Why AI Adoption Splits by Industry, and What to Do About It

Executive summary

• US Census Bureau survey data analyzed by the Peterson Institute found AI use in the information sector running near 73% among firms with more than 250 employees in early 2026, while retail, accommodation and food services showed the slowest adoption of any sector measured.
• The Peterson analysis notes a positive correlation between a sector’s average hourly earnings and its AI adoption rate, and concludes that industries employing more highly skilled workers adopted earliest.
• That correlation is real. As an explanation it is incomplete, because it describes who adopted rather than why the technology worked for them.
• The mechanism underneath is simpler: the leading sectors run on documents. Their work product is already text, already digital, already the exact material a language model consumes. The lagging sectors run on physical process and tacit judgment that nobody ever wrote down.
• A longitudinal survey of oilfield services and equipment leaders shows how large the gap between intent and result gets. In 2023 they expected over 70% of companies to be piloting or scaling generative AI by 2025. In 2025, 1% reported having reached significant scale.
• Two different problems get conflated in these industries, and only one of them is solvable by spending money. Fragmented data is an integration problem. Knowledge that was never recorded is a capture problem. Companies buy the first fix and are surprised the pilots still underwhelm.
• For a company in a low-corpus industry, this changes the sequence. The first AI investment is not a use case. It is the corpus.
• That is inconvenient and it is also the opportunity. A corpus your competitors cannot buy is the one AI asset that does not commoditize.

The split, measured

Start with the size of the gap.

Gary Clyde Hufbauer and Ye Zhang of the Peterson Institute for International Economics, writing in May 2026 and drawing on the US Census Bureau’s Business Trends and Outlook Survey, found average AI use in the information sector running at roughly 73% among firms with more than 250 employees in early 2026. Across all firm sizes, information, finance and insurance, and professional, scientific and technical services were the sectors above 30% as of late 2025. Retail, accommodation and food services adopted most slowly.

They also report a positive correlation between average hourly earnings in a sector and that sector’s adoption rate, and conclude that industries employing more highly skilled workers were among the earliest adopters.

Gartner has published a matrix showing that adoption strength varies not just by industry but by use case within each industry, so that the same capability lands very differently depending on where it is deployed. I am not going to quote specific cells from it, because I have not read the underlying research and the version I saw was a low-resolution image. The directional finding, that use-case strength is industry-dependent rather than universal, is consistent with the Census data and with what I have seen operating in physical industries.

Why the wage correlation is a description, not an explanation

The Peterson finding is solid and the correlation is almost certainly real. My argument is with what people do with it.

Read as an explanation, it implies that high-skill industries are more capable of adopting AI, or more willing. That framing leads somewhere unhelpful, because it makes the gap a question of organizational quality. The board of an industrial company reads it and concludes their people are behind, and the next move is training, a hiring push, or an innovation program.

Those responses assume the constraint is capability. In my experience running commercial operations in feed, food and industrial equipment, it usually is not. The people in those businesses are not less capable of using a tool. They are trying to point a tool at a body of knowledge that does not exist in a form the tool can read.

Consider what the leading sectors actually do all day.

Information, finance, insurance and professional services produce text as their work product. Contracts, filings, policies, claims, research notes, code, memos, case files. That output is already written, already structured enough, already digital, and already stored. When a language model arrives, the corpus is sitting there. It was created as a byproduct of doing the job.

Now consider a feed mill, a dealer network, a fleet operation, a farm, a construction site, a processing plant.

The decisive knowledge in those businesses is: which supplier ships late in the rainy season, which machine sounds wrong before it fails, which customer pays on time when the market turns, how a specific field behaves in a dry year, which dealer will actually move inventory versus park it. That knowledge is real, it is valuable, and it is expensive to be without.

None of it is written down. It lives in the heads of people who have been there fifteen years, and the only record of it is the decision they made, not the reasoning behind it.

The gap is not skill. It is corpus.

The corpus test

One question tells you which side of the split a given use case sits on.

If I gave a capable new hire access to everything we have recorded, could they reach the right answer?

If yes, the corpus exists and AI can work on it now. The knowledge is in your systems, your documents, your transcripts, your records.

If no, the corpus does not exist, and no model will fix that. The new hire would have to go ask someone, watch the process, or spend a season learning it. A model has the same problem and, unlike the new hire, it will not tell you it is guessing.

Run that question across ten decisions your business makes weekly and you will get an honest map of where AI can help you this quarter, and where it cannot until something changes.

This is the same principle that decides which sales tasks AI improves and which it degrades, described in where AI helps in sales and where it hurts. At the task level the question is whether the answer is in the input. At the industry level it is whether an input exists at all.

Two problems that look the same

There is a distinction inside all of this that is easy to miss and expensive to miss, and one industry’s data shows it clearly.

McKinsey has surveyed oilfield services and equipment leaders on generative AI twice, in 2023 and again in 2025, and published the comparison in April 2026. The 2023 respondents expected more than 70% of companies to be piloting or scaling the technology by 2025. When the same population was asked in 2025, fewer than 25% had progressed beyond pilots, and 1% reported having achieved significant scale.

Intent was not the problem. Over three quarters of those leaders believed the technology would deliver operational efficiencies, and 65% said they would put a hypothetical billion dollars into digital infrastructure and AI capability ahead of anything else.

Asked what was stopping them, more than half named data fragmentation and legacy system integration.

That answer is true and it is only half the picture, because it describes a different problem from the one this article has been making.

Fragmented data is data that exists. It sits in six systems that do not speak to each other, in incompatible formats, behind different owners. That is an integration problem. It is expensive, it is tedious, and it is entirely solvable with money, time and a competent team. Vendors sell this, which is why it is the barrier leaders name.

A missing corpus is knowledge that was never written anywhere. No integration project reaches it, because there is nothing to integrate. It is solvable only by starting to record, and then waiting.

Both are real, both are present in the same companies, and they feel identical from the executive floor: “our data is not ready.” The difference matters because the fixes have nothing in common and the timelines differ by years.

A company that diagnoses a capture problem as an integration problem will fund a data platform, integrate everything that exists, and discover at the end that the decisions it most wanted to improve are still not supported, because the reasoning behind those decisions was never in any of the six systems.

Do the integration work. It is necessary and it pays for itself. Just do not expect it to produce a corpus that never existed, and start the capture in parallel rather than after.

The uncomfortable sequence

If most of your decisions fail the corpus test, the standard AI roadmap is wrong for you.

The standard roadmap says pick a high-value use case, run a pilot, measure, scale. That works when the corpus is already there, which is why it was written by and for companies where it is.

In a low-corpus business, running that roadmap produces a familiar sequence. The pilot performs adequately on the one process that happens to be documented, usually something administrative. It fails or underwhelms everywhere the real money is. Leadership concludes AI is overhyped for their industry, and the program quietly ends. The conclusion is wrong, but the evidence that produced it was real.

The correct first investment is capture. Not a use case. Not a platform. Not an agent.

Capture means starting to record the reasoning behind decisions that currently leave no trace:

• Why was this supplier chosen over that one, in the words of the person who chose
• What did the technician actually observe before calling for a shutdown
• Why was this deal priced below list
• What did the customer say in the visit, not just what was entered in the CRM
• Which conditions preceded the last four quality failures

None of that requires an AI project. It requires a form, a voice memo, a required field, a five-minute debrief that produces text instead of nothing. The cost is process discipline, which is why it is unpopular, and it is also why it works.

Why this is the better position to be in

Here is the part that gets missed, and it is the reason this is a commercial argument rather than a technology one.

In the document-rich sectors, the corpus is largely commodity. Every bank has similar contracts. Every firm has similar filings. Models were trained on enormous quantities of exactly that kind of text. When a capability arrives, competitors get it at the same time, at the same price, with similar results. The advantage is real and it is brief, which is the pattern I described in why AI spending rarely reaches earnings.

In a physical industry, the corpus you build is yours. Nobody else has fifteen years of your dealers’ behavior, your equipment failure conditions, your regional demand patterns, your customers’ payment behavior through two commodity cycles. A competitor cannot buy that from a vendor, because no vendor has it.

That is the only category of AI asset that does not commoditize on the vendor’s release schedule. It maps exactly onto the quadrant worth serious capital in the build, buy or partner decision: a capability a competitor cannot buy, applied to something that changes how you make money.

Being late, in this specific case, is not the same as being behind. The sectors that adopted first got a fast, shallow, competed-away advantage. The sectors adopting later have the option of a slow, deep, defensible one. Very few of them are taking it, because building a corpus looks like administrative work and buying a platform looks like strategy.

What to do in the next quarter

Run the corpus test on ten decisions. The ones that actually move money in your business. Sort them into corpus-exists and corpus-missing. This is a one-afternoon exercise with your operating leaders and it usually surprises people.

Take the corpus-exists list and act on it now. Those use cases are available, the standard roadmap applies, and the wins fund the rest of the work. In most industrial businesses this list is smaller than expected and it is not empty.

Take the top three from the corpus-missing list and start capture. Choose the three where the knowledge is most valuable and most concentrated in the fewest heads. Concentration is the risk measure: knowledge held by one person retiring in three years is both the most valuable to capture and the most likely to be lost.

Set a capture standard, not a capture project. A required field, a structured debrief, a voice note after a site visit that gets transcribed. Something that produces text as a byproduct of work already being done, in the way that a contract is a byproduct of a deal.

Give it eighteen months before expecting a model to use it. This is the part that requires a board that can hold a position. A corpus is an asset that accrues, and the first twelve months produce data rather than results. Say that out loud at the start, because a program that promises results in quarter two will be killed in quarter three.

Five ways this goes wrong

1. Diagnosing a corpus problem as a capability problem. Training people to use AI better does not help when the thing they need is not recorded anywhere. It produces frustration and a conclusion that the technology does not work here. The same oilfield services survey shows how strong this instinct is: 80% of executives rated themselves mostly to fully ready to adopt the technology, and those same executives judged 90% of their frontline employees to be slightly ready or not ready. McKinsey’s broader workplace research points the other way, finding that the number of employees using generative AI for a third or more of their work is roughly three times what their leaders believe, and that C-suite leaders are 2.4 times more likely to name employee readiness as the barrier than to name their own leadership alignment. The workforce is usually further along than the executive floor thinks, and the constraint is usually somewhere else.
2. Buying a platform to solve it. Platforms are excellent at operating on a corpus and cannot create one. A vendor whose demo works on your data is showing you the corpus-exists list, which is the smaller half.
3. Capturing everything. Recording all decisions produces volume, not signal, and it collapses under its own process weight within a quarter. Three high-value decision types, done consistently, beats everything done briefly.
4. Capturing outcomes without reasoning. Your systems already store what was decided. The missing asset is why. A field that records the decision adds nothing you do not have.
5. Expecting the adoption gap to close on its own as models improve. Better models read a corpus better. They do not conjure one. The gap between document industries and physical industries narrows only where somebody deliberately builds the missing record.

The shape of it

The published data shows a wide and persistent split in AI adoption between sectors, and the honest reading of that split is not that some industries are behind.

It is that a language model is a machine for operating on recorded knowledge, and industries differ enormously in how much of their knowledge is recorded. Where the corpus already existed, adoption was fast and the advantage was shared. Where it does not, adoption looks slow, and the companies that build the record will hold something their competitors cannot purchase.

For an industrial, distribution or agribusiness company, that is a better strategic position than the adoption statistics suggest. It requires accepting an eighteen-month horizon in a market that is being sold on ninety-day pilots.

Frequently asked questions

Which industries have the highest AI adoption?

Information, finance and insurance, and professional, scientific and technical services lead. US Census Bureau survey data analyzed by the Peterson Institute put AI use in the information sector near 73% among firms with more than 250 employees in early 2026, with those three sectors above 30% across all firm sizes as of late 2025. Retail, accommodation and food services adopted most slowly.

Why is AI adoption lower in manufacturing, energy and agriculture?

The common explanation is workforce skill, following the observed correlation between sector wages and adoption. A more useful explanation is that the decisive knowledge in physical industries was never written down. Language models operate on recorded knowledge, and in these industries the critical expertise lives in operator judgment and field experience rather than in documents.

What is the corpus test?

One question, asked about any decision: if I gave a capable new hire access to everything we have recorded, could they reach the right answer? If yes, AI can work on that decision now. If no, the record does not exist, and no model will substitute for it.

Should an industrial company wait for better AI?

No, but it should sequence differently. Act now on the decisions where the record already exists, which is usually a smaller list than expected but not an empty one. In parallel, start capturing the reasoning behind the highest-value decisions that currently leave no trace. Better models will read that record more effectively. They will not create it.

Our data is fragmented across systems. Is that the same problem?

No, and conflating the two is expensive. Fragmented data exists and needs integration, which is costly but entirely solvable and is what vendors sell. A missing corpus is knowledge that was never recorded anywhere, and no integration project reaches it. Both are usually present in the same company and both feel like “our data is not ready” from the executive floor. Do the integration, and start capture in parallel rather than after it, because the timelines differ by years.

Is being late to AI a disadvantage in a physical industry?

Not necessarily. In document-heavy sectors the corpus is largely commodity, so competitors gained similar capability at similar times and the advantage compressed quickly. In a physical industry, a corpus built from your own operations cannot be purchased by a competitor from any vendor. That is a slower and more defensible position, and very few companies are pursuing it.

Sources

1. Gary Clyde Hufbauer and Ye Zhang, “The adoption of AI by industrial sectors,” Peterson Institute for International Economics, May 21, 2026. Drawing on the US Census Bureau Business Trends and Outlook Survey, with supplementary Federal Reserve and NBER data. piie.com
2. US Census Bureau, Business Trends and Outlook Survey (BTOS), the underlying data source for the above.
3. Bill Ambrose and Spandan, with Priyank Singh and Sri Kandala, “Gen AI in the OFSE industry: Progress lags behind intent,” McKinsey and Company, April 1, 2026. Data from the OFSE Leaders Gen AI Survey, 2023 and 2025 editions. mckinsey.com
4. McKinsey and Company, “Superagency in the workplace: Empowering people to unlock AI’s full potential,” 2025. Cited for the gap between employee generative AI use and leadership perception of it.
5. Gartner, “GenAI use-case adoption trends across industries,” referenced for the directional finding that use-case adoption strength varies by industry. No specific figures from this research are cited. See the note in the text.

About the author

Andre Magrini is a chief revenue officer and fractional CRO based in the Greater Chicago Area. He led North America for Ag Growth International and, as general manager in Brazil, scaled an operation more than 4x in three years. He served as Vice President of the Marketing and Communications Committee at the American Feed Industry Association, and is the author of seven books, including five on sales, marketing analytics and corporate governance.