train the people who will own it.

We're paid for a fixed 90-day engagement, so the only way for us to look good is to ship. (More on that in why not a consulting firm.) Not every company needs 90 days. If you only need a ranked roadmap or a second opinion on a vendor contract, AI consulting is shorter.

Rewiring a workflow doesn't require pouring a new building, so I expect AI to move much faster than four decades. David himself found that the growth in factory electric motors accounts for only about half of the 1920s productivity jump. Vendors aren't villains either: the same MIT study found that tools bought from specialized vendors succeeded about two-thirds of the time, far more often than internal builds. And one operator can't fix a leadership team that refuses to change anything; they can only make the cost of not changing visible.

I don't have the answers for your company. These are the questions I'd want answered before buying one more license.

• Which process would you design completely differently if it had been invented this year?
• Who on your leadership team can explain both that process and what today's models can do with it?
• How many AI licenses do you pay for, and how many people would notice if you canceled them tomorrow?
• What was the last pilot that reached production, and who owned it the day after the demo?
• What is your line shaft: a system, an approval chain, or a report everyone works around?

If you have answers, or better questions, I'd like to hear them.