See how AI adoption by industry compares in 2025 — adoption rates, leading sectors, generative and agentic AI use, ROI, and the top barriers holding firms back.
Roughly one in five European enterprises now runs at least one AI technology in production, most large US firms have moved past pilots, and the median OpenAI enterprise customer's usage grew more than 6x year-over-year. AI adoption in 2025 is landing unevenly across sectors, functions, and company sizes, and the returns are concentrating in a narrower band of firms than the headline growth rate suggests.
Who's adopting, what they're deploying, what's working, what's stuck: the industry-by-industry picture matters more than the aggregate if you're deciding where to place your next AI bet.
The State of AI Adoption in 2025
The gap between "we're exploring AI" and "AI is a line item on the P&L" widened sharply this year. Adoption is broad, but depth of use is what separates the leaders from the rest.
How Many Companies Are Actually Using AI
Eurostat's 2025 enterprise survey found that 19.95% of EU enterprises with 10 or more employees used at least one AI technology, spanning text mining, speech recognition, natural language generation, image and video generation, and related capabilities. Roughly one in five European businesses of any meaningful size now has AI in production somewhere.
The percentage of companies using AI climbs steeply with size. Among EU enterprises that hadn't yet adopted, 36.54% of large firms had considered doing so, compared to 22.26% of medium-sized businesses and 12.65% of small ones. Large enterprises are roughly three times more likely to be on the AI runway than small ones by that measure.
In the US, the picture is denser. OpenAI reports more than 1 million business customers, with AI moving from isolated pilots into workflows, products, and internal systems across most sectors.
How Fast AI Adoption Is Growing
Growth rates are what make the 2025 numbers unusual. OpenAI's data shows the median industry expanded its enterprise usage more than 6x year-over-year, growth broad enough and fast enough to distinguish this cycle from the 2018–2022 machine-learning wave, which stayed concentrated in tech and finance.
The EU is moving too, but more slowly. Eurostat reports that 14.21% of non-adopting enterprises are considering AI, and the year-over-year change against 2024 is modest. Consideration is not deployment, and Europe's lag on the latter is one of the defining features of the 2025 landscape.
AI Adoption Rates by Industry
No single industry curve fits everyone. KPMG's Q1 2026 Global AI Pulse frames sectors along two axes, AI maturity and agentic orchestration capability, and finds that sector positioning reflects AI maturity and agentic coordination capability rather than a single linear path to scale. The AI adoption rate by industry now varies by an order of magnitude between leaders and laggards.
Leaders: Technology and Financial Services
Technology, Media and Telecom lead the pack. TMT is one of the sector spotlights KPMG uses to illustrate how these dynamics manifest in practice, with mature orchestration capabilities and deep integration into core workflows. Software firms in particular have embedded AI into code generation, customer support, and internal knowledge systems.
Financial services sit right behind, and in some risk and compliance workflows ahead. Banks, insurers, and asset managers have spent a decade building the data infrastructure that AI now runs on: clean transaction data, well-governed customer records, compliance controls that make GenAI deployments defensible. Fraud detection, document processing, and analyst copilots are already generating measurable returns.
Fast Movers: Healthcare, Retail, and Insurance
AI in healthcare is scaling faster than most incumbents expected. Clinical documentation, radiology triage, and prior authorization, all functions with abundant unstructured text, are the wedge use cases. The bottleneck isn't model quality; it's integration with electronic health records and clinical workflows.
Retail is running AI hard on merchandising, search, and personalization, with generative product descriptions and image generation cutting content production costs meaningfully. Eurostat found that 34.70% of EU enterprises using AI apply it to marketing or sales, the single largest functional category. AI in financial services has a close cousin in insurance, where claims processing, underwriting, and customer service automation are all moving from pilot into production.
Laggards: Manufacturing, Construction, and Energy
AI in manufacturing is a more complicated story. The technology fits (predictive maintenance, quality inspection, supply chain optimization), but the biggest constraint is the deployment surface itself: sensors and PLCs on the plant floor, OT-locked data, integration cycles measured in quarters not sprints. Only 6.08% of AI-using EU enterprises apply AI to logistics, a fraction of the marketing figure. Construction and energy lag for similar reasons: data lives in silos, safety and regulatory review add friction, and the ROI case has to compete with heavy capital projects. These sectors are moving, but on multi-year timelines rather than multi-quarter ones.
Generative AI and Agentic AI in the Enterprise
Two waves are overlapping in 2025. Generative AI has become baseline infrastructure at most large firms; agentic AI is where the frontier work, and most of the disappointment, is happening.
Generative AI Adoption Across Business Functions
Generative AI enterprise adoption moved from IT and marketing into legal, finance, HR, and operations over the past year. EXL's 2025 US enterprise study found that GenAI has progressed rapidly, but the speed of adoption may be interrupted by talent, user adoption, and data quality obstacles. That's the pattern most CIOs describe: the easy wins landed in months, and the next tier of use cases requires cleaner data and better change management than most organizations have.
Eurostat's functional breakdown confirms the pattern: after marketing and sales at 34.70%, 31.05% of AI-using EU enterprises apply it to business administration and management. The functions with the highest GenAI penetration are the ones where the raw material is text: customer support, marketing content, legal review, internal knowledge search. Pricing, planning, and complex operations lag because they need structured data and reliable reasoning, which is where agentic systems come in.
What Agentic AI Is and Which Industries Are Piloting It
Agentic AI adoption is where 2025's real experimentation is happening. An agent is a system that can plan, call tools, take actions, and iterate, not just answer a prompt. KPMG's framing treats agentic coordination capability as one of the two axes of AI maturity, and finds it concentrated in TMT, financial services, and parts of insurance.
Building agents in production is where most teams get stuck. The abstractions (memory, tool use, orchestration strategies, evaluation) don't come with a database and an auth system. That gap is exactly what we built Powabase's agents and orchestration layer to close, with supervisor, sequential, and other execution patterns available as native primitives.
Where AI Is Used: Adoption by Business Function
The AI use cases by business function that show up most consistently across industry surveys cluster into a handful of areas:
| Function | Typical use cases | Adoption depth |
|---|---|---|
| Customer support | Chatbots, ticket triage, agent copilots | High |
| Software engineering | Code generation, review, documentation | High |
| Marketing | Content generation, personalization, SEO | High |
| Sales | Lead scoring, email drafting, call summaries | Medium-high |
| Finance & accounting | Invoice processing, forecasting, reporting | Medium |
| HR | Resume screening, internal Q&A, onboarding | Medium |
| Legal & compliance | Contract review, policy Q&A, risk flagging | Medium |
| Operations & supply chain | Demand forecasting, anomaly detection | Low-medium |
| R&D | Literature review, hypothesis generation | Varies by industry |
Customer support, engineering, and marketing are where GenAI has moved fastest because the output is text and the tolerance for imperfection is high enough for humans-in-the-loop to be efficient. Operations and R&D lag because the cost of a wrong answer is higher and the data is messier. For a deeper look at how these functions interact in production, our enterprise AI workflow automation guide walks through the highest-return patterns.
How AI Adoption Differs by Company Size
The size gap in AI adoption is the most consistent finding across every survey. Eurostat's consideration data shows 36.54% of large non-adopting enterprises weighing AI versus 12.65% of small ones, nearly a three-to-one ratio on intent alone, before any deployment gap on top. The same pattern shows up in US data.
Large enterprises pull ahead on the strength of three things: dedicated budget, dedicated headcount, and existing data infrastructure. A failed pilot at a Fortune 500 is a rounding error; at a 30-person company it's the AI budget for the year. Enterprise teams have data engineers to wire up retrieval pipelines. In a small business, that job typically falls to a founder writing Python at midnight.
AI adoption at small businesses is real but shallower, mostly off-the-shelf SaaS with AI features baked in (support tools, marketing platforms, coding assistants). That's changing as platforms collapse the stack. When you can spin up a database, RAG pipeline, auth, and an agent runtime in one project without hiring a platform team, the cost floor for a bespoke AI product drops sharply. That's the wedge for smaller companies, and it's why we bundle Postgres, retrieval, agents, and workflows into a single Powabase backend instead of five separate services.
The Business Impact and ROI of AI Adoption
AI ROI at enterprises is finally moving from anecdote to data, but the distribution is skewed. A minority of deployments generate outsized returns; a majority generate modest ones; a meaningful tail generates none. OpenAI's case evidence shows AI associated with revenue growth and improvements in customer experience across a range of operational and strategic challenges, with impact reflecting specific applications rather than a one-size-fits-all pattern.
OpenAI's enterprise report is direct on what separates the returns: the data suggest that depth of use matters, with workers and firms making more consistent use of advanced tools (reasoning models, data analysis, Custom GPTs, Projects, and APIs) pulling ahead. The ROI variable is how deeply a firm uses the tools it already pays for, not how many seats it buys.
How Companies Are Measuring Productivity Gains
The measurement problem is real. Most firms track a mix of time saved per task (minutes per ticket, contract, or report aggregated into FTE equivalents), throughput (tickets closed, documents processed, code merged), quality proxies (customer satisfaction, error rates, rework), and revenue lift on AI-assisted workflows (conversion, cross-sell, retention).
The firms with the cleanest ROI stories are the ones that instrumented workflows before rolling out AI, so they had a real baseline. The firms with the worst ones deployed first and tried to measure later.
The Biggest Barriers to AI Adoption
The barriers to AI adoption in 2025 have shifted. Model quality is rarely the blocker anymore; the constraints are organizational and infrastructural. EXL's survey names talent, user adoption, and data quality as the three most likely to slow the next phase of GenAI deployment. KPMG's sector work adds orchestration capability, the ability to coordinate multiple agents, tools, and data sources reliably, as the frontier challenge for firms trying to move beyond point solutions.
The barriers that come up most consistently:
- Data quality and access. Retrieval-augmented generation is only as good as the corpus behind it. Most enterprises underestimated the work to clean and structure their own documents.
- Talent. Not just ML engineers, but product managers who understand AI, platform engineers who can run agent infrastructure, and evaluators who can measure quality.
- Integration. Getting AI outputs back into the systems of record where decisions actually happen.
- Governance and security. SOC 2, ISO 27001, data residency, audit logs, RBAC. Table stakes for regulated industries, and often a six-month project on their own.
- Change management. Users have to adopt the tools. Deployments that skip enablement stall.
The infrastructure barrier is the one platforms can actually solve. When RAG, agents, auth, and a Postgres database with pgvector come in one project, most of the integration and security work is already done. That's the case for consolidating the stack.
Regional Differences: US vs. EU Enterprise Adoption
US and EU enterprise adoption diverge on both pace and posture. The US is deployment-heavy: OpenAI's million-plus business customers, dense penetration across TMT and financial services, and a willingness to ship AI features before governance is fully settled. The EU is more measured, with 19.95% adoption among enterprises with 10+ employees and steady but slower growth in consideration.
The gap isn't just regulatory. European enterprises weigh data residency, DPA compliance, and the EU AI Act more heavily in vendor selection, which lengthens procurement cycles but produces more durable deployments. It's also why platforms that offer regional data residency, self-hosting, and standard enterprise certifications, terms we ship with our Powabase Enterprise tier, see meaningfully different traction in the EU than pure US-cloud offerings.
What Comes Next for AI Adoption
Three shifts define the next 12 months. Agentic systems move from demos into production in TMT and financial services, and start appearing in real deployments in healthcare and insurance. The gap between leaders and laggards widens, not because laggards stall but because leaders compound. And the platform layer consolidates: the teams that shipped fastest in 2025 stopped assembling seven vendors and started building on backends where the AI primitives are native.
The practical implication for a team planning its 2026 roadmap: pick a workflow with clean data, an obvious baseline, and a user population that will actually adopt the tool. Then pick infrastructure that doesn't force you to become a platform team on the side. Firms that skip either step tend to spend 2026 rebuilding what they shipped in 2025.