AI agents are spreading quickly through large companies, but the share of companies seeing a profit impact from AI has not moved. The surveys published this year point the same way: companies are counting agents deployed and roles removed, when they should be counting verified outcomes and what each one costs. For a CFO, that is the scorecard to fix first.
Adoption is up
McKinsey's 2026 State of AI survey, with 1,719 respondents in the field from May to June, found that 40% of respondents from organizations with more than $1 billion in revenue were scaling AI agents, up from 27% a year earlier. Among smaller organizations the figure stayed at 22%.
KPMG's quarterly pulse of 314 US business leaders, taken from late July to late August, found 62% building, deploying or developing agents, up from 53% the quarter before. The share working on multi-agent systems jumped to 25%.
Profit is flat
The same McKinsey survey found 37% of respondents attributing at least some EBIT impact to AI, about the same as last year. Only about 6% qualified as high performers. About one in five said AI operating costs, including token costs, had limited how much they used it.
KPMG was more upbeat, with nearly six in ten leaders reporting measurable business value. But "measurable value" is a lower bar than a change in EBIT, and the two surveys are measuring different things.
Cutting jobs is not the same as a return
Gartner surveyed 350 executives on autonomous business and published the results in May. About 80% of organizations piloting or deploying autonomous capabilities had reduced their workforce, and the rate was nearly the same for those reporting higher returns as for those reporting modest or negative ones. Gartner's conclusion: "Workforce reductions may create budget room, but they do not create return."
Removing people before the work is reliably done elsewhere often moves the cost rather than removing it. The exceptions land on whoever is left, and quality slips.
Controls are slipping as agents spread
The KPMG numbers carry a warning. The share of organizations building controls into their agents fell to 30% from 43% two quarters earlier, even as deployment rose. Fewer than half (49%) have defined the high-risk cases where an agent may not act alone. 43% have set usage or token budgets.
More agents, fewer controls and unmanaged usage costs are a fast way to an expensive incident and a cancelled program.
The value is still sitting in the ledger
The opportunity has not gone anywhere. The Hackett Group's 2026 working capital survey of the 1,000 largest non-financial companies in North America found that days sales outstanding worsened by 2.1 days, the third year in a row, and got worse in 40 of 50 industries. Receivables made up 40% of a record $1.94 trillion working capital opportunity. That is exactly the kind of result agents can move, if it is measured.
Five measures that show whether agents pay
- Verified outcomes. Count only items that met a written definition: an invoice matched and posted and not reversed within 30 days, an order entered without correction. See how to write an outcome definition.
- Cost per verified outcome. Everything you spend (licenses, model usage, the people who handle exceptions and the oversight) divided by verified outcomes. Compare it with the cost before.
- Exception rate and causes. The share of items that needed a person, and why. A falling rate means the process is improving. A rising one means drift.
- The ledger effect. DSO, cost per invoice, close days, credit notes. If agents work, these numbers move.
- Control coverage. What each agent is allowed to do alone, what needs approval, and whether that is enforced rather than just written down. See how to set permissions and approval limits for AI agents.
Gartner expects 60% of organizations deploying AI to have a dedicated function tying AI cost to business value by 2029. A mid-market company does not need a new team for that. It needs those five numbers on one page each month.
Why we price this way
These measures are how we bill. Agentic MSP charges per verified outcome or a capped share of verified gains, and every billed item has a record in the Evidence File. If an agent does not deliver a verified outcome, you do not pay for it.