AI will have an impact on work between 2025 and 2035 similar to that of the internet between 2000 and 2010.
The automation of this or that has grown into an under-the-radar redesign of entire practices, including recruiting and customer care, as well as writing, research, and graphics.
The data we have for 2024 to 2025 already show the curve: adoption levels topping 70%, concrete improvements in productivity, and concerns over creativity, equity, and reliability.
This paper gathers the latest international evidence on AI at work, including the extent of its use, the industries and occupations where it is spreading fastest, and workers’ own attitudes toward it.
It examines the metrics that can be observed: the specific applications most commonly used, the influence on productivity and innovation, and early hints of ROI.
The sum of these numbers does more than paint a picture of where AI stands today; they point to the building blocks of a future economy of work in which the distinction between humans and algorithms will be irrelevant.
Global rate of workplace AI adoption (2019 to 2025)
Over the past seven years, AI has gone from being a fringe idea to something that most companies claim to be adopting in practice.
The proportion of companies using AI in at least one function, as measured by McKinsey’s long-term global survey, dipped during the COVID-19 crisis but has since rebounded with the emergence of generative AI: 72 percent in the early 2024 survey and 78 percent in our latest (2025) survey.
This uptick is confirmed by the 2025 AI Index report from the Stanford Institute for Human-Centered AI, which says that 78 percent of organizations were using AI in 2024, compared with 55 percent in 2023.
Snapshot of Adoption
| Year | % of organizations using AI in ≥1 business function |
| 2019 | 58% |
| 2020 | 50% |
| 2021 | 56% |
| 2022 | 50% |
| 2023 | 55% |
| 2024 | 72%* |
| 2025 | 78% (latest) |
*Early-2024 reading; several end-of-year reports indicate that the proportion was about 78 percent by the end of 2024, in line with the “latest” level indicated by the 2025 survey.
Sources: McKinsey State of AI in the Enterprise survey series (see sidebar, “Survey demographics,” for details on methodology and respondent profiles), with the 2024 and 2025 levels confirmed by the Stanford AI Index.
What the data mean
The adoption rate flattened out at about 50 to 56 percent from 2020 through 2022 before increasing in 2023 and 2024, as companies began deploying pilots of generative AI into production, primarily in IT, marketing and sales, and customer-service applications.
This means a lot of embedding or plugging in of AI but not necessarily deep transformation of business processes; nevertheless, it is a significant step toward expanding the “tool kit” of the typical employee.
Analyst commentary
My interpretation of this trend is that it represents the S curve we typically see in the evolution of technology adoption, with one caveat: in this case, it was faster for companies to flip the switch on adoption than it will be to squeeze out the commensurate value.
Many companies opted for “easy” or embedded AI (for instance, via plug-ins or copilot functionality), so the numerator (the proportion of companies that have adopted AI) has been growing faster than the denominator (value).
Looking ahead over the next year or two, I would expect the adoption rate to continue rising, albeit only modestly, given that there isn’t much more headroom, while the focus and emphasis shifts to embedding these tools, rationalizing applications and tools, and refining the operating model of a small number of mission-critical business processes where AI can drive material efficiency gains.
Companies that approach AI as they would any other capability (by defining an owner, budget, key performance indicators [KPIs], and a process for retiring investments that fail to meet their business case) will be those that reap lasting productivity improvements from their adoption of AI.
Industry use cases (2025)
The industry patterns in 2025 are sufficiently distinct to guide investments.
The top two sectors in terms of deployment of gen AI in daily operations are tech and professional services, followed by media and telecom and advanced industries (which includes electronic, aerospace, and automotive companies), with consumer, finance, and the typically heavily regulated and asset intensive trailing (see sidebar “Our survey on the state of AI in 2025”).
Seventy-one percent of respondents to the latest global survey from McKinsey on the state of AI reported that their organizations deploy gen AI in at least one function, but the rates vary widely by sector.
Snapshot by sector
| Industry sector | % using gen AI in ≥1 function (latest 2025 survey) |
| Technology | 88% |
| Professional services | 80% |
| Advanced industries | 79% |
| Media & telecom | 79% |
| Consumer goods & retail | 68% |
| Financial services | 65% |
| Healthcare, pharma & medical products | 63% |
| Energy & materials | 59% |
| Overall (all sectors) | 71% |
Source: McKinsey Global Survey on the State of AI (fielded H2 2024, published 2025). Figures represent generative AI usage across industry, which is the percent of organizations using generative AI in at least one business function.
The implications
In my view, these results reflect the relative ease of incorporating AI into business processes.
Tech companies tend to build AI into products and services; professional services rely on knowledge management and writing. Media and telecom use AI in service operations.
Healthcare and energy appear to be lagging, not because there is a lack of use cases but because for them, to achieve production readiness, requirements such as safety, data governance, and integration with legacy systems must be met.
Finance tends to prioritize governance over deployment, a slower process, though probably one that is more sustainable in the long term.
In addition, to put the sector breakdown into perspective, the global business landscape more broadly has experienced a significant uptick in the use of AI in recent years (see sidebar “Rise of AI across business globally”).
Analyst perspective
For planning purposes, I would assume the rates of adoption will not be the constraint; depth will be.
The leaders in every sector are shifting from “flipping switches” to fundamentally reimagining a few high throughput processes (for example, claims adjudication, marketing content production, level one customer support, bug fixing).
The next level of differentiation will come from data infrastructure and risk management: data fetch integrated with governed data sets, usage tracking, and human review points.
Two operational markers of AI maturity that I look for are (1) a single owner for AI operating risk and (2) an investment portfolio approach that sunsets anything that is not paying for itself in computing resources.
More heavily regulated sectors will catch up as model reliability and explainability become routine, and both of those are closer than you might think.
AI Adoption by Function (2025)
Here are the percentages for 2025, broken down by function. Gen AI is most prevalent in externally facing, content-rich, and software-writing functions, and less common in functions involving capital or rigorous oversight.
These figures come from a new global survey by McKinsey (conducted H2’24, published 2025) on the functions where gen AI is being used regularly. The functions identified in the survey are what I call “functional roles.”
Overall, we see marketing and sales at the top, product and service development, IT, and service operations in the middle, and risk and compliance, supply chain, and manufacturing bringing up the rear (at least for now).
Snapshot by role type (share of organizations regularly using gen AI)
| Role type (function) | % of organizations |
| Marketing & sales | 42 |
| Product & service development | 28 |
| IT | 23 |
| Service operations | 22 |
| Knowledge management (other corporate) | 21 |
| Software engineering | 18 |
| Human resources | 13 |
| Risk, legal & compliance | 11 |
| Strategy & corporate finance | 11 |
| Supply chain & inventory management | 7 |
| Manufacturing | 5 |
Source: McKinsey Global Survey on the State of AI (H2’24 data, published 2025). Percentages represent “companies that use gen AI regularly in at least one use case” by function. “Knowledge management” is the label used by McKinsey to aggregate other corporate functions.
Key takeaways
I observe two patterns here. First, functions that inherently involve text, images, code, and structured data (marketing, product, IT, and software development) are relatively easier to deploy models into.
Second, risk and compliance, supply chain, and manufacturing are lower down on the list because they involve more stringent gating, data access, and safety justifications.
Beyond that, the fact that overall gen AI adoption has shot up in 2024 (see business-wide gen AI adoption) provides some context for even the lowest functions to show some uptick.
Analyst’s take
In my view, 2025 is the last year that penetration will be the story.
These front-running functions won’t just “adopt AI” in these roles; they will end-to-end automate a few select, high-frequency tasks in these areas, such as campaign planning to A/B testing in marketing, bug fixing in software engineering, and knowledge lookup with auditing in support functions.
Three signs will signal gen AI maturity: (1) a named owner of AI operating risk, (2) data lookup functions connected to well-governed data sources (and not the open internet), and (3) a pipeline with sunset rules for prototypes that don’t generate sufficient value to justify their energy consumption.
I’d also expect functions on the right-hand side of the graph (supply chain and manufacturing) to increase as insurance against tooling maturity and synthetic data pipelines becomes more robust. In short: we’re done with gen AI breadth. It’s time for gen AI depth.
Employee Exposure to AI Tools (2023 to 2025)
In 2025, we see clear separation between role types. Generative AI permeates jobs with an external, content-rich footprint or those that are involved in software development. It is less prevalent in capital- or permission-intensive areas of the business.
McKinsey’s most recent global survey (conducted H2’24; published 2025) shows where companies are now regularly deploying gen AI, by function. I’m treating these functions as the practical proxies for “role types” in the organization.
The headline result: marketing and sales are the most exposed, followed by a big cluster of product or service development, IT, and service operations. The roles least likely to use gen AI today are in governance, supply chain, and manufacturing.
Snapshot by role type (share of organizations regularly using gen AI)
| Role type (function) | % of organizations |
| Marketing & sales | 42 |
| Product & service development | 28 |
| IT | 23 |
| Service operations | 22 |
| Knowledge management (other corporate) | 21 |
| Software engineering | 18 |
| Human resources | 13 |
| Risk, legal & compliance | 11 |
| Strategy & corporate finance | 11 |
| Supply chain & inventory management | 7 |
| Manufacturing | 5 |
Source: McKinsey Global Survey on the State of AI (H2’24 data; published 2025).
Numbers represent the “regular use of gen AI in at least one use case” per function. “Knowledge management” is the bucket of other business functions.
Takeaways. What are the key observations here?
There are two for me. The first is that functions that have heavy text, image, code, or knowledge representation as inputs already (i.e., marketing, product, IT, software) are easier to penetrate. The models are a drop-in.
The second is that more permission- or asset-intensive areas of the business (risk/compliance, supply chain, manufacturing) are lower on the list because assurance, access control, and legal safety are more important than generating novel outcomes.
It’s worth noting that business adoption overall rose throughout 2024. So even those trailing in adoption are seeing at least some lift.
Analyst’s take.
How do I interpret these results? My view is that 2025 is the year the penetration stops being the story.
The most-advanced companies will have found a way to penetrate gen AI use into these roles, but more importantly, will use it to rewrite a few key high-volume tasks from end-to-end (e.g., idea generation through A/B testing in marketing, bug fixing in engineering, or knowledge search with tracing in support).
I’ll be looking for three proxies for maturity here: (1) a single person named as owning AI operational risk, (2) search or other retrieval processes using managed corpora (as opposed to the internet), and (3) a portfolio of use cases that quietly retires projects that don’t pay back for the computation.
More heavily regulated industries will close the gap as soon as assurance and provenance issues become routine, and that’s closer than most suspect.
The Productivity Effect of AI Tools (2024 to 2025)
As I digest the newest evidence on the effect of AI on productivity, I notice two patterns: there are some hard benefits to be found here (when you do it right) and there is a large footnote that says “it depends on how you do it.”
The Federal Reserve Bank of St. Louis finds that, “among employed users of generative AI in the United States, the new technology helped with 6 percent to 24.9 percent of their total work hours (during their usage week) in late 2024.”
Another paper finds that “each hour of generative AI use was about 33 percent more productive than a typical hour of work.”
Here’s a simple table to pull these findings together:
| Period | Metric | Observed impact |
| Late 2024 (Nov survey) | % of all work hours assisted by generative AI (users) | 6 % to 24.9 % |
| Late 2024 (Nov survey) | Productivity gain per hour of generative AI use | ≈ 33 % more productive |
| 2025 forecast / aggregate | Productivity growth potential from AI (economy-wide) | 0.3 to 3.0 percentage points added to annual productivity growth |
What do these figures tell us?
In my reading, the implication of these findings is that when employees meaningfully interact with these tools, there is indeed a productivity dividend to be found.
However, the fact that this dividend tops out at 24.9 percent “hours assisted” implies that most people are not allowing these tools to consume their every working moment.
The 33 percent per hour productivity boost is terrific, but it only pertains to those hours when the tool is in use, not for the week overall.
And when aggregated to the whole economy, the benefits are in the range of 0.3 to 3.0 percentage points of annual productivity growth.
In other words, this is all still in its infancy; the benefits are real but still concentrated in a few pockets.
My view
The upshot of this for executives in my view is that the easy wins of AI-enabled productivity are here for the taking, but realizing these gains broadly will take process redesign, upskilling and management.
Executives need to shift from “great, let’s just put these tools in everyone’s hands,” to “which processes are we prepared to overhaul?” “Which hours will the tool actually support?”
“Which business processes can we inject the tool into and where will we be able to measure hour-by-hour productivity enhancements?”
Until we are able to do this, we will be stuck in a world of partial bars (6 percent to 24.9 percent of hours assisted), not full bars.
The challenge ahead is not about finding the productivity dividend; it is about institutionalizing it, internalizing it and diffusing it throughout the organization.
Job postings already list AI skills as requirements
The job market does not have time to debate the matter; it already has added “AI literacy” as a requirement to many job postings.
LinkedIn recently reported that job postings listing “AI literacy” as a requirement – including experience with ChatGPT, GitHub Copilot, and prompt engineering – grew more than sixfold in the past year.
While such job postings are still relatively rare (i.e., 0.2% of all paid job postings globally), the growth rate is unmistakable.
Similarly, Indeed reported in January 2025 that job postings in the United States that reference generative AI have grown 170% in the past year, while the percentage of postings remains relatively low at about 0.3%.
Snapshot of AI-skill mentions in job ads
| Period | Platform/Scope | What’s measured | Value | Notes |
| 2023 Q3 | LinkedIn (global) | Share of paid jobs listing an AI-literacy skill | ~0.03% | Implied by 2024Q3 being >6× higher and at ~0.2% (1 in 500). |
| 2024 Q3 | LinkedIn (global) | Share of paid jobs listing an AI-literacy skill | ~0.2% | “1 in 500” jobs requested AI-literacy; up >6× YoY. |
| 2025 Jan | Indeed (U.S.) | Share of postings mentioning GenAI terms | ~0.3% | About 3 in 1,000; ~170% YoY growth from Jan 2024. |
My interpretation
This is a classic “thin tail, steep trend” situation. The base level of job postings that require “AI literacy” skills is still low (i.e., well below 1%) but the trend is very strong (i.e., a sixfold increase in one year on LinkedIn and a 2.7-fold increase on Indeed).
Clearly, many jobs are moving from requiring “nice to have” experience with AI tools to requiring a baseline level of “must have” literacy in using AI tools.
We see the language first appearing in job postings for technical jobs (e.g., software development, data science) and consulting jobs and then spreading to other knowledge worker roles as use of AI tools becomes more standardized within organizations.
Analyst’s take
If I were managing a team, I would view “AI literacy” as I now view “spreadsheet literacy.” It is not required for all jobs, but it is expected for many jobs that involve analysis, writing, or serving clients.
To address the need for AI literacy, hiring managers should do two things. First, they should identify the roles that require proficiency with specific AI tools and include that in the job description.
This keeps the job requirements grounded in reality and helps applicants decide if they have the requisite skills.
Second, they should provide training on AI tool use for new employees, including tutorials on using the most popular tools, examples of accepted use cases, and tools for measuring the benefits of each use case.
This is because the writing is on the wall: Job descriptions increasingly will include proficiency with AI tools as the reality of how work gets done catches up with job descriptions.
Sentiment among workers about AI (2025)
I have been looking at the recent research into worker attitudes about AI in work. From what I have seen there is a sense of optimism, a sense of unease and a sense of complexity.
On the one hand, workers know that things will change, on the other hand, they are unsure what that will mean for them personally.
The numbers
According to a new Pew Research Center survey of U.S. workers (early 2025):
52% say they are worried about how AI will be used in the workplace. 36% say they are hopeful about AI’s impact on their work. 16% say some of their work is currently being done with AI. 25% say they could imagine some of their current work being done with AI.
A global study by KPMG International and University of Melbourne (48340 participants, across 47 countries) found that 57% of employees admit they have hidden their use of AI tools at work.
Table of worker sentiment metrics
| Metric | Value | Notes |
| Worried about how AI will be used in the workplace | 52% | U.S. workers survey |
| Hopeful about AI’s impact on their work | 36% | Same source |
| Workers whose job currently involves AI | 16% | U.S. workers survey |
| Workers who admit hiding AI use at work | 57% | Global KPMG/University of Melbourne study |
What these numbers are telling us
From my perspective, I see two tracks in the workforce. Many workers know about AI and what it can do, but fewer feel completely secure or prepared.
That over half are worried suggests that deployment and communication is not yet where it needs to be.
That a solid one-third feel hopeful suggests that the opportunity is evident and palpable.
The statistic about “hiding use” is especially interesting; it suggests a disconnect between deployment and worker comfort (or disclosure), as workers are using tools but perhaps don’t feel safe or supported to say so openly.
My take
I think these mixed results are a wake-up call. I think organisations should not assume that worker confidence will come simply because the technology is available.
Rather, organisations need to work with employees to build trust, to clarify use and to train workers.
My advice is to invest in clear policies around how AI will be used, involve workers in the development of these policies and to track confidence alongside usage.
The tech is ready, but the people aren’t yet.
In a nutshell, we are through the shock-and-awe phase of AI in work, and now it is time for the alignment phase.
Most frequently used AI tools (2025)
What about AI tools employees actually access on the job? There, we see two snapshots — one of developers and another of enterprise users.
Developer focused: Within the developer segment, we see two out-of-the-box productivity tools leading the pack.
In the 2025 Stack Overflow survey, 82% of developers report using ChatGPT while 68% use GitHub Copilot, followed by Gemini (47%), Claude/Claude Code (41%), Microsoft Copilot (31%), and Perplexity (16%). Note that these are percentages of developers who use any AI tool, not of all users.
Enterprise: On the enterprise side, we have real-world browser telemetry data that suggests a similar story: according to LayerX’s 2025 report, ChatGPT comprises 92% of all enterprise GenAI usage, followed by Gemini (15%), Claude (5%), and Copilot (2-3%). The report also estimates that 45% of employees use some GenAI tool or another, a reminder of just how common these tools have become.
Snapshot: most-used AI tools at work (2025)
| Tool | Share & scope |
| ChatGPT | 82% of developers using out-of-the-box AI (Stack Overflow 2025); ~92% of enterprise GenAI usage by traffic (LayerX 2025). |
| GitHub Copilot | 68% of developers (Stack Overflow 2025). |
| Google Gemini | 47% of developers (Stack Overflow 2025); ~15% of enterprise GenAI usage (LayerX 2025). |
| Claude / Claude Code | 41% of developers (Stack Overflow 2025); ~5% of enterprise GenAI usage (LayerX 2025). |
| Microsoft Copilot | 31% of developers (Stack Overflow 2025); ~2–3% of enterprise GenAI usage (LayerX 2025). |
| Perplexity | 16% of developers (Stack Overflow 2025). |
Context: The Stack Overflow data are based on developer self-reporting of tool usage, while the LayerX data are based on enterprise browser telemetry data. Both reports were published in 2025.
Interpretation
My interpretation is simple: While the long tail is long indeed, the tools employees actually use at work are relatively few.
ChatGPT is still the front door for most users, both developers and (in terms of relative traffic) the broader enterprise.
Copilot has particular traction with developers, but trails in the enterprise because most use cases occur in consumer-facing chatbots accessed via personal accounts.
The second story is one of fragmentation: While tools like Gemini, Claude, and Perplexity have solid use cases, they have not replaced the general-purpose pattern of “open chatbot, get answer.”
Planning implications
If I were planning rollouts, I would plan for a single generalist tool and a small number of specialty tools to cover most use cases.
The key will be governance and integration: Route most everyday queries through a centrally managed chatbot with access to company data and logs, but make it seamless for employees to call up specialty tools (e.g., coding copilots or search-heavy retrievers) from within the same interface.
Track not just MAUs, but percent of tasks assisted and time per task saved; that’s where the value lies.
Having most of your employees concentrated on a single front-door tool isn’t a problem; it’s an opportunity to standardize prompts, logs, and guardrails so that those benefits can scale without the chaos.
Productivity vs. Creativity Outcomes (2024 to 2025)
After reviewing the latest research, I see that generative AI tools are having a solid, positive impact on productivity, while the impact on creativity is more nuanced.
That is, businesses are moving faster, but it’s less clear that they’re becoming more creative.
Key findings
In 2025, the Organisation for Economic Co-operation and Development (OECD) found that workers who used generative AI tools were about 40% faster when writing or summarising text, and the quality of their work was about 18% higher, as rated by evaluators.
A 2025 meta-analysis of 28 studies (involving more than 8,000 participants) found that people who worked with AI were better at producing creative work (with an effect size of g ≈ 0.27), but that the diversity of the ideas they generated decreased (with an effect size of g ≈ -0.86).
Snapshot: productivity vs. creativity outcomes
| Metric | Value (2024–25) | Context / Notes |
| Time-to-complete writing/summarising tasks | ~-40% faster | OECD experiment for mid-level professionals. |
| Quality improvement of output | ~+18% | As judged by external evaluators in the same study. |
| Creative performance boost (human + AI) | g ≈ 0.27 | Meta-analysis of 28 studies. |
| Idea diversity change (human + AI) | g ≈ -0.86 | Indicates less idea variety when AI plays a role. |
What the numbers suggest
In my reading of the data, the boost to productivity is more straightforward. With the help of AI tools, workers are getting their jobs done faster and producing higher-quality work in some areas, like writing and summarising. But the impact on creativity is more subtle.
It’s true that people who collaborate with AI on creative work perform better (g ≈ 0.27) but the diversity of the ideas they come up with suffers as a result (g ≈ -0.86). This implies that AI systems may be leading humans to similar solutions rather than truly original ones.
Analyst’s view
In my view, this means businesses should treat productivity and creativity separately. If you want to improve the speed, consistency and quality of repeating tasks, then AI is a no-brainer.
If you want to radically innovate, generate ideas or explore brand new possibilities, however, you will need to manage the interplay between humans and machines more carefully.
You will want to allow people to retain their agency, ensure a diverse array of inputs, and make sure the machines don’t dominate the idea generation process.
If you want to reap the full benefits of AI by mid-2025, I would suggest you follow a twin-track approach: In the short term, you should focus on tasks where you can speed up and improve quality; in parallel, you should be investing in spaces for experimentation where divergence (not convergence) is the objective.
AI in Recruitment & HR Automation (2025)
It’s now 2025 and HR teams are using AI and automation in hiring, onboarding and managing employees. The once experimental pilot projects are now must-haves.
In fact, 99% of hiring managers say they’re already using AI in their hiring process, and 98% are reporting “significant improvements” because of it.
Similarly, 65% of small businesses say they’re already leveraging AI for HR purposes (primarily recruitment) and more than half plan to increase investment in the next year.
Snapshot: key metrics in 2025 for hiring & HR automation
| Metric | Value | Context / Notes |
| Hiring managers using AI in hiring process | ~99% | From “AI in Hiring 2025” survey. |
| Hiring managers seeing significant efficiency improvements via AI | ~98% | Same survey as above. |
| Small businesses using AI for HR, primarily recruiting | ~65% | According to Paychex / RBJ article. |
| Organisations planning further investment in HR/AI functions | ~53% | From same small business-study; more intend to invest. |
| HR departments using AI for talent acquisition / monitoring engagement | ~54% / ~62% | From a broader “AI in workplace” dataset. |
So what does this really mean?
I’ll be the first to say I’m somewhat surprised how rapidly AI in HR has taken hold, even if it’s in its earliest form of applying to hiring. But the rapid progress from experimental to nearly ubiquitous is unexpected.
There are numerous processes currently being automated such as candidate sourcing, resume screening, interview scheduling, and even aspects of the interview process itself.
No doubt, these processes are more efficient. The work is often completed more quickly and with less human intervention.
But there’s also a flip side. Once 99% of hiring managers are using AI, it no longer becomes a differentiator in and of itself.
The competitive differentiator is going to be how you implement it, how thoughtfully you implement it, how fairly you implement it, and how well you integrate it with human judgment.
A case in point: 65% of small businesses have adopted AI for HR processes (mainly recruitment), which indicates that this is not an enterprise-only phenomenon.
Our View
Two important analyst ramifications come to mind.
First, governance, risk, and control need to move out of the background. Because AI is now an integral part of the majority of hiring decisions, issues such as bias, explainability, candidate experience, and compliance become business as usual.
While automation can certainly help accelerate the front-end of recruiting (sourcing), it also fundamentally alters decision rights, auditing, and equity. That’s a transformation not an optimization.
The second thing is that you will need to focus on human and AI collaboration.
AI can definitely speed up and optimize recruitment processes, but at the end of the day, the value is in human evaluation; the cultural fit, strategic fit and the future fit.
Firms that employ AI to analyze and suggest but retain human debate on who should be hired will do better than those that delegate the whole decision to AI.
We’re past the stage of asking if HR can be automated. It can. The question is how well it’s being automated, governed, and integrated with human processes.
2024 – 2025: Value & ROI in AI
Now for the 2024-25 ROI picture. For starters, inside business functions, leaders are claiming direct (revenue increases, cost reductions) ROI outcomes from their generative AI investments.
However, if you step back and look at it at an enterprise level, the reality is more mixed. Although there have been some successes, almost none have yet really made a significant dent in EBIT across the enterprise.
According to our latest global survey, a majority of executives across business functions report having experienced both cost savings and revenue gains in the second half of 2024.
More than 80 percent also report no material enterprise-level EBIT impact. Just 17 percent indicate that 5 percent or more of total EBIT over the past year was driven by generative AI.
Take a snapshot: Where is ROI appearing (share of respondents reporting value, by business unit, H2 2024) Source: State of AI in 2024, McKinsey Global Survey (2025).
| Function | Revenue increase | Cost decrease |
| Strategy & corporate finance | 70% | 56% |
| Supply chain & inventory | 67% | 61% |
| Marketing & sales | 66% | 47% |
| Service operations | 63% | 58% |
| Software engineering | 57% | 52% |
| Product / service development | 51% | 43% |
Take a snapshot: Where is ROI appearing (share of respondents reporting value, by business unit, H2 2024) Source: State of AI in 2024, McKinsey Global Survey (2025).
In-function metrics show the share of respondents who report that their generative AI applications have brought in new revenue or cut costs in the last 12 months.
When it comes to business results, however, the picture is very different: more than 80 percent say that they haven’t seen a significant EBIT impact from generative AI yet, and only 17 percent report that generative AI has accounted for 5 percent or more of total EBIT in the past year.
What these figures represent
In my opinion, this is what we should be seeing. The point solutions are delivering. The question now is, can they scale?
The most impactful results at the functional level are being achieved in areas that are inherently digitized and metric-driven: strategy and finance, supply chain, service operations, and engineering. After all, it’s easier to measure gains in efficiency and performance.
However, to achieve enterprise-wide EBIT from these point optimizations, you need to do more than add a few more tools.
You need to standardize APIs to core systems, assign AI risk and ROI accountability, and, importantly, reinvest time saved into value added activities and not have it simply leak away.
This is also why most teams can proudly report business impact, but the CFO still isn’t ready to say the whole company has been transformed.
The analyst’s take
If I were to strategize for 2025–2026, I would strategize for ROI on three connected fronts:
- at the use case level (e.g. reduction in minutes, reduction in errors, % upsell per call) b. at the portfolio level (e.g. consolidation of tools, killing of a slow-burning pilot, AI resource allocation) c. at the enterprise level (e.g. EBIT, CCC) with a clear process for releasing resources back into the business
We know there’s value to be had. The data tells us that. You just have to create the pipes. You have to attach those local benefits to the bottom line through governance, process simplification, and analytics.
Taking a step back and examining all of these different indicators, we can see one thing very clearly: AI is not just on the doorstep of the workplace. It’s already inside.
We see high adoption. We see tool usage. We see some mixed signals on ROI, but we do see ROI.
Productivity is rising faster than creativity, which means the initial phase of this transition is more about efficiency than radical innovation.
Underlying these productivity measures, though, is a more subtle evolution: a steady movement of work from doing things to deciding, managing, and improving things.
By 2035, we will no longer be discussing our use of AI. We will be talking about the extent to which we have successfully amplified trusted systems, secured our data, upskilled our employees, and — in addition to productivity — assessed our collective intelligence and innovation.
Having learned to coexist with AI in the 2020s, we will learn to collaborate with AI in the 2030s.
