Most Companies Use AI. Far Fewer Have Scaled Agents Across the Organization

Ask survey respondents whether their organization uses AI and almost all of them say yes: in McKinsey's latest survey, nearly nine in ten did. Now picture a narrower question. Is an agent doing work on its own in one of your business processes, with someone checking the result? No survey asks it in exactly those words, but the closest questions get a yes far less often.

That gap is the useful part of this year's numbers. This piece is for people who own an AI rollout and want to know two things: how far other companies have actually got with agents, and what to watch for before putting one into a process you are accountable for.

How far companies have got

The broadest source is McKinsey's 2026 State of AI survey, published on 25 August. It was in the field from May 4 to June 8, 2026, with 1,719 respondents across 97 countries. The nine in ten are respondents who say their organization regularly uses AI in at least one business function. A smaller group, 44%, say AI is scaling across the enterprise, up from 38% a year earlier.

Agents are a smaller and more uneven picture.

• About two in ten respondents say their organization has reached the scaling phase with AI agents across the organization. A similar share say so for software coding agents.
• The next figure answers a different question: scaling agents in at least one function, not across the organization. The two cannot be converted into each other, and neither lines up with last year's 23% (scaling "somewhere") below.
• Size matters a lot. Among organizations with more than $1 billion in revenue, the share scaling agents in at least one function rose from 27% to 40%. Among smaller organizations it stayed at 22%.
• For comparison, the most widely scaled AI tool is the chatbot, at 47%.

In the text of the 2026 article I could read, there is no per-function breakdown, so the per-function figures here are from last year's survey (fielded mid-2025). Then, 23% said their organization was scaling an agentic system somewhere and another 39% were experimenting. But most of those scaling were doing it in one or two functions, and in any single business function, no more than 10% reported scaling agents. Stanford's AI Index 2026, which charts that same McKinsey data (the 2025 survey), puts it more bluntly: in most functions a majority of respondents reported no agent use at all, and scaled use was in single digits nearly everywhere. The exception was the technology sector, where scaled agent use reached 24% in software engineering, 22% in IT and 21% in service operations.

Two cautions. These are self-reported answers from people describing their own organization; Stanford itself calls them directional rather than comprehensive. And "scaling somewhere" is a statement about one place in a company, not about how much of the company's work runs through agents.

One vendor's own numbers

OpenAI publishes usage figures for its own enterprise customers. Its Enterprise Signals report, last updated 12 August, says that as of June, Codex generated 64% of combined Codex and ChatGPT output tokens among those customers. It also ranks customers by output tokens per active user each month: the top 10% now generate 8.3 times as much as firms in the middle (45th to 55th percentile), up from 2.6 times in January.

This is real data, but it answers a different question. It covers one vendor's paying customers, not the market. What the 64% does tell you is that most of these customers' combined Codex and ChatGPT output tokens come from the agentic tool; OpenAI counts Codex tokens as agentic use. The same report says that use is spreading beyond engineering: since February, weekly active Codex users at enterprise customers grew 108 times in legal, 41 times each in sales and recruiting, and 26 times in marketing, against 5 times in engineering. These figures measure usage; the report does not measure results. OpenAI says tokens are an imperfect measure of business value, and also calls them a useful proxy for the depth of AI use. Depth of use is what they show.

Do not stack these figures into one funnel. "Nine in ten use AI", "two in ten have scaled agents" and "64% of tokens" come from different samples, different questions and different units. No chain of percentages connects them.

Cost and value

Two other findings deserve more attention than the adoption rates.

The first is cost. About one in five McKinsey respondents say operating costs, tokens included, have limited their AI use. Gartner's 1 October summary of its 2026 AI Hype Cycle points the same direction: the question has moved from whether a company can build with AI to whether it can account for what its AI is doing and what it costs.

The second is value. 37% of respondents attribute any EBIT impact to AI, the same share as last year, even as more of them report scaling. Most still expect their organizations to increase AI investment over the next year.

What I have seen in small tests

I run agents and local models on fixed tasks and grade the output. Two recent pieces are relevant here, with a caveat up front: both are small samples on models running on my own machine, not production systems.

In one, I graded 120 runs of an agent whose checklist asked it to record how it had verified its own work. With the checklist advisory, none of the 60 runs with public records wrote that note. On the summary task (qwen3:14b), all 17 runs that finished the work declared done anyway. Enforcing the checklist made the note far more common. It did not get more bugs fixed. What enforcement bought was paperwork, written by the same agent whose work it was supposed to check.

In the other, a local model drafted support replies for six synthetic cases, one run each. In three, it chose a return policy before it had the information that decides which policy applies; two of those later corrected themselves in the same output or switched to a conditional. Two of the three hinged on fields like a receipt date, which a few lines of ordinary code can check, and refuse on when the date is missing.

Neither result says anything about adoption rates. They are failures I saw in my own tests, in what counts as done and in which decisions get left to the model, and they are worth checking for in your own process first.

Before an agent goes into your process

If you are deciding where an agent goes next, these are the questions I would want answered first. None of them needs a framework.

1. Which one workflow, in which function? In last year's survey, most companies scaling agents did it in one or two functions. If you cannot name the workflow, you are experimenting, which is fine. Budget and report it as an experiment.
2. Who decides that the work is done, and is it someone other than the agent? An agent's own report that it finished, or a note saying it checked, is a claim. Decide what evidence a person or a test will look at.
3. What can it do without asking? Write down which actions it can take alone (read, draft, look up) and which need a named person to approve (send, refund, change a record). Start with the short list.
4. Which decisions should not be the model's? Anything that follows from a date, a version, an amount or a customer's channel can usually be a rule. Leave the model the parts that actually need language.
5. What does one finished task cost? Track cost per completed and accepted task, not per seat or per month. About one in five respondents already say cost is limiting their use.
6. What set of cases will you rerun? Keep twenty or thirty real examples with known answers. Rerun them when the model, the prompt or the vendor changes, and compare.
7. How do you switch it off, and who picks up? Know who takes over the queue on the day you turn the agent off.

When someone shows you an adoption number, from a vendor, a consultant or your own team, ask what the denominator is: respondents, customers, tokens or tasks. One common source of confusion about how far agents have spread comes from comparing numbers that count different things.

If you have put an agent into a workflow and worked out how you check its results, I would like to hear what you settled on.