Why AI adoption lags hardest where the work is physical

Three measurements, three methods, one gradient

Three organizations have looked at this using methods with almost nothing in common. A survey asks firms what they do. A bank watches what they buy. A model provider counts who shows up. Three different blind spots, one answer.

The Census figures are the sector detail behind the Federal Reserve’s summary above; the national rate has moved between 17% and 20% across the Bureau’s own reporting, which is why the reference period matters more than the decimal.

JPMorganChase’s economists, Chris Wheat, Chi Mac and Andrea Passalacqua, give the conventional reading in their April 2026 report:

“The concentration in knowledge-intensive industries may reflect the fact that AI primarily impacts cognitive and knowledge-intensive activities, while tasks with strong physical components in construction, manufacturing, and personal services remain less affected.”

— Chris Wheat, Chi Mac and Andrea Passalacqua, JPMorganChase Institute

That is half right. The half it misses is where the money is.

The gap was wider before the question got easier

One detail inside the measurement is worth more than any of the headline numbers.

In November 2025 the Census Bureau rewrote its AI question, from AI used in producing goods or services to AI used in any business function. The Federal Reserve’s Jeffrey S. Allen documented the effect. Professional services rose 47%. Manufacturing rose 159%, roughly from 4.7% to 12.2%, the largest proportional jump of any sector.

Same firms, same fortnight, wider question, four times more AI.

The AI those companies were already running sat almost entirely outside the making of things. The Peterson Institute for International Economics, on the same data, states it directly: “most AI use by manufacturing firms is applied to tasks outside the production of goods, such as sales and marketing.” Census’s own working paper says where: sales and marketing 52%, strategy and business development 45%, IT 41%, with 57% of users running it in three or fewer functions.

So this is not a standing start. In these industries AI is already bought, already running, and pointed at the commercial functions rather than at the work the company gets paid for. Which means the first mover advantage here is unusually cheap: the tools are in the building, and nobody has walked them to the floor.

What is actually different about frontline work

In knowledge work, the record and the work are the same object. A lawyer’s output is the contract. An analyst’s output is the model. Point AI at that and you are pointing it at something that already exists as text, already lives in a system, already carries its context.

In physical work, the record is a second, separate job.

A quality engineer inspects a part, then writes the 8D report. A driver completes the run, then logs the hours. A technician closes the work order, then photographs the punch list and files the SLA report. A recruiter fills the role, then rewrites the job order for a dozen client VMS portals. Every one of those is a person who has already finished the work, describing it again, under time pressure, into a system designed to satisfy somebody else.

The only place anyone has counted this properly is healthcare, where a 2021 review in the Journal of the American Medical Informatics Association, covering 35 studies, found physicians spending twice as much time on documentation and clerical tasks as on direct patient care, and nurses spending more than half of every shift entering and retrieving data.

Nobody has built that instrument for a shop floor, a job site or a loading dock. Which is why the cost of it never appears in anyone’s plan.

What it costs, in people

The single number anybody has published on this: OFS surveyed 500 manufacturing professionals in June 2026 and found operators losing an average of 21 minutes a shift to manually logging information, transcribing data and preparing reports. Its CEO James Magee made the obvious point: “21 minutes does not sound dramatic until you scale it.”

So scale it. A hundred people on shift, 250 working days: 8,750 hours a year. Four full-time equivalents, occupied entirely with writing down work that had already been done. At 300 on shift it is closer to twelve.

Four people. Finding out whether that is actually your number, in the one department carrying most of it, starts at $15,000 and reports in six weeks.

That is operators. It excludes the quality engineer whose 8D takes most of a morning, the dispatcher rebuilding a detention claim from three systems, the coordinator chasing sign-offs across a dozen sites, the recruiter retyping the same job order into the eleventh client portal.

None of it shows up as a line item anywhere. All of it is payroll.

Worth knowing, while we are on unmeasured things: the claim that 80% of the global workforce is deskless, roughly 2.7 billion people, traces to a 2018 venture capital microsite that no longer resolves, cited by the firm that produced it. The category everyone uses to describe this work is defined by a number nobody can source. That is the state of measurement here, and it is exactly why the cost stays invisible in the businesses carrying it.

The workers are further behind than the companies

Company-level numbers hide the operative problem.

Workvivo, a Zoom company, commissioned research across 4,736 frontline and desk employees in 2026 and found:

Manufacturing is starker. PwC and the Manufacturing Institute surveyed 102 HR and operations leaders in Q3 2025 and found AI meaningfully integrated into daily operations for 8% of frontline workers and 2% of frontline leaders. Only 19% of firms offer any AI-related training. And 3% report that most suggestions coming from the frontline get implemented.

That last figure explains the rest. Decisions travel down after they are made. Almost nothing travels up.

“Rather than easing pressure on the frontline, AI often introduces new complexity by changing how decisions are made, how performance is measured, and how work is executed day to day.”

— PwC and the Manufacturing Institute, Frontline Leadership in Manufacturing’s AI Adoption

Which is what an unowned rollout looks like from the floor. A new tool arrives, the old process stays, and the person doing the work now keeps two things current instead of one. Gardner Carrick, Chief Program Officer at the Manufacturing Institute, names the fix: “As AI becomes more important to manufacturing, leaders need to offer additional training so that workers are equipped to utilize AI on the shop floor.”

One shape, five industries

This gradient resolves into a single, specific claim: in every one of these five businesses, the highest-value AI opportunity is a documentation workflow. The evidence lines up industry by industry.

For manufacturers, PwC’s 2026 operations survey found 90% of industrial products companies implementing an AI strategy and 20% saying it is embedded. The gap closes first in quality and production, where the writing is heaviest and the output is a dated document, so a before-and-after is provable in weeks.

For construction firms, JPMorganChase puts 8.9% as having ever paid for an AI service, while AGC of America and Sage’s 2026 outlook finds 61% using AI or planning to invest more. Of the firms using it, 45% put it into office and administrative work first. Not the field. The paperwork.

For carriers and warehouse operators, at 5.4% the lowest measured anywhere, the money is unambiguously in the record rather than the driving. The US Department of Transportation’s Inspector General found detention costing for-hire truckload drivers between $1.1 billion and $1.3 billion a year in earnings, and reported in the same audit that accurate industry-wide detention data does not exist. The loss and the missing record are the same fact.

For staffing firms, Bullhorn’s research finds two thirds experimenting with AI and 10% running it across the full workflow. The experimentation concentrates on screening resumes. The hours sit in job order copy rewritten for a dozen client portals, one recruiter hour at a time.

For facilities services providers, 65% of business leaders report using AI for facility operations, per Johnson Controls’ 2026 report, almost all of it inside a single tool. Meanwhile the US Department of Energy’s own maintenance guidance puts reactive maintenance at three times the cost of a reliability-centered program, and the federal deferred maintenance backlog rose from $171 billion to $370 billion between 2017 and 2024. The economics of missing a preventive date are federally documented. The thing that misses it is a work order nobody owns end to end.

Five industries, one shape: high-value work, recorded manually, after the fact, by the person least able to spare the time.

Where to start, and what it costs

If the bottleneck is the record, the useful question stops being which model to buy and becomes which record gets made first, by whom, and into what.

That has a practical shape, and only the first move needs deciding now.

One department, measured. One department, measured, from $15,000, in the function where the documentation load is heaviest.

What that actually involves: a week timing how the department’s core documents get produced today, two weeks configuring agents against your own systems and your own past documents, three weeks of the team using it on live work, and a written comparison at the end. Nothing new to maintain, and the decision about going further gets made on the numbers rather than on a demonstration.

One workflow, rebuilt. An embedded pod rebuilds one workflow end to end, against live production data, with a go or no-go gate before anything scales.

Both start small on purpose. In industries where the last five technology rollouts landed as extra work for the same people, a bounded first step is the proposal that earns belief.

The tools are already in your building. They are in sales, in marketing, in IT, bought and running, and nobody has walked them to the floor. The company in your sector that does it first gets a cost structure the rest cannot match on price.