McKinsey State of AI 2026: Agents Are Scaling at Big Firms, but Profit Is Not
Large companies are scaling AI agents faster than a year ago, yet the share reporting profit impact has barely moved. Here is how to read the gap.
In this story 9 sections
McKinsey's State of AI 2026 survey finds that 40% of large organizations are scaling AI agents in at least one function, up from 27% a year ago. Yet only 37% of respondents report any profit impact from AI, roughly unchanged. Adoption is moving faster than financial results, and smaller firms are stuck at 22%.
This analysis is for executives, investors, and builders who want a clear read on enterprise AI adoption. Emergent Wire covers the AI industry, and we use outside summaries of the survey to separate the solid findings from the shaky comparisons.
What Did McKinsey's 2026 Survey Find?
The survey shows broad adoption and uneven results. FM Magazine's summary of McKinsey's tenth annual survey, "The State of AI in 2026: On the Road to ROI," says 80% of respondents report better individual productivity and 50% say AI helps them make better decisions. For most enterprises, those gains have not yet turned into earnings growth.
On scaling, 44% of organizations report scaling AI, up from 38% the year before. Among organizations with at least $1 billion in revenue, 54% report scaling AI across the enterprise, versus about one-third of smaller ones.
The report draws on 1,719 responses from 97 nations, with fieldwork from May 4 to June 8, 2026, according to Digital Applied's evidence guide. That guide stresses that the numbers are respondent reports, not audited deployment counts.
Why Are Large Companies Pulling Ahead on Agents?
Large firms are scaling agents faster because scaling takes resources. FM Magazine reports that the share of large companies scaling agents in one or more functions rose from 27% to 40%, while smaller organizations stayed flat at 22%. A gap that wide suggests something other than interest.
The survey does not explain the gap, so any reason is an interpretation. A plausible one is capacity. Agents need clean data, access controls, monitoring, and staff who can fix them when they fail. Large firms have more of each, and they can spread the cost across more users.
That fits a wider pattern in enterprise software. New tools reach the biggest buyers first, and smaller buyers follow when packaged products arrive. Our look at the difference between AI agents and assistants explains why agents demand more setup than a chat tool.
Why Has Profit Impact Barely Moved?
Profit impact lags because productivity gains are hard to turn into earnings. FM Magazine says 37% of respondents attribute at least some EBIT impact to AI, roughly unchanged from last year. EBIT means earnings before interest and taxes. High performers, defined as those with at least 5% of EBIT attributed to AI and a significant impact, are about 6% of respondents, in line with 2025.
Individual productivity does not automatically become company profit. If every analyst saves two hours a week, the company only gains if it uses those hours for something that earns money, or if it needs fewer analysts. Otherwise, the time disappears into more meetings and slower work.
Another reason is cost. About 20% of respondents say AI-related operating costs constrain their use of AI. When a company scales agents, inference bills scale too, and a project that looked cheap in a pilot can look expensive in production.
There is also a definition problem. The word "scaling" can mean very different things. One company may call a rollout to 500 employees scaling, while another reserves the word for company-wide use. FM Magazine notes that its summary does not say how scaling is defined beyond the contrast with the experimental stage. That makes year-over-year changes more reliable than comparisons across firms.
Is Gartner's 40% the Same Number?
No, and mixing them up is a common mistake. Digital Applied points out that Gartner's figure is a forecast that task-specific agents will appear in 40% of enterprise applications by the end of 2026, up from under 5% in 2025. It is about applications, not about customers who have turned the feature on.
McKinsey's 40% is a survey result about large organizations scaling agents in at least one function. One counts software features, and the other counts companies' self-reported behavior. The guide warns that the two should not be averaged, compared, or treated as corroborating each other.
| Source | Figure | What it measures |
|---|---|---|
| McKinsey 2026 | 40% of large firms | Scaling agents in at least one function (self-reported) |
| McKinsey 2026 | 22% of smaller firms | Scaling agents in at least one function (self-reported) |
| Gartner forecast | 40% of enterprise apps | Apps with task-specific agents by end of 2026 |
| McKinsey 2026 | 37% of respondents | Any EBIT impact attributed to AI |
The Gartner release does not disclose a sample size or collection window, according to the guide. That is another reason to treat it as a forecast and not a measurement.
What Should Executives Take From the Gap?
The main lesson is to measure outcomes, not activity. A company that counts agents deployed is measuring effort. A company that counts hours saved, errors avoided, or revenue added is measuring results. Only the second group can tell whether scaling is paying off.
Start with a few concrete practices:
- Pick two or three workflows where you can measure cost or revenue before and after.
- Track inference cost per completed task, not only per token.
- Assign an owner for each agent, with a clear business metric.
- Review each deployment quarterly and shut down those that do not pay.
Emergent Wire's editors would also be cautious about vendor claims that cite adoption without outcomes. A claim that most large firms now use agents says little about whether those firms are better off.
What Does the Cost Constraint Mean for the AI Industry?
It means price will keep shaping demand. If one in five organizations says operating costs limit their AI use, then cheaper tokens and smaller models widen the market. That ties to the shift we covered in our piece on why AI model prices keep falling.
FM Magazine adds that 60% plan to increase AI investment over the next year despite cost pressures. Budgets are growing, but buyers are also watching returns more closely. That is a healthier setup than the early rush, and it favors vendors who can show results.
Productivity data points the same way. Our summary of AI productivity data in 2026 shows that measured gains are real but uneven across tasks and teams.
Investors face a similar test. Announcements about agent rollouts are easy to make and hard to verify. Earnings calls that cite specific cost savings, headcount changes, or revenue lifts carry more weight than slides about the number of pilots. Look for companies that give before-and-after numbers for a named workflow and that repeat those numbers across quarters.
How Should Smaller Companies Respond?
Smaller firms should focus on narrow, high-value uses before they try to scale broadly. A single well-chosen workflow, such as support triage or invoice processing, can show a return without a large platform. Buying a packaged tool is often cheaper than building one.
They should also learn from the large-company numbers without copying them. A 22% scaling rate for smaller firms is flat, but it does not mean they are behind in value. A small company that saves real money with one tool may beat a large one that runs dozens of unproven pilots.
Governance still matters at any size. Decide who can approve an agent, what data it can touch, and how you will detect errors. These steps cost little and prevent the incidents that make leaders pull back.
How Much Should You Trust the Survey?
Trust it as a directional reading. FM Magazine notes that the year-over-year changes are not dramatic, that the figures rely on respondents' own assessments, and that several are rounded. EBIT attribution and high-performer status are self-reported.
Survey samples also lean toward organizations that engage with consulting research. They may be more active in AI than a random company, so overall adoption could be lower than the headlines suggest. Treat percentages as comparisons across years, not as precise counts of the economy.
Finally, keep an eye on the high-performer group. Only about 6% of respondents qualify, and that share has not grown. If scaling agents were reliably profitable, you would expect that number to climb as adoption spreads. Its stillness suggests that the winners are doing something hard to copy, such as redesigning processes around the tools instead of bolting the tools onto old ones.
Conclusion
McKinsey's 2026 survey shows agents spreading among large companies, while profit impact stays flat and smaller firms hold at 22%. The Gartner 40% forecast is a different measure and should not be blended in. The honest reading is that adoption is ahead of proven returns.
Your next step is to measure results in the workflows you already automate, and to ignore adoption statistics that do not say how they were counted. Emergent Wire will track the next McKinsey and Gartner releases to see whether profit catches up.