Electric power reached American factories in the 1880s, and for something close to four decades afterward the productivity statistics barely registered it. Paul David, an economic historian at Stanford, worked out why and laid it out for the American Economic Association in 1990.1

A nineteenth-century factory was built around a single steam engine turning one long shaft that ran the length of the ceiling, with belts dropping down to every lathe and press below. Where a machine sat was decided by where a belt could reach. When electricity became available, most owners swapped the steam engine for a large electric motor and left the shaft, the belts, and the floor plan exactly where they were. Fuel bills improved. Everything else about the building carried on as before. The productivity acceleration showed up in the 1920s, once factories gave each machine its own motor and could finally arrange the work in whatever order the work wanted to go.1

Eighteen percent, three functions deep

The Census Bureau has been measuring something structurally similar. Its 2026 artificial intelligence supplement to the Business Trends and Outlook Survey found that 18 percent of American firms used AI in a business function during the November 2025 to January 2026 reference window, rising to 32 percent when weighted by employment.2 The interesting numbers sit underneath that headline. Among firms that had adopted, 57 percent confined it to three or fewer business functions, most often sales and marketing at 52 percent and strategy at 45 percent.2 At the level of individual workers, 23 percent of firms reported any work-related task use at all, and 65 percent held that use to three or fewer tasks.2 Employment decreases attributable to AI turned up in 2 percent of firms.2

Read that as a floor plan rather than an adoption curve. A great many organizations have put the motor in one corner of the building while the belts still run everywhere else.

Automation pays in runs, not in scattered steps

A working paper released through the National Bureau of Economic Research in February gives a reason that pattern yields so little.3 Mert Demirer, John Horton, Nicole Immorlica, Brendan Lucier and Peyman Shahidi model production as a sequence of steps, each of which can be done by hand, assisted by AI, or handed over completely, and they argue that full automation happens inside contiguous runs of steps they call chains.3 Their empirical work supports the shape of that idea. Steps executed by AI cluster together rather than scattering across a job, and the two related findings are about position: where a job's automatable steps sit dispersed among human steps, AI execution at the job level goes down, while a step sitting next to an already-automated step becomes more likely to be automated itself.3 Horton discloses in the paper that he is a paid advisor to Anthropic, worth knowing when you read anything modelling AI, though the finding is an awkward one for anybody selling capability, since it says capability was never the binding constraint.

The unit of automation is the run of steps that nobody has to interrupt. Everything shorter than that is a demonstration.

The question that costs something

So the question most leadership teams are asking is the cheap one. Ask which tasks AI can do and you have asked something that requires nothing from anybody. Every function answers it in an afternoon, produces a slide, and leaves the building as it was. Ask instead which sequence of steps your organization owns from one end to the other, and the question acquires a price, because a real sequence crosses more than one budget line and more than one vice president, and somebody has to give up a piece of their territory before the run can close.

I run into a small version of this in my own practice. The automations I have built tend to work cleanly for a few steps and then sit there waiting on me. A file lands in the wrong folder. An interface hits its quota partway through a batch and stops. Each individual step does its job. What I own is a collection of well-built fragments with a person standing in every gap, and the gaps cost more than the steps they connect.

Most organizations are running a much larger version of the same arrangement and reporting it as adoption. This sits close to the argument I made about where value actually gets decided in delivery work, and it is a cousin of the failure that turns an AI rollout into an announcement nobody acted on. The tool arrives and the sequence stays where it was.

What the data is pointing at

The Census researchers also ran regressions and found a positive correlation between a firm's commercial performance and the breadth of its AI integration, across functional deployment and task-level use.2 Correlation, and they treat it as such, so which way the causation runs is open. The pattern still lines up with what the chaining model predicts, and it points somewhere uncomfortable for an executive handling this as a procurement decision. The firms getting something out of it appear to be the ones who moved the machines.

There is a second implication that tends to get missed. If adjacency drives automation, whoever controls the boundaries between steps controls the pace of the whole thing. In most companies those boundaries were drawn by the org chart rather than by the work, set years ago by people who had no idea they were capping something that did not exist yet.

Here is the exercise I would put in front of a leadership team this quarter. Take one sequence a customer would recognize, from the arrival of a request through to the moment somebody is satisfied, and count the places where the work stops and waits for a person to pick it up again. Then find out who has the authority to remove one of those stops, and whether anyone has ever asked them to. My read is that the count runs higher than anyone expects and the authority sits further from the work than anyone would like to admit. What is your count?