AI Adoption by Industry: Banking 33.9%, Healthcare 20%, and Why the Gap Persists

Finance and insurance firms reported 33.9% AI use in the Census collection period ending 3 May 2026. Health care sat near 20%, against a national rate of 19.8%. So the real finding in AI adoption by industry is not that healthcare trails badly. It is that banking is the outlier above the line, and healthcare is close to ordinary.

The comparable numbers, and what they actually measure

Only one source measures both sectors the same way. The Census Bureau's Business Trends and Outlook Survey asks a large, repeated sample of US firms whether they used AI to produce goods or services in the past two weeks.

That question is blunt. It is also consistent, which is the property that matters when you want to compare two industries.

In the collection period ending 3 May 2026, information led at 39.7% and finance and insurance followed at 33.9%, against a national rate of 19.8%. Retail trade sat near 14%.

The Minneapolis Fed's read of the same survey puts health care at roughly 20%, with manufacturing near 12% and construction, transport and food service all under 10%. Healthcare is not at the bottom of that list. It is in the middle, sitting almost exactly on the national average.

That reframing is the whole post. The interesting question is not why healthcare is behind. It is why finance is ahead of everything except pure information work.

Where the 47% comes from, and why it is not the healthcare comparison

The figure circulating as proof of a banking lead is 47%. It is a real number from a real survey. It is not measuring what a sector comparison needs it to measure.

What EY measured

EY-Parthenon fielded a generative AI survey to 100 senior decision-makers across 50 retail and 50 commercial banks. Of those respondents, 47% reported having rolled out generative AI applications, against 10% in 2023. A wider 77% had launched or soft-launched something.

Every respondent is a bank large enough to have a senior decision-maker with a strategy remit. That is a survey of the top of the industry, and it answers a good question. It does not answer "what share of firms in this sector use AI".

What the Census measures

BTOS samples firms of every size, including the thousands of small credit unions, insurance brokers and independent advisers that also sit inside finance and insurance. The Census working paper on AI diffusion found 18% of firms using AI, rising to 32% on an employment-weighted basis. Weighting by employment nearly doubles the number, because large firms adopt and small firms do not.

So the 47% and the 33.9% are not in conflict. They are the same industry measured at two different cut points, one of which excludes almost everyone.

Where the 18% probably came from

The pairing of 47% for banking and 18% for healthcare circulates widely. I could not verify any survey reporting 18% specifically for healthcare. What I could verify is that 18% is the Census figure for all US firms in the 2026 AI supplement.

My read is that a national rate has been relabelled as a healthcare rate somewhere upstream, and then repeated. This is the failure the source-checking rule exists to catch. Six blog posts quoting the same number is one source, not six.

Only rows two and four are directly comparable. Row one is a different population. Row three is the national baseline, not a sector.

Banking had a model governance regime before AI had a name

Now to the part of the gap that is real. The first condition is regulatory clarity, and banking's version of it is fifteen years old.

SR 11-7 predates the term AI governance

The Federal Reserve and the OCC issued supervisory guidance on model risk management in 2011. It defines a model broadly, as any quantitative method that turns input data into estimates. A machine learning system meets that definition without argument.

So when a bank wants to deploy a model, the questions are already written down. Who validates it. Who owns it. What the monitoring cadence is. What happens when performance drifts.

That is an enormous head start, and it is almost never the thing people credit when they explain banking's lead. The usual explanation is budget. Budget is not scarce in healthcare either at the system level. Process is what was scarce.

A bank deploying its first generative model in 2024 was extending an existing control framework. A hospital doing the same thing was writing one. The same is true of the human oversight requirements now appearing in law, which map closely onto the architecture choices behind keeping a person in the loop.

Healthcare's rules arrived in 2026, and arrived by state

Healthcare got its regulatory clarity late and got it fragmented. The FDA published guidance in January 2026 reducing oversight of certain low-risk digital health products, while confirming that software influencing clinical judgement will generally need review.

State legislatures moved faster than the federal government. Holland & Knight's review of 2026 sessions records continued state activity regulating AI in healthcare, with prior authorisation and mandatory clinician oversight as the recurring themes. A common bill structure requires that a licensed clinician actively supervise any AI used in clinical decision support.

Read that as an operator, not a lawyer. A national health system now has to satisfy a different rule set in every state it operates in. Banking's federal framework applies once. Healthcare's compliance surface multiplies with geography, which is a cost that scales the wrong way. Firms already tracking the EU transparency obligations landing on AI systems will recognise the shape of that problem.

The data structure gap does most of the remaining work

Regulatory clarity tells you whether you are allowed to build. Data structure tells you whether you can.

The Minneapolis Fed attributes sector variation to how far existing workflows are digitised and how much data is available. Banking scores well on both by construction. A transaction is a structured record with a timestamp, an amount, a counterparty and a code, produced by a system that has to reconcile.

A hospital network runs several record systems

Healthcare produces enormous volumes of data and very little of it is normalised across the organisation. A hospital network typically runs multiple electronic health record instances across its facilities, with differing data models and inconsistent clinical coding. Acquisitions add more.

The consequence is specific rather than vague. A model trained on one facility's records does not transfer cleanly to another facility in the same network. The integration work has to be done before any of the AI work starts, and it is the sequence covered in the order in which data remediation actually has to happen.

Remediation costs money healthcare does not have spare

This is where the argument becomes financial rather than technical. Kaufman Hall put the adjusted year-to-date hospital operating margin at 1.3% at the close of 2025, with performance bifurcating sharply between large academic systems and smaller community hospitals.

A multi-year data integration programme is straightforward to approve at a 20% margin and hard to approve at 1.3%. Banking's lead is partly a story about who could fund the unglamorous prerequisite. That is the least discussed and probably the most decisive of the three conditions.

Where healthcare did adopt, and why it was documents

Healthcare's adoption is not evenly thin. It is concentrated in one category, and that concentration is informative.

Ambient clinical documentation moved from pilot to scale faster than any other healthcare AI category. Epic, Oracle Health and athenahealth are shipping native capabilities alongside third-party tools, and vendors including Abridge are deployed across large academic systems.

The pattern holds because documentation clears every barrier at once. The output is a draft a clinician reviews and signs, so the human oversight requirement is satisfied by the workflow itself. The input is a conversation, not a fragmented record, so the data problem does not apply. The saving is clinician time, which is the most expensive input in the building.

Compare that to the Census function data for adopters generally, where sales and marketing leads at 52%, strategy at 45% and IT at 41%. Most sectors adopted AI where the work is text. Healthcare did the same thing. It just has less text work per employee than a bank does, because a large share of its labour is physical.

My view is that this makes the sector comparison less useful than people want it to be. Adoption rate is tracking the share of a workforce doing document-shaped work, more than it is tracking sector ambition or competence.

The classification artefact that widens the gap on paper

One more mechanical factor, and it is rarely mentioned. Federal Reserve analysis notes that healthcare's measured adoption is depressed partly by industry classification. Pharmaceutical and life sciences firms are classified under manufacturing and professional services, not under health care.

Those are among the most AI-intensive organisations in the economy. Every one of them counted somewhere else lowers the health care number and raises another sector's.

The same analysis points to firm size composition. Industries made up of larger firms report higher adoption, because adoption correlates strongly with headcount. Health care and social assistance contains an enormous tail of small practices, home care agencies and social service providers. Finance and insurance has a tail too, but a smaller and better capitalised one.

None of this makes the gap fictional. It does mean a meaningful slice of it is an accounting property of the survey rather than a behavioural difference between sectors.

Where this argument is weakest

Three problems with what I have just written, stated plainly.

Adoption rate is not value

Every figure in this post counts whether a firm used AI. None of them counts whether it worked. A sector could report 40% adoption and negative return, and the survey would look identical to a sector reporting 40% adoption and real savings.

Banking's own numbers hint at this. Only 28% of banking automation use cases in development or implemented employ generative or agentic AI, per the same EY survey. Deployment breadth and deployment depth are different things, and the distinction is the subject of the analysis of where AI deployments returned less than they cost.

A different survey says 75% of health systems use AI

This is the strongest counter to my framing. Eliciting Insights reported in 2026 that 75% of US health systems use at least one AI application, up from 59% a year earlier. Read against that, calling healthcare "average" looks wrong.

Both can be true. Health systems are large organisations, and the Census figure is firm-weighted across a sector dominated by small providers. A survey of health systems is closer in spirit to the EY survey of banks than to BTOS. The honest position is that healthcare's large institutions look a lot like banking's, and its long tail does not.

If you take that seriously, the sector gap shrinks further, and my third condition about margins gets stronger rather than weaker. Small providers cannot fund this. Large ones can.

Two of my figures are read off charts. The health care figure of about 20% and the accommodation and food figure are read from published Federal Reserve charts, not from a tabulated release. They are directionally reliable and they are not precise. I would not build a business case on the second decimal place of either.

What would actually close the gap

The useful version of this question is not "when will healthcare catch up". It is "which specific condition has to change, and what would you see when it does".

The fourth row is the one I would watch. The regulatory and data conditions get discussed constantly. The financing condition decides whether the tail of the sector ever moves, and nobody has solved it.

If you sell into either vertical, the practical read is narrower. In banking you are selling into an organisation that already has a validation process and will apply it to you. Expect model documentation requests and a longer procurement cycle, and treat passing that review as the product requirement it is. The gates involved look a lot like the tests that separate a proof of concept from a production system.

In healthcare you are selling into an organisation whose blocker is usually upstream of you. If your deployment depends on clean data from several record systems, you are quoting a project the buyer has not budgeted. The vendors winning in healthcare picked workflows where the data arrives with the task, which is exactly why documentation went first.

Frequently asked questions

Where to start this week

Two things, and the first takes twenty minutes.

Pull the last three AI adoption statistics you have quoted in a deck or a board paper. For each one, write down the sample, the question and the period. If you cannot find all three within a few minutes, take the number out. That single test would have caught the 47 against 18 comparison this post started from.

Then pick the condition that binds you rather than the one that binds your sector. If you are in healthcare, that is almost certainly data integration cost, and the answer is to find a workflow where the input arrives with the task. If you are in banking, it is more likely to be depth. Broad rollout with 28% of automation use cases touching AI is a starting position, not a finished one, and the question of where measurable return has actually shown up is still the one worth answering.

Related reading