How Companies Are Adopting AI

(2025 Research)

What fraction of companies globally have adopted AI in any function (pilot, partial, or full deployment)?

According to recent surveys, ~75–80 % of companies globally report that they use AI in at least one business function (even in pilot form). But under stricter definitions (deep integration, production systems), adoption rates are likely much lower and concentrate in large, digitally mature firms.

What the data say

In McKinsey’s State of AI survey, 78 % of respondents say their organizations use AI in at least one business function (i.e. pilot, partial, or full deployment). McKinsey & Company (2024)
That figure is up from 72 % reported earlier in 2024, and 55 % a year prior. McKinsey & Company
The survey defines “use of AI” broadly (from early experiments to deeply embedded systems) across functions such as IT, marketing, service operations, etc. McKinsey & Company
Another source, IDC / IDCA’s Q1 2025 survey, reports 76 % of organizations now use AI in some capacity, and 69 % use generative AI in at least one function. International Data Center Authority
In a separate report, PwC’s 2025 “AI Agent Survey” found 79 % of executives say their companies are already adopting AI agents (i.e. autonomous or semi-autonomous AI modules). PwC

Interpretation & caveats

These adoption percentages are self-reported survey data, and typically include any use (from pilots to full production). They don’t necessarily imply deep integration or broad use across all functions.
The definition of “using AI” is quite loose in many surveys (e.g. “in at least one business function”) so the bar for inclusion is low. McKinsey & Company
There is likely sampling bias: respondents to AI surveys tend to be more AI-savvy or from more digitally mature firms.
Adoption rates vary dramatically by industry, region, and firm size. For example, McElheran’s research suggests that among large U.S. firms, >50 % may be using AI in production, while many smaller firms are still well behind. MIT Sloan
In more “hard production” contexts (manufacturing, retail, logistics), the penetration is often much lower than the headline number.
Also, some government or census data in the U.S. suggest that only ~7 % of U.S. firms currently report using AI in their operations (as of recent surveys) when using stricter definitions. arXiv

What fraction of companies reports “full deployment” (not just pilots) of AI in at least one business function?

Only about 10–30% of organizations have pushed at least one AI use case into true full deployment beyond pilots, though broader surveys using looser definitions sometimes report 30–40%. Most companies are still in partial or experimental stages rather than enterprise-wide adoption.

What the evidence says

In the Riverbed Global Survey (2025), only 12 % of AI projects are reported to be fully deployed enterprise-wide (i.e. has scaled beyond pilot) in IT operations. Riverbed
The report “The GenAI Divide: State of AI in Business 2025” claims that “nearly 40 % report deployment” when considering generative AI tools (though this likely includes partial deployment or non-enterprise scale). MLQ
In PwC’s AI Agent Survey (2025), among organizations adopting AI agents:

35 % say they’re being used broadly
another 17 % say AI agents are “fully adopted in almost all workflows/functions.” PwC

Some tech-industry surveys (e.g. EY) suggest that 48 % of tech executives claim to be “already adopting or fully deploying” agentic AI (though “fully deploying” is bundled with “adopting”) in their domain. EY

Interpretation & guidance

The 12 % figure from the Riverbed survey is probably among the more conservative and rigorous estimates (enterprise-wide scale, not just departmental).
The ~35 % + 17 % split in the PwC survey suggests a hybrid: some organizations have broad deployments, but fewer have everything fully adopted across all workflows.
The ~40 % “deployment” in the GenAI Divide report is looser and likely includes partial or feature-level deployment, not full scale.
So a reasonable “headline estimate” is that somewhere between 10 % and 30–40 % of organizations have pushed at least one AI use case into full deployment (beyond pilot) in a meaningful way.

What is the typical ROI that adopters see from their AI investments?

In reported cases, many adopters see a 2× to 4× return on their AI investments (i.e. $2–$4 of value per $1 spent), especially in generative AI or analytics use cases. In high-performing cases, ROI can reach 8× to 10× or more. However, many less mature or poorly executed projects report minimal or zero P&L impact.

What studies & reports say

IDC / Microsoft, via VentureBeat
- Businesses report an average 3.5× return on AI investments (i.e. for every $1 spent, $3.50 of value) Venturebeat
- Some individual cases in that survey saw as much as 10× ROI Venturebeat
IDC (Generative AI ROI)
- In a generative AI context, some organizations report $3.70 of return per $1 spent (i.e. ~3.7×) Technology Record
- The same study notes that in some cases, ROI reaches ~10.3× for high-performing implementations Technology Record
Snowflake Research
- 92 % of early adopters say their AI investment is “paying for itself” or profitable Snowflake
- But Snowflake doesn’t publish a standard multiple in that press release; it reports prevalence, not magnitude Snowflake
Strategy Software / Microsoft-sponsored blog
- Claims “AI investments yield an impressive return of $3.50 for every $1 spent” in analytics settings (i.e. ~3.5×) Strategy
Older / more skeptical views
- ESI ThoughtLab (2020) found that many firms report very modest returns: in that sample, average returns across AI investments were only ~1.3 % (i.e. a very low “multiple” or marginal gain) for the broader population, though “AI leaders” saw higher (~4.3 %) Econsult Solutions, Inc.
- That suggests for many firms, ROI is very modest unless they are advanced in deployment.
- Also, some recent commentary (e.g. MIT/“GenAI Divide” critiques) suggest that many AI projects deliver no measurable P&L returns (i.e. 0×) in practice, particularly among less mature adopters. Blueflame AI+1

Interpretation, caveats & what to expect

Generative AI tends to show higher multiples (3× to 4×) in early adopter reporting, especially in marketing, customer support, personalization, etc.
Outliers push the upper bound (10×+), but those are less common and often in well-resourced, data-rich settings with tight execution discipline.
Many implementations show minimal returns (or even negative ROI) — especially pilots, under-resourced projects, or poorly scoped ones.
ROI depends heavily on scale, integration, execution, domain fit, and measurement rigor. A “good multiple” is much more achievable in mature, core use cases than in fringe or speculative experiments.
Time to realize value matters a lot — early returns may be modest, while compounded or cumulative gains emerge over 1–2+ years.
Also, what counts as “return” is variable: cost savings, revenue growth, improved retention, reduced error, time savings, etc. Some “returns” are soft or indirect.

What are the main barriers that firms or product teams cite to AI adoption?

AI fails or stalls for several reasons. Here is a ranking of barriers by how frequently they are cited in surveys:
1. Skills / knowledge gap / training
2. Budget / cost constraint
3. Security / integration / technological complexity
4. Immature tools / performance concerns
5. Regulatory / compliance / uncertainty
6. Cultural resistance / change management

Main Barriers to AI Adoption
Here are the most commonly reported obstacles, grouped by theme, along with sources and nuance:

Barrier / Theme	Description & nuance	Supporting Evidence / Quotes
Lack of clear strategy / vision / executive buy-in	Many companies launch AI pilots in silos without connecting them to business goals. Without a guiding AI vision or roadmap, efforts stall.	McKinsey’s Notes from the AI Frontier lists “lack of clear strategy for AI” as the top barrier. (McKinsey & Company) McKinsey in their “Gen AI’s Next Inflection Point” piece argues many organizations lag behind employee experimentation because they haven’t transformed operating models or governance. (McKinsey & Company)
Talent and skills gap	Shortage of people with AI, data science, ML ops, prompt engineering, and domain knowledge is often cited. Even when such people exist, they may not understand product use cases or integration.	Surveys indicate ~63 % of organizations cite technical skills shortage as a barrier. (LinkedIn) In AI in healthcare, respondents report “insufficient expertise or technology” as a key barrier. (PMC) McKinsey comments that early adopters are better at diagnosing and closing skill gaps; many experimenters do not. (McKinsey & Company)
Data maturity, quality, silos, and infrastructure	Even when data exists, it may be fragmented across systems, inconsistent, or poor quality. Legacy systems may not support real-time or integrated usage.	Berkeley CMR article cites “data complexity and data silos” as top barriers to adoption. (California Management Review) OECD report emphasizes that “lack of data maturity” is a fundamental barrier. (OECD) Deloitte’s AI Trends article warns foundational infrastructure gaps remain for many firms. (Deloitte)
Organizational silos, process misalignment, change resistance	AI often requires reworking business processes (not just dropping in a model). Silos between IT / data / business units impede end-to-end solutions.	In McKinsey’s older frontier report, “functional silos that constrain end-to-end AI solutions” is among top obstacles. (McKinsey & Company) BCG’s “AI adoption in 2024” report notes ~70 % of implementation challenges come from people/process issues rather than technical ones. (Boston Consulting Group) PwC notes that “mindset, change readiness and workforce engagement” are bigger hurdles than technology itself. (PwC)
Trust, governance, risk, regulation, ethics	Particularly in regulated or mission-critical domains, users and leaders may distrust AI outputs, worry about liability, or regulatory compliance. Ethical concerns (bias, transparency) complicate adoption.	In PwC’s AI Agent Survey, “trust lags for high-stakes use cases” is singled out. (PwC) In healthcare AI studies, “regulatory uncertainty” and lack of leadership support are named key barriers. (PMC) The CMR adoption article mentions the difficulty of establishing reliable “sources of truth” in agentic systems, which implicates trust and governance. (California Management Review)
Cost, ROI uncertainty, and resource constraints	AI projects often require investment in compute, tooling, personnel, and change management. If firms can’t predict returns or have constrained budgets, they hesitate.	OECD report highlights that uncertainty over ROI is a critical obstacle. (OECD) McKinsey’s generative AI maturity commentary notes that while many invest in AI, only a small share believe they are mature; the risk of overinvestment is real. (McKinsey & Company) BCG notes that many companies struggle to scale value; 74 % say value/scale is a bottleneck. (Boston Consulting Group)
Employee resistance, culture, and change management	Fear of job loss, lack of understanding of AI, resistance to shifting workflows, or distrust of automated decisions can slow adoption. Cultural inertia is hard to overcome.	PCBB article cites employee resistance due to fear of job loss and uncertainty about AI’s role. (PCBB) PwC’s survey frames the barrier this way: “The biggest barrier isn’t the technology; it’s mindset, change readiness and workforce engagement.” (PwC) In healthcare, “low clinician adoption” is listed as a barrier. (PMC)
Integration, complexity, and deployment friction	Integrating AI into existing systems, managing model deployment, monitoring, continuous retraining, scalability, and infrastructure are nontrivial. Many pilot projects never survive “to production.”	Deloitte’s AI Trends commentary reminds that many firms are not ready for “agentic / sovereign AI” because of integration and operational complexity. (Deloitte) CMR article emphasizes the challenges of aligning real-time sources of truth and consistency in agentic systems. (California Management Review)
Legacy systems / technical debt	Older architectures, monoliths, tightly coupled systems make embedding AI more difficult or expensive.	McKinsey often notes that digitization maturity is a precursor to AI—firms in early digitization stages struggle most. (McKinsey & Company) Deloitte’s “foundational barriers remain” phrase speaks to legacy technical infrastructure. (Deloitte)
Small scale / lack of pilot scaling / fragmentation	Many AI experiments stay as pilots; lacking governance or metrics, they never scale or cohere across domains.	PwC notes that “few businesses are connecting agents across workflows and functions, yet that’s where the real value lies.” (PwC) McKinsey’s generative AI paper shows many firms have started use cases but have not scaled them or aligned them holistically. (McKinsey & Company)

Relative Importance & Anecdotes

In many surveys, people / process / culture issues (strategy, buy-in, silos, mindset) outweigh purely technical issues, even though companies often focus first on algorithms or models. For example, BCG says ~70 % of challenges derive from people/process issues versus 10 % from algorithmic shortcomings. Boston Consulting Group
The McKinsey “AI adoption advances” report (older, but commonly cited) places lack of strategy, lack of talent, and functional silos as the top three barriers. McKinsey & Company+1
The CMR article on agentic AI adoption underscores that data sources and consistency become even more critical when systems act autonomously; it’s not enough to have data—systems must share a unified, trustworthy “source of truth.” California Management Review
PwC’s survey (2025) is a more recent lens: it highlights that mindset, trust, and readiness are often bigger blockers than technical constraints. PwC

21 Executive Principles to Scale Your AI Adoption

This guide defines 21 Principles that executives can adopt and measure inside their organization to scale the adoption of AI technologies.

Read the Guide