AI budgets are running ahead of the receipts

KPMG’s 2026 US CEO Outlook, published October 15, 2026, has 65% of 400 US CEOs¹ saying they will spend at least 10% of their capital budget on AI¹ in the year after the survey. Two days earlier, Wharton and GBK Collective reported that 40% of managers at large US companies²2Wharton Human-AI Research and GBK Collective, published October 13, 2026. US online survey, July 9-29, 2026: 1,007 managers and above at companies with 1,000+ employees and over $50M revenue. Self-reported. Read the source are very confident their organization can track total AI spend² once costs bundled into other software and cloud services are counted.

The samples differ, and both are self-reported. Read together, they describe the position going into budget season: at least a tenth of the capital budget is a large sum to approve on a total fewer than half of those managers feel very sure about. 87% of them expected AI budgets to rise² in the 12 months after the July 2026 fieldwork.

On September 16, 2026, Gartner forecast worldwide AI spending of $2.7 trillion in 2026 and $3.64 trillion in 2027³3Gartner forecast, September 16, 2026. Worldwide, all segments including hyperscaler infrastructure. A market size, not corporate AI budgets. Gartner sells research. Read the source, a market total including hyperscaler infrastructure, and placed generative AI in the Trough of Disillusionment³ in the same release. In the Census Bureau’s reading for October 5 to 18, 2026, 24.6% of US businesses⁴4US Census Bureau, Business Trends and Outlook Survey, published October 22, 2026. About 1.2M US nonfarm employer businesses, voluntary, prior two weeks. Experimental data product. About 9.6% answer “do not know.” Read the source reported using AI at all.

The evidence supports neither the sales pitch nor the obituary. Most firms in the broadest executive survey reported no productivity effect as of mid-2026⁵5NBER Working Paper 34836, Firm Data on AI, revised October 2026. Nearly 9,000 executives in seven countries, May-July 2026 (Germany January 2026); US N=964. Self-reported; productivity means sales per employee. Not peer reviewed. Read the source, a small share of organizations clears a high bar², and gains that do appear tend to arrive late⁶6NBER Working Paper 35684, August 2026. US public firms, 2010-2024, from job posting and resume data. Covers in-house AI talent, not firms using only off-the-shelf tools. Estimated effects. Not peer reviewed. Read the source. So the useful question before a 2027 line is approved is what this company will count, by when, and what happens if it does not show up.

US adoption is 24.6% or 95%, depending on who is asked

Census’s Business Trends and Outlook Survey, covering about 1.2 million US nonfarm employer businesses⁴, shows AI use rising from 17.3% in the period ending November 30, 2025, to 24.6% in the period ending October 18, 2026⁴. An NBER paper revised in October 2026, drawing on the Atlanta Fed’s survey of executives, puts US use at 95% in June 2026, weighted by employment⁷7NBER Working Paper 34836, revised October 2026. US data from the Atlanta Fed’s Survey of Business Uncertainty, N=1,012 in the mid-2026 wave, employment-weighted, any listed AI technology. Not peer reviewed. Read the source.

Neither is “the” number. About 75% of US firms have fewer than 10 employees⁸8US Census Bureau working paper CES 26-25, April 2026. Nationally representative US employer firms, November 2025-January 2026. Self-reported; performance link is correlation. Authors’ views, not official Census positions. Read the source, and Census counts them. In an April 2026 paper, Census researchers attribute most of the gap to the Atlanta Fed survey’s focus on large firms in AI-intensive sectors⁸. Restrict Census’s data to firms with 200 or more employees in AI-heavy sectors, answered by top executives, and it reaches 58% of firms, or 76% weighted by employment⁸.

The NBER paper’s first version put US use at 78%, counting firms that skipped the question as non-users⁷. Treating those firms as missing raised it to 94% for November 2025⁷. One decision about blank answers moved the result by 16 points.

A slide claiming that every competitor already does this needs a sample behind it. And the peer figure should resemble the company: Census’s October 2026 reading has use at 24.3% of firms with one to four employees and 43.6% of firms with 250 or more⁹9US Census Bureau, BTOS by employment size, collection October 5-18, 2026, published October 22, 2026. Standard errors for the 250+ class ran 1.51 to 3.63 points across these cycles, so single readings are noisy. Read the source. A lender weighing a borrower’s plan can ask the same thing.

Later readings need care. Census updated the question wording from the October 19, 2026 collection¹⁰10US Census Bureau, BTOS About page, read November 5, 2026. Wording updated from the October 19, 2026 collection. Comparability of later readings with this series has not been established. Read the source. When the wording changed in November 2025, it saw a level shift and started a new series¹¹11US Census Bureau, BTOS AI Core Question Updates, December 3, 2025. New AI question from November 17, 2025, and a new time series as a result. The series should not be spliced onto the older one. Read the source. A move after the October change may reflect the question as much as the economy.

One survey returns 80%, 28% and 7%, depending on the bar

Use says only who has the tool. On payoff, the large US reading published in October 2026 is the Wharton-GBK study, fielded in July 2026. Among managers and above, 80% report a positive return², the strongest pushback here on any claim that nothing is paying off. 28% report a strongly positive one². And 7% of organizations are successful across all three of the study’s markers².

Earlier global surveys set other bars. McKinsey’s, fielded May 4 to June 8, 2026, found 37% attributing at least some EBIT impact to AI, about the same as a year earlier¹²12McKinsey, The state of AI in 2026, August 25, 2026. Online survey, May 4-June 8, 2026: 1,719 participants in 97 nations, GDP-weighted. Self-reported shares of respondents, not companies. Not US-only. McKinsey sells AI services. Read the source, even as enterprise scaling rose from 38% to 44%¹². PwC’s, fielded September 30 to November 10, 2025, found 56% of CEOs reporting neither higher revenue nor lower costs¹³13PwC’s 29th Global CEO Survey, January 19, 2026. 4,454 CEOs in 95 countries and territories, fielded September 30-November 10, 2025. GDP-weighted, global, self-reported. Gain means a change of 2% or more. PwC sells consulting. Read the source, with a gain defined as a change of 2% or more. That 56% does not mean AI lost money.

The NBER paper covers nearly 9,000 executives in seven countries⁵, mostly surveyed from May to July 2026. 83% of firms report no productivity impact over the prior three years, the same share in the US⁵. The estimated effect is 0.2% a year, with 0.6% a year expected over the following three⁵. Reports of positive effects rose from 8% to 15% in six months⁵, so the direction is upward.

These figures cannot be averaged or read as a trend, because neither the respondents nor the questions match. For a budget meeting, the point is that “positive” and “measurable” are different claims, and a proposal should say which one it is making.

The common yardstick may be the wrong one

The Wharton-GBK summary reports that hours saved is the most common way departments judge AI spending, at 40%², and net-new revenue the least used of the major measures, at 27%². Among departments tracking a growth measure, 35% report strongly positive returns, against 13% of those tracking efficiency alone². The authors warn that some of the difference “may reflect conviction rather than capability.”²

In an NBER study fielded from November 2025 to mid-January 2026, nearly 750 US financial executives¹⁴14NBER Working Paper 34984, March 2026. Nearly 750 US financial executives, November 2025-January 2026; median firm $46M revenue. Implied figures still come from CFOs’ answers, not accounts. 2026 is projected. Not peer reviewed. Read the source reported that AI raised labor productivity 1.8% in 2025, while their own revenue and headcount answers implied 0.6%¹⁴. The paper finds little evidence of large near-term effects on headcount or costs¹⁴, and associates gains more with innovation and demand than with cost reduction alone¹⁴.

The authors lean toward a lag explanation¹⁴. On this evidence a lag cannot yet be told apart from optimism, which they list among the limits of survey data¹⁴.

Neither finding proves that growth measures cause better returns. Both suggest that a case resting only on hours saved or lower costs deserves harder questions. A revenue or margin measure named in advance gives the company a test it can fail.

Returns may take longer than the plan assumes

KPMG’s CEOs are specific about timing: 14% expect returns in six months to one year, 68% in one to three years and 13% in three to five¹. The strongest outcome data points to something slower. An NBER paper from August 2026 ties a one-standard-deviation rise in the growth of the AI-skilled worker share at US public firms to roughly one percentage point of extra productivity growth a year over 2018 to 2024⁶, after no relationship from 2010 to 2018⁶.

The authors call the gains “substantially lagged and gradually accumulating.”⁶ Their measure “does not capture firms that rely exclusively on off-the-shelf external AI tools,”⁶ so it maps poorly onto a subscription bought in one budget cycle. A plan promising payback within a year sits with the 14% in KPMG’s survey¹. Cutting a line at month 12 because nothing shows may be just as premature.

The practical answer is a dated checkpoint with an agreed measure and a stop rule. That last part is a real choice: in BCG’s January 2026 report, 94% of chief executives said they would keep investing at current or higher levels even if the spending did not pay off within a year¹⁸18BCG AI Radar 2026, January 15, 2026. 2,360 executives including 640 CEOs in 16 markets. Field dates not stated. Plans and expectations, not results. BCG sells AI services. Read the source.

Firms reporting gains tend to have redesigned the work

Egon Zehnder’s study of 843 global chief executives¹⁹19Egon Zehnder, The CEO Response 2026, press release, October 21, 2026. 843 global chief executives. Field dates not stated. Full report not read. Egon Zehnder is an executive search firm. Read the source, published October 21, 2026, came with a line from the firm’s global CEO, Francesco Buquicchio, that names the constraint: “CEOs can invest in AI much faster than they can change how an organization works.”¹⁹

The survey evidence fits that line, though these are associations. In McKinsey’s data, nearly three-quarters of high performers report fundamentally redesigning workflows, against one-quarter of other respondents¹², and those high performers, about 6% of respondents¹², are twice as likely to have defined processes to measure impact¹². Census researchers found 57% of AI-using firms in three or fewer business functions⁸, with broader use connected to better firm performance⁸ and no causal claim.

So read the composition of the line as closely as its size. The BIS paper finds larger gains when AI is paired with complementary investment¹⁵. If most of the money buys licenses, someone should say who is redesigning the work and on what budget.

The size of the line is a judgment for each company. The evidence supports one narrower request: write down the answer to this memo’s opening question before the money is approved.

What will this company count, by when, and what happens if it does not show up?