Where AI has not landed: 77% report no use case

TL;DR. Eleven articles in this series covered domains where AI is deployed. This one covers the ones where it is not, because the pattern is more informative than any single deployment. Census data puts transportation at 7.5% adoption, accommodation and food at 8.3%, and construction at 9.5%, against roughly 73% for large information-sector firms. In some sectors, including utilities and construction, adoption has declined during 2026. The most common reason businesses in those sectors give is not cost, regulation or skills. It is that no use case applies, reported by 77% of small businesses concentrated in construction, food service and the trades. The property those domains share is not that the work is physical. It is that the output is not a document.

---

Eleven articles into this series, a question worth asking is where AI has not gone.

The numbers are unambiguous. United States Census Bureau survey data from May 2026 puts overall business AI use at 19.5%, and the distribution is extreme. Transportation reports 7.5%. Accommodation and food service, 8.3%. Construction, 9.5%. Agriculture and construction have both been recorded near 1% in earlier rounds.

At the other end, average AI use among firms with more than 250 employees in the information sector reached roughly 73% in early 2026.

And in some sectors, including utilities and construction, adoption has declined during 2026. That is the more interesting number, because it means firms tried and stopped.

The reason given is the part worth sitting with. Asked why, the most common answer from small businesses in these sectors is not that the technology is too expensive, too risky or too complicated. It is that no use case applies: 77% of small businesses report exactly that, and the response is concentrated in construction, food service, the skilled trades and local services.

The property those domains share

The obvious explanation is that the work is physical, and the obvious explanation is not quite right. Healthcare is physical and has high adoption. Manufacturing is physical and has substantial adoption. Transportation is physical and sits at 7.5%.

The distinguishing property is what the work produces.

Look at the eleven domains covered in this series. Medicine produces notes, images and reports. Law produces filings. Education produces explanations and assessments. Journalism produces articles. Software produces code. Translation produces text. Customer service produces messages. Hiring produces rankings and decisions on paper. Government produces records. Finance produces scores and documents. Science produces papers and predictions.

Every one of them has a symbolic artifact as its output.

Now the low-adoption list. Construction produces a building. Transportation produces a delivery. Food service produces a meal. Agriculture produces a crop. The trades produce a repair.

AI has landed where the artifact is symbolic and has not landed where the artifact is physical, and this explains the apparent exceptions. Healthcare adopts heavily because the documentation surrounding care is symbolic even though the care is not, which is why the most deployed medical AI application is ambient note-taking rather than anything clinical. Manufacturing adopts where the work is scheduling, inspection and quality records, and not where it is assembly.

The correct statement is not that AI cannot do physical work. It is that AI produces symbols, so it has been adopted wherever the valuable output of a job is already a symbol.

What "no applicable use case" actually means

This is the most honest signal in the data and it deserves to be taken at face value rather than reinterpreted as ignorance.

A plumber diagnosing a leak in a wall is doing something with a large tacit component, an irreducible physical step and an outcome verified by whether water stops. No part of that produces a document that anyone needs. There is an invoice, and invoicing software has existed for thirty years.

A site foreman coordinating six trades against a delivery schedule has a genuine information problem, and it is a scheduling problem with hard physical constraints, not a text-generation problem. The tools that would help are the ones that have been sold to construction for two decades with limited uptake, for reasons that predate AI entirely.

When a sector reports no applicable use case, the most likely explanation is that they have looked and there is not one, at a price and reliability that makes sense for them. Treating that as a failure of imagination on their part is the assumption that has produced most of the pilots that quietly ended.

The declining adoption figures are the same finding from the other direction. Sectors that tried and reduced usage did not fail to understand the technology. They evaluated it against their work and stopped.

The size divide, which is larger than the sector divide

Worth separating, because it is frequently confused with the sector effect.

Adoption at firms with 250 or more employees runs at 36.1%, and in the information sector at that size it reaches roughly 73%. Small firms in the same sectors adopt far less.

Three mechanisms explain most of it, and none is about the technology.

Fixed costs. Evaluating a tool, changing a process and training people costs roughly the same whether you have twenty employees or two thousand. The return scales with headcount and the cost does not.

Someone whose job it is. Large firms have people who assess and deploy tools. A twelve-person contractor does not, and the owner evaluating AI is the person also doing the estimating.

And structured data. The sectors with high adoption already had digitised, structured records because they had already been through an earlier wave of software. The sectors with low adoption frequently still work from paper, photographs and phone calls, and the prerequisite for AI is not AI.

So a large share of what looks like sector resistance is a small-business effect wearing a sector's clothing, and the two are hard to separate because the low-adoption sectors are also the ones with the most small firms.

The employment correlation, handled carefully

One 2026 analysis compared sector-level adoption against employment change and found that job growth appeared negatively correlated with AI adoption. Construction, healthcare, transportation and hospitality added jobs. Information and financial activities shed them.

This is early, correlational, and confounded, and it should not be read as AI causing job losses. Those sectors differ in interest-rate exposure, post-pandemic recovery position and cyclical demand, any of which could produce the same pattern.

The analysis itself flagged two outliers that are more interesting than the headline. Healthcare and professional services had high adoption and outperformed the hiring trend, and the proposed explanation is that jobs in those sectors are high-dimensional: AI augments a wide bundle of tasks rather than replacing a narrow one, which raises productivity in a way that supports more hiring rather than less.

If that mechanism is real, the exposure is not about how much AI a sector adopts. It is about how narrow the jobs are. A role consisting of one repeated symbolic task is exposed. A role consisting of thirty different things, several of them physical, is not, regardless of how much AI the sector buys.

That is a considerably more useful frame than counting adoption rates, and it is a hypothesis rather than a finding.

What this means for the eleven domains

Reading the series backwards from here changes the emphasis.

The domains where AI has landed are the domains that were already document factories, and much of what has been achieved is the automation of documentation rather than of the underlying work. The clearest case is medicine: the deployed success is ambient note-taking, and diagnosis is not deployed at all.

Which reframes the productivity findings. If AI mostly automates the symbolic layer around work, then the software result makes sense: coding is a symbolic task and it got faster, while the whole job includes design, review, coordination and incident response, and overall output moved about 10%. The same shape should be expected wherever the symbolic layer is a minority of the job.

And it suggests where the next wave lands, if there is one. Not in construction or the trades directly, but in the paperwork attached to them: permits, inspections, compliance records, insurance claims, scheduling. That is symbolic output produced by physical industries, and it is the largest untouched surface visible in the data.

What twelve articles found, in one table

Territory 5 covered eleven domains where AI is deployed and one where it is not. Read together, the measurement problem is more consistent than the technology.

Medicine. 1,524 cleared devices, 1.6% citing a randomised trial, under 1% reporting patient outcomes, about ten reimbursed. Clearance certifies resemblance to an existing product.

Law. 1,313 documented court proceedings involving fabricated content. Not because law is worse, but because an adversary reads every filing.

Education. Satisfaction 0.93 and confidence 0.91 against knowledge 0.53, with the authors rating certainty of all three as very low.

Hiring. Eighteen bias audits across 391 employers, nearly all passing, under a law that lets employers decide whether they are in scope.

Science. A Nobel Prize, an unchanged experimental rate, and 736 synthesised compounds against 380,000 predicted.

Finance. The only domain with governance predating AI, whose regulator looked at generative systems in 2026 and placed them outside scope.

Customer service. 90% deflection reported against 40% resolution, with every party to the measurement benefiting from the same answer.

Government. 126 use cases, 65 not public, and an inventory the auditors found incomplete despite being a legal requirement.

Journalism. 45% of AI news answers carrying a significant issue, from the one field that audited the technology rather than adopting it.

Translation. Fifty years of evaluation, ending in a shared task titled "Stop using BLEU" and a measurability ceiling.

Software. 19% slower measured, 20% faster reported, from developers on their own code.

And the trades. 77% reporting no applicable use case, which is the only domain in the series where the people involved were asked directly and answered plainly.

The pattern across all twelve is not about capability. It is that each domain measured AI with an instrument built for something else, and the two domains that produced the clearest evidence did so for structural reasons rather than deliberate ones: law because it has an adversary, and software because its subjects can read the study.

What is unresolved

Whether robotics changes the boundary. The argument above is about a system that produces symbols. Systems that produce physical actions are a different proposition, and essentially none of the adoption data describes them.

Whether the small-firm gap closes. If the fixed cost of adoption falls far enough, the size divide should narrow. If the binding constraint is structured data or someone to run it, cheaper tools will not help.

Whether the employment correlation survives. It is one quarter of one analysis using survey-derived adoption rates. It could be an artifact of sector cyclicality, and it will take several more quarters to distinguish.

And whether "no applicable use case" is stable. It is currently the most common answer in the low-adoption sectors. Whether that reflects a durable property of the work or a temporary state of the tools is exactly the question nobody can answer from adoption data.

The counter-argument

Low adoption today is not evidence of an inherent limit. Every general-purpose technology reached labour-intensive sectors late. Electrification took decades to move from factories to farms, and reading a 9.5% construction figure as a property of construction rather than a point on a curve is the mistake that history usually punishes.

The survey measures the wrong thing. Asking a firm whether it uses AI captures deliberate organisational adoption and misses a foreman using an assistant on his phone to write a client email. Shadow usage is real, unmeasured, and probably larger in exactly the sectors reporting low official adoption.

The symbolic-output argument may be circular. AI has been deployed where its output is useful, and its output is symbols; observing that it landed in symbolic domains risks restating the definition rather than explaining anything. The test is predictive: if it is a real constraint, physical-output sectors should stay low as tools improve, and if it is circular, they will not.

And "no applicable use case" may reflect the tools rather than the work. The people answering are describing products currently offered to them, most of which are chat interfaces and document assistants sold by firms with no understanding of their sector. A tool built for scheduling trades against weather and material deliveries might find a use case that does not currently exist because nobody built it.

The short version

Census data puts overall US business AI use at 19.5%, with transportation at 7.5%, accommodation and food at 8.3%, and construction at 9.5%, against roughly 73% for large information-sector firms. In utilities and construction, adoption has declined during 2026, meaning firms tried and stopped.

The reason given is the most useful datum in this article. It is not cost, regulation or skills. It is that no use case applies, reported by 77% of small businesses, concentrated in construction, food service and the trades.

The property the untouched domains share is not that the work is physical. It is that the output is not a document. Every domain in this series where AI is deployed produces a symbolic artifact: notes, filings, explanations, articles, code, messages, rankings, records. The low-adoption sectors produce a building, a delivery, a meal, a crop, a repair. This explains the apparent exceptions, since healthcare adopts heavily around its documentation while its clinical work stays manual.

AI produces symbols, and it has been adopted wherever the valuable output of a job was already a symbol.

A large share of the sector effect is also a size effect. Adoption reaches 36.1% at firms of 250 or more and far less below that, driven by fixed evaluation costs that do not scale down, the absence of anyone whose job it is, and the fact that low-adoption sectors frequently lack the structured data that is a prerequisite rather than a consequence.

And the most interesting hypothesis in the data concerns narrowness rather than adoption. One analysis found employment growth negatively correlated with sector adoption, with healthcare and professional services as outliers that adopted heavily and hired anyway. The proposed explanation is that their jobs are high-dimensional, so AI augments a wide bundle of tasks rather than replacing a narrow one. If that holds, exposure is a property of how narrow a job is, not of how much AI its industry buys.

Common questions

Which industries have the lowest AI adoption? Transportation at 7.5%, accommodation and food service at 8.3%, and construction at 9.5%, according to Census Bureau survey data from May 2026, against an all-business average of 19.5%. Agriculture and construction have been recorded near 1% in earlier rounds. For comparison, average use among firms with more than 250 employees in the information sector reached roughly 73% in early 2026.

Why do some industries not use AI? The most common reason given is not cost, regulation or lack of skills. It is that no use case applies, reported by 77% of small businesses and concentrated in construction, food service, the skilled trades and local services. That answer is worth taking at face value: those firms have looked at what is on offer and concluded it does not address their work at a price and reliability that makes sense.

Is AI adoption actually falling anywhere? Yes. In some sectors, including utilities and construction, recorded AI usage declined during 2026. That is a more informative figure than low adoption, because it means firms tried the technology, evaluated it against their work, and reduced usage rather than never starting.

Why has AI reached office work but not the trades? Because AI produces symbols, and it has been adopted wherever the valuable output of a job was already symbolic. Notes, filings, code, messages and reports are all things a model can produce directly. A building, a delivery, a meal and a repair are not. This explains the apparent exceptions too: healthcare adopts heavily around its documentation while its clinical work remains manual, and the most deployed medical AI application is ambient note-taking.

Is low AI adoption about company size rather than industry? Substantially, and the two are hard to separate because the low-adoption sectors have the most small firms. Adoption runs at 36.1% among firms with 250 or more employees and far lower below that. Three mechanisms explain most of it: the fixed cost of evaluating and deploying a tool does not scale down, small firms have nobody whose job it is, and low-adoption sectors frequently lack the structured digital records that are a prerequisite rather than a result.

Does AI adoption cause job losses? The available evidence does not support that claim. One 2026 analysis found employment growth negatively correlated with sector adoption, with construction, healthcare, transportation and hospitality adding jobs while information and financial activities shed them. It is early, correlational and confounded by interest-rate exposure and cyclical demand. The analysis itself flagged healthcare and professional services as outliers that adopted heavily and outperformed the hiring trend.

What is the high-dimensionality hypothesis? The proposal that jobs consisting of many varied tasks are less exposed to AI than jobs consisting of one repeated task, regardless of how much AI the sector adopts. It was offered to explain why healthcare and professional services adopted heavily and still added jobs, on the reasoning that AI augments a wide bundle of tasks rather than replacing a narrow one. It is a hypothesis rather than a finding, and it is a more useful frame than counting adoption rates.

Where might AI reach these sectors next? Not in the physical work, but in the paperwork attached to it. Permits, inspections, compliance records, insurance claims and scheduling are symbolic outputs produced by physical industries, and they represent the largest untouched surface visible in the adoption data. That is also the pattern that already played out in medicine, where the deployed success is documentation rather than anything clinical.