Four Forecasts, One Year, and the Free Data That Beats ...

Gartner has published four different forecasts of worldwide IT spending for calendar year 2026. All four appeared inside a ten-month window. Three of them landed after 2026 had already started.

Published	2026 IT spending	Growth over 2025
22 October 2025	$6.08 trillion	9.8%
3 February 2026	$6.15 trillion	10.8%
22 April 2026	$6.31 trillion	13.5%
27 July 2026	$6.37 trillion	14.2%

Every one of those is verbatim from a Gartner press release, and every one was delivered in the same flat declarative voice. "Worldwide IT spending is expected to reach $6.37 trillion in 2026, up 14.2% from 2025." No hedge, no range, no revision note.

Subtract the first from the last and the upward revision is about $290 billion. That is my arithmetic on two of Gartner's own published figures, not a number Gartner published. Here is why it is worth doing the subtraction: in the same July 2026 release, Gartner forecasts the entire worldwide infrastructure-as-a-service market at $287 billion for 2026. The correction to the estimate is larger than a market being estimated alongside it.

I want to be careful about what that does and does not prove. Forecasting a year that is happening around you is genuinely hard, revising upward when the data moves is the correct behaviour, and a firm that never revised would be worse, not better. The problem is not that the number changed. The problem is that the delivery did not change with it. Nothing in the July release tells you the same team said $6.08 trillion nine months earlier. You have to go and find that yourself.

The AI-specific forecasts move the same way. In January 2026 Gartner put worldwide AI spending at $2.52 trillion for 2026 and $3.34 trillion for 2027. In May 2026 it put them at $2.59 trillion and $3.49 trillion. The 2027 figure moved by about $150 billion before 2027 began, which is again my subtraction rather than a number Gartner published.

The confidence is identical whether the evidence is vendor sales or a webinar poll

Here are two Gartner outputs from the same period.

The first: worldwide IT spending will grow 14.2% in 2026. Gartner describes that forecast as resting on analysis of the sales of over a thousand vendors. That is a real evidentiary base, assembled at real expense, and nobody else has quite the same one.

The second, and the one Gartner AI statistic secondary coverage reaches for most often: over 40% of agentic AI projects will be canceled by the end of 2027. The underlying survey is disclosed in the release itself, openly and without spin. It was a January 2025 poll of 3,412 Gartner webinar attendees. Nineteen percent said their organisation had made significant investments in agentic AI, 42% conservative investments, 8% none, and the remaining 31% were waiting or unsure.

People who sign up for a Gartner webinar on agentic AI are not a random sample of businesses. They are a sample of people interested enough in agentic AI to give up an hour. Gartner says so plainly. It is the secondary coverage that strips the caveat, not Gartner.

But both outputs reach the reader in the same font, in the same format, with the same certainty. There is no visual or verbal marker separating "we analysed a thousand vendors' revenue" from "we polled the people on our webinar." If you are going to quote one of these in a board deck, you have to go and look up which kind it is, every single time.

The failure predictions, in order, and the two that get merged

Read as a sequence with dates attached, the prediction lineage is more interesting than any single line of it.

July 2024: at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025.
June 2025: over 40% of agentic AI projects will be canceled by the end of 2027, on escalating costs, unclear business value or inadequate risk controls.
September 2025: less than one in five generative AI projects will achieve their desired business value through 2026.
May 2026: by 2027, 40% of enterprises will demote or decommission autonomous AI agents because of governance gaps found only after production incidents.
July 2026: through 2028, at least 50% of generative AI projects will overrun their budgeted costs.

Those two 40%-by-2027 lines are different predictions. One is about projects being canceled before they ship, published June 2025. The other is about agents being pulled back after they shipped, published May 2026, with a different mechanism and a different failure mode. Secondary coverage merges them constantly, usually into a single sentence of the form "Gartner says 40% of AI agents will fail by 2027." If you see that sentence, the writer has not read either release.

Alongside the cancellation prediction sits the sharpest thing Gartner has published on the AI vendor landscape: an estimate that only about 130 of the thousands of self-described agentic AI vendors are real. Gartner coined "agent washing" for the rest. That is a genuinely useful number for a buyer, and it is free.

Gartner's own navigation disagrees with Gartner's own page by nearly five times

On 21 August 2026, the site navigation across gartner.com labels its AI vendor hub "Competing in the $1T AI market." Click through to the destination page and the body copy invites you to "win a share of the 4.7-trillion-dollar AI market opportunity." Same site, same day, one click apart, off by a factor of 4.7.

Neither of those is a forecast in the sense the press releases are. There is no Gartner release anywhere anchoring a "$1T AI market" figure; the closest published primary numbers are the $2.59 trillion total AI spending forecast from May 2026 and the "opportunity" framing on that vendor page. The same page also describes an "AI growth supercycle" projected to add up to $48 trillion to the global economy over the next fifteen years.

Treat the $1T label as stale site furniture rather than a Gartner position. The lesson is not that somebody forgot to update a menu. It is that trillion-scale numbers circulate as decoration, including inside the firm that produces them.

The boilerplate contradicts the SEC filing

Two inconsistencies live in the standard paragraph at the bottom of Gartner's AI press releases.

The expert count. The press releases say "more than 2,500 business and technology experts." The 10-K, which is signed and filed with the SEC, says "more than 2,400 business and technology experts and 920 experienced consultants" as of 31 December 2025. Gartner's own AI Hub says "2,400+." The marketing number is a hundred experts above the legally binding one.

The AI use-case count. The 15 January 2026 release says "more than 1,000 AI use cases and case studies." The 19 May 2026 release says "more than 4,000." Same paragraph, same claim, four months apart, four times the number, with no announcement that anything changed.

Neither is fraud and I am not calling it that. It is what happens when a stat block becomes a template that different teams edit at different times. But it tells you something about how much weight the stat block will bear, and the answer is: not much.

The AI Hub stat block, and the one number in it that is not an AI number

Gartner's AI Hub leads with a set of figures. I checked every one against the live page on 21 August 2026. They are all there, exactly as printed:

2,400+ analysts providing guidance
6,000+ AI-related written insights
4,000+ AI use cases and case studies
510,000+ client interactions
80K+ business executives and 10K+ IT executives engaged, plus thousands of technology providers
1M+ proprietary data points from client interactions, vendor briefings and proposal reviews
over 50 internal AI applications of Gartner's own

And the quote the page hangs it all on, from Daryl Plummer, Distinguished Vice President, Chief of Research: "By 2030, CIOs expect 75% of IT work will be done by humans augmented with AI, and 25% will be done by AI alone. That means by 2030, 0% of IT work will be done without AI."

Now the caveat, which is the only thing I would flag to a buyer. The 510,000+ client interactions figure is not an AI number. The 10-K uses the identical figure for all client interactions across the entire company in 2025: "our experts had more than 510,000 direct client interactions in 2025." On the AI Hub it appears inside an AI stat block with nothing qualifying it. A reader scanning that block comes away believing Gartner had half a million AI conversations last year. What Gartner disclosed is that it had half a million conversations.

The same care applies to the analyst count in the other direction. Gartner had 20,244 employees globally at the end of 2025. More than 2,400 of them are the experts. That is about 12% of the company, by my arithmetic on two disclosed figures. The rest is sales, conferences, consulting and operations, which is a perfectly normal shape for a subscription business and is only worth knowing because "2,400+ analysts" reads, in a stat block, like the whole firm.

A date trap in the Hype Cycles

If you cite a Hype Cycle placement, check which year's edition you are actually looking at.

As of today, the page at gartner.com/en/articles/hype-cycle-for-artificial-intelligence is still serving the 2025 edition, bylined July 2025. That is where "GenAI enters the Trough of Disillusionment" comes from, and where the two biggest movers are AI-ready data and AI agents, both at the Peak of Inflated Expectations, with ModelOps expected to reach the Plateau of Productivity. Those are 2025 placements on a page that reads as current.

The genuinely 2026 material sits on two other pages. The Agentic AI Hype Cycle article, 15 April 2026, puts agentic AI itself at the Peak of Inflated Expectations and adds governance, security and cost-control entries (agentic AI governance, agentic AI security, FinOps for agentic AI) spread across the curve rather than clustered. The GenAI Hype Cycle article, 22 July 2026, says large language models are the most mature technology on the curve and that GenAI virtual assistants and GenAI-enabled applications are nearest the Plateau of Productivity. Neither ungated article publishes a full placement map; those are behind the client gate.

One thing does carry forward independently. Gartner's January 2026 AI spending release states that AI is in the Trough of Disillusionment throughout 2026, and draws a commercial conclusion from it: AI "will most often be sold to enterprises by their incumbent software provider rather than bought as part of a new moonshot project." That is a 2026 statement, and for a small business it is the most directly useful sentence in the whole set. Your existing vendors are about to add AI to your bill.

Three more free numbers are worth having, all from that same 2025-edition page, so date them accordingly. Gartner reports an average spend of $1.9 million on generative AI initiatives in 2024, with fewer than 30% of AI leaders saying their CEOs are happy with the return. It says 57% of organisations estimate their data is not AI-ready. And it produces more than 130 Hype Cycles a year covering over 1,900 innovations, which should calibrate how much any single placement means.

What Gartner's own annual report says about Gartner

This is the part that makes the whole piece worth writing, and it is all in filings anyone can read on EDGAR for nothing.

Gartner's 2025 revenue was $6.50 billion, up 4% over 2024. The research business, renamed from "Research" to "Business and Technology Insights" in the second quarter of 2025, brought in $5.07 billion at a 77% gross contribution margin. Those are healthy numbers.

Then the direction. Second-quarter 2026 total revenue was $1.676 billion, down 1% year over year, with Insights at $1.290 billion, Conferences at $244 million and Consulting at $142 million. A chunk of that decline is explained: Gartner sold its Digital Markets business on 5 February 2026 for approximately $110.0 million.

Contract value, which Gartner defines as the annualised value of every subscription contract in effect at a point in time, was $5.155 billion at the end of 2025, up 1% excluding currency. At 30 June 2026 it was $5.283 billion, up 2% excluding currency. Growing, barely.

The number that actually matters is wallet retention, and it is below 100% in both sales segments.

Segment	Client retention, Jun 2026	Wallet retention, Jun 2026	Wallet retention, Jun 2025
Global Technology Sales	85%	97%	99%
Global Business Sales	86%	99%	104%

Across full-year 2025 the fall was steeper: GTS wallet retention went from 102% to 96%, a six-point drop, and GBS from 106% to 99%, a seven-point drop. GTS client retention went the other way, from 84% to 85%.

Read those two rows together. Clients are staying and spending less. The 10-Q says exactly that, in its own words: the decrease "was largely due to lower levels of spending by existing clients."

Then the risk factor, which I am quoting in full because paraphrase would soften it:

In addition, we face competition from free sources of information that are available to our clients through the internet. We anticipate encountering more competition with increased adoption of AI services in the markets in which we compete. Limited barriers to entry exist in the markets in which we do business.

That is the firm forecasting 47% AI spending growth, telling its shareholders that free information plus AI is coming for its own book of business. It is the thesis of this article, written by the subject of it, and filed with the SEC.

One more disclosure that got no press release. Gartner's Insights contract value with the US federal government was approximately $126.0 million at 31 December 2025, and less than half of the 31 December 2024 federal contract value was retained during 2025. A majority of a nine-figure book of business, gone in twelve months, disclosed inside a risk-factor paragraph.

If you want a sense of scale: dividing the $5.283 billion of contract value by the "over 13,000 distinct client enterprises" the 10-K describes gives roughly $406,000 per enterprise per year. That is my arithmetic on two disclosed figures, not a Gartner disclosure, and because "over 13,000" is a floor rather than a count, the true average is lower than that. Gartner does not publish an average client spend anywhere I could find in either filing.

What the subscription is actually for, said fairly

None of the above means Gartner sells nothing. It sells something real, and I would rather name it precisely than sneer at it.

Contracts are per-user, defined-term, with a minimum period of twelve months, and 77% of them were multi-year at the end of 2025. The majority are paid in advance. Gartner held 53 in-person conferences in 2025 with more than 83,000 attendees, including 12 Symposium/Xpo events. Its experts logged more than 510,000 direct client interactions.

What a buyer gets for that is not primarily a number. It is an outside name attached to a recommendation, an analyst who will take a call about your specific shortlist, and a defensible answer to "who else says this" when a committee is signing off on a seven-figure purchase. That is a genuine product and it is priced like one. If your problem is that six executives need to agree on a platform and nobody wants to own the decision alone, a research subscription solves your actual problem.

If your problem is "should my eleven-person firm bother with AI this year," it does not. You would be buying committee cover you do not need, at a price set by enterprises that do.

The Census Bureau runs the survey, and gives it away

Here is the free thing that answers the small-business version of the question better than any forecast will.

The Business Trends and Outlook Survey is run by the US Census Bureau against a sample of roughly 1.2 million businesses, in biweekly cycles. It asks whether the business used artificial intelligence in any business function in the prior two weeks, and whether it expects to in the next six months. The most recent published cycle at the time of writing is 202616: reference period 13 to 26 July 2026, collected 27 July to 9 August, published 13 August 2026. The next cycles were scheduled for 27 August and 10 September 2026, and the file addresses stay the same, so you re-download the same spreadsheet and get the new column.

The headline: 21.8% of US businesses said they used AI in any business function in the prior two weeks, with a standard error of 0.33%. 25.9% expect to in the next six months.

Be honest about the weakness, because it is a real one. The unit response rate for that cycle was 12%, about 24,500 responses out of the 1.2 million-business sample, and the estimates are non-response adjusted. That is low. What it has that vendor surveys do not is a sampling frame drawn from the actual population of US businesses. A Federal Reserve Board staff note put it plainly: the BTOS "generally mirrors the firm population distribution," so its firm-weighted estimate "is likely a strong representation of the AI adoption rate across all U.S. businesses."

The size gradient is not the one you have been told

This is the finding I did not expect, and it is the reason to read the size-class file rather than the headline. Question 7, current AI use, by employment size class, cycle 202616:

Employees	Using AI now	Expect to within six months
1 to 4	21.9%	25.8%
5 to 9	20.0%
10 to 19	20.2%
20 to 49	22.4%
50 to 99	27.8%	33.9%
100 to 249	30.3%	36.3%
250 or more	41.5%	50.6%

The smallest firms in the country use AI at essentially the same rate as firms with 20 to 49 employees, and at a higher rate than firms with 5 to 19. The gradient only appears above 50 employees. The trough is the 5-to-19 band, not the bottom.

That reframes the usual story. "Small business is being left behind" is not what this data says. What it says is that a solo operator or a four-person shop can pick up a chatbot with nobody's permission, and a twelve-person firm has enough process to make that awkward and not enough staff to run a project. Census's own analysts noted that between December 2025 and May 2026, AI use rose among firms with at least 20 employees but did not change significantly among firms with fewer than 20.

Anyone charting this series naively is manufacturing a hockey stick

On 17 November 2025 the Census Bureau changed the wording of the core AI question, from AI use "in producing goods or services" to AI use "in any business function." That is a break in the series.

Under the old wording the firm-weighted rate had crawled from 3.5% to about 10% between September 2023 and late 2025. Post-revision it reads around 18%, and 21.8% in the latest cycle. Under the current wording alone, the move is 17.3% in the first post-revision cycle to 21.8% now, about 4.5 points in nine months. That is a real trend and a modest one.

Census researchers attribute the late-2025 jump to three things at once: the revised wording, continued adoption during the federal funding lapse when no data was collected at all, and the introduction of the second AI supplement, which itself likely raised respondents' propensity to report use. Three confounds at one point in a series. If you see a chart running smoothly from 2023 to 2026 with no break marker on it, whoever drew it either did not know or did not care.

What "adopting AI" means for the median adopter

The 2026 BTOS AI Supplement pooled more than 117,000 distinct firms across six biweekly panels from 17 November 2025 to 8 February 2026. Two of its findings are worth more to an owner than any spending forecast.

64.3% of businesses that use AI made no changes at all to adopt it. No staff training, no hardware or software purchase, no new workflows, no vendor. For comparison: 15.4% developed new workflows, 15.0% trained current staff, 8.5% purchased cloud services, 3.8% used vendors or consultants, and 1.3% hired AI-trained staff.

95.7% of AI-using businesses reported no change in total employment from AI use over the prior six months. 2.3% reported an increase, 2.0% a decrease.

And on what the AI is doing: only 10.1% of AI-using firms used it to perform a task previously done by an employee, while 43.7% used it to supplement or enhance a task an employee already performed. 10.6% used it for a new task. 51.5% said none of the above.

Put those together and the median AI adopter in the United States is a business where somebody opened a chatbot. That is not a sneer. It is the single most useful calibration available, and it also explains why survey numbers diverge so wildly: the threshold for answering yes is close to zero.

Why non-adopters say no, and it is not the price

Among businesses with no plans to use AI in the next six months, 61.6% said "AI is not applicable to this business." Lack of knowledge of AI capabilities came in at 22.0%, privacy or security concerns at 20.7%, not mature enough at 13.0%, lack of a skilled workforce at 7.1%. Only 6.9% said it was too expensive, and only 3.2% said a previous attempt did not meet expectations.

"Not applicable" also falls steadily as firms get bigger, from 63.3% at 1 to 4 employees down to 42.2% at 250 or more. Cost is not the barrier and disappointment is not the barrier. Relevance is, and relevance is a perception that shifts with size rather than a fact about the technology.

If you are on the fence, those two numbers are your decision aid. The people who tried it mostly changed nothing to do so, and the people who declined mostly did so because they could not see the point, not because it burned them.

The same question, five answers, all defensible

Ask "how many US businesses use AI" and you get, depending on who you ask:

Source	As of	Figure	What kind of sample
Census BTOS, cycle 202616	July 2026	21.8% of firms	Probability sample, 1.2M frame, 12% response
Fed Small Business Credit Survey	Sept to Nov 2025	46% of small employer firms	Convenience sample, stated as not random, n=6,525
Atlanta Fed Survey of Business Uncertainty	November 2025	69.4% equal-weighted	Stratified, phone-recruited
Atlanta Fed SBU, employment-weighted	November 2025	78% of the labor force at AI-adopting firms	Same survey, weighted by employees
McKinsey State of AI	June to July 2025	88% of organisations	Online, n=1,993, GDP-weighted, 38% from firms over $1B revenue

None of these is lying. They weight by firm, by employee, or by whoever answered the email. A Federal Reserve review found that back in mid-2024, sixteen different surveys of work-related AI adoption produced point estimates ranging from about 5% to 40%. An eight-fold spread on the same phenomenon.

The Small Business Credit Survey is the one to watch closely, because it is the source of the 46% small-business figure that gets quoted everywhere, and its own methodology section says: "The SBCS is not a random sample; results should be analyzed with awareness of potential biases that are associated with convenience samples." It is still worth reading. Among its AI-using small firms, just 7% had fully integrated AI into business processes, about half described themselves as still experimenting, and 44% were partially integrated. Those same users self-report 71% increased productivity, 39% improved quality and 31% higher sales, with most reporting no change in labor costs.

Even Stanford's AI Index, which republishes McKinsey's 88%, attaches this caveat to it: the results "are self-reported and should be viewed as directional rather than comprehensive." When the index republishing a number will not stand behind it as a measurement, that tells you how to use it.

There is also a gap of roughly two and a half times between asking firms and asking workers. BTOS puts employee generative AI use at roughly a fifth of firms. The Census Household Trends and Outlook Pulse Survey asked workers directly in March 2026 and 55% said they had used AI on the job for at least one of eleven tasks. Both are Census probability samples. The Census working paper's explanation is that "worker task use sometimes occurs without formal firm-level adoption." That is shadow AI, measured by the government, and it is happening in your business whether you have a policy or not.

Two studies, same weeks, opposite conclusions

The most-quoted AI failure statistic in circulation is MIT NANDA's finding that 95% of organisations are getting zero return from generative AI, against $30 to $40 billion of enterprise investment, with only 5% of integrated pilots extracting millions in value.

Its methodology, from its own notes page: a review of over 300 publicly disclosed AI initiatives, structured interviews with 52 organisations, and 153 survey responses collected at four industry conferences, over January to June 2025. The report carries its own disclaimer that the figures are "directionally accurate based on individual interviews rather than official company reporting" and that "success definitions may differ across organizations." MIT no longer hosts it at its original address; that address now serves the NANDA group's overview page, so I read the report from a third-party mirror rather than from MIT. Both of those are real weaknesses in the citation.

Now the Wharton and GBK Collective enterprise study, fielded 26 June to 11 July 2025, n=801: "Nearly three-quarters already see positive ROI." Same country, overlapping weeks, opposite conclusion.

The resolution is in the screening criteria. Wharton required respondents to be senior decision makers at US organisations with 1,000+ employees and more than $50 million in revenue. No small business is in that sample at all. And even inside it, belief in ROI tracks distance from the work: 81% of VP-and-above respondents believe ROI is positive versus 69% of mid-managers.

If you only carry one "AI is not paying off yet" statistic, do not carry the 95%. Carry this one instead. The Atlanta Fed and Bank of England surveyed nearly 6,000 CFOs, CEOs and executives from stratified firm samples across the US, UK, Germany and Australia, and more than 80% of firms reported no impact from AI on either employment or productivity over the past three years. Broken out by country, the US sits at the top of that range, with about 91% of US firms reporting no impact against an 89% four-country average. The estimated average realised productivity gain across all firms is about 0.29%. Same headline as the 95%, roughly forty times the sample, and a real sampling frame.

The same executives forecast much larger effects ahead than they have observed: AI boosting productivity 1.4%, raising output 0.8% and cutting employment 0.7% over the next three years. Employees surveyed separately predicted a 0.5% employment increase.

A Richmond and Atlanta Fed survey with Duke of nearly 750 corporate executives named the gap directly, calling it a productivity paradox in which "perceived productivity gains are larger than measured productivity gains." CFOs' perceived AI-attributed labor productivity gain averaged 1.8% in 2025. Measured against their own reported revenue and headcount, the implied figure is about 0.8% in high-skill services and finance and roughly 0.4% in low-skill services, manufacturing and construction.

What the official statistics do and do not say

The Bureau of Labor Statistics publishes no productivity statistic attributable to AI. It says so on its own page: BLS "implicitly captures AI use through its capital measure of software used in production." Anyone citing a BLS AI productivity figure is citing something BLS does not produce.

What BLS does show is the money going in. Software investment grew at an 11.1% compound annual rate from 2019 to 2024, up from 7.9% over 2007 to 2019. Total factor productivity over 2019 to 2024 rose 1.1%. Capital in, output not yet visible. That is the productivity paradox in two official numbers and it costs nothing to read.

On the spending side, the concentration is extreme and almost nobody reports it. More than half of firms surveyed by the Atlanta Fed expect to spend no more than $200 per employee on AI in 2026. The top 10% plan at least $2,800 per employee. Any aggregate AI-spend figure you see is describing the behaviour of a handful of firms, not the median one.

Which is worth holding next to Gartner's own framing. Gartner says AI infrastructure will be the largest segment of the AI market, over 45% of all AI spending, "which will be driven by vendors." And it says enterprises "have yet to really flex their spending potential," calling 2026 the inflection year. Read carefully, the largest single slice of the trillions is vendors buying compute from each other. Gartner also observes that organisations "show limited appetite for using AI to drive disruptive enterprise change" and instead "favor tactical AI initiatives with incremental improvements in efficiency and productivity." Census's 64.3% who changed nothing is the same finding, arrived at from a probability sample rather than an analyst's read.

The one Gartner finding aimed at firms anywhere near your size

There is no Gartner press release on AI spending by small or midsize businesses specifically. The nearest published thing is a May 2026 analysis of 101 efficient-growth companies, which found that high-performing firms under $3 billion in revenue deployed twice as many AI use cases as comparable peers. Gartner's framing is that AI acts as "a scale multiplier when resources are constrained."

Note that "under $3 billion in revenue" is Gartner's idea of small. It is not yours. But the direction of the finding, that the constrained firm gets more out of many small deployments than one large one, matches what the Census supplement shows about how adoption actually happens.

What I would actually read, in order

Six free sources. About an hour in total, and you will know more about whether AI matters for a business your size than any forecast will tell you.

Census BTOS, Employment Size Class file. Answers: do firms my size actually use this, and is that changing? Look at Question 7 and Question 24 in the latest column. Download at census.gov/hfp/btos/downloads/Employment%20Size%20Class.xlsx. New column roughly every two weeks.
Census BTOS AI Supplement table. Answers: what does adopting actually involve, and what happened to the firms that did it? Questions 3, 6, 7 and 13 carry the "made no changes," "no employment change" and "not applicable" figures. census.gov/hfp/btos/downloads/AI_Supplement_Table_2026.xlsx.
The Fed's "Monitoring AI Adoption in the U.S. Economy" note. Answers: why do the numbers I keep seeing disagree so much, and which one should I believe? federalreserve.gov, 3 April 2026.
Atlanta Fed Working Paper 2026-3, "Firm Data on AI." Answers: has this shown up in anyone's actual results yet? Nearly 6,000 executives, stratified samples, four countries.
Atlanta Fed macroblog on AI spending per employee. Answers: what would a normal firm spend, as opposed to what the aggregate implies? The median and top-decile split is the whole point.
Gartner's own newsroom. Answers: what is the vendor-side money doing, and what is the current sales motion? The press releases are ungated and the forecasts are in them. Read the date on every one, and check whether a newer revision exists before you quote a figure.

Then, on Monday: pick one task somebody in your business already does by hand every week, and try it. That is what 64.3% of adopters did. It cost them nothing to change.

If you want the longer version of the argument that outside firms charge retainer prices for work that is now a subscription and an afternoon, I wrote a short Kindle book about it called The $20 Dollar Agency. This post is the free version of the same reasoning, pointed at research subscriptions instead of marketing ones.

Fact-check notes and sources

The four 2026 IT spending forecasts are each from a Gartner newsroom release: 22 October 2025, $6.08 trillion at 9.8%, 3 February 2026, $6.15 trillion at 10.8%, 22 April 2026, $6.31 trillion at 13.5%, and 27 July 2026, $6.37 trillion at 14.2%. The $287 billion 2026 IaaS figure is in the July release, alongside the $822 billion Data Center Systems forecast. The roughly $290 billion revision is my subtraction of the first figure from the last, not a Gartner disclosure.
The AI spending forecasts are 15 January 2026, $2.52 trillion for 2026 and $3.34 trillion for 2027 and 19 May 2026, $2.59 trillion and $3.49 trillion. The over-45% infrastructure share, the "driven by vendors" phrasing, the inflection-year quote and the "limited appetite for disruptive change" quote are all from the May release. The Trough of Disillusionment sales-motion quote is from the January release.
The 40% agentic cancellation prediction and its webinar-poll basis are both in the 25 June 2025 release, which also carries the "about 130 of the thousands of agentic AI vendors are real" estimate. The separate 2027 demote-or-decommission prediction is in the 26 May 2026 governance release. These are two different predictions and are routinely merged in secondary coverage. The 30% abandonment prediction is from 29 July 2024; the less-than-one-in-five prediction is from 29 September 2025; the at-least-50% budget overrun is from the 2026 GenAI Hype Cycle article, 22 July 2026.
The description of the IT spending forecast as resting on analysis of the sales of over a thousand vendors is Gartner's own account of its forecast methodology as published with its newsroom releases, not an independent audit. I am relying on the wording as published. I have not seen the underlying vendor list, and nobody outside Gartner has. Gartner publishes that methodology line alongside its newsroom forecasts rather than on a dedicated page, so there is no single deep link for it; the wording appears with the 27 July 2026 release.
The navigation label conflict was observed on gartner.com on 21 August 2026: the site menu says "Competing in the $1T AI market" while the destination page, Race Ahead With AI, says "4.7-trillion-dollar AI market opportunity" and carries the $48 trillion supercycle framing. Both are marketing copy aimed at technology vendors rather than forecasts in the sense the press releases are, and no Gartner release anchors a $1T AI market figure.
The boilerplate inconsistencies are between the 27 July 2026 release ("more than 2,500 business and technology experts") and the 10-K for fiscal 2025 ("more than 2,400 business and technology experts and 920 experienced consultants"), and between the January 2026 release ("more than 1,000 AI use cases") and the May 2026 release ("more than 4,000"). Both pairs are verbatim.
The AI Hub stat block and the Daryl Plummer quote were read on the live page at gartner.com/en/ai on 21 August 2026 and every figure listed here appears there. These are Gartner's own marketing claims about itself, not independently established facts. The caveat about the 510,000+ figure is mine: the 10-K uses the identical number for all client interactions company-wide in 2025, and the AI Hub does not qualify it. The 12% analyst share is my arithmetic on the 10-K's 2,400 experts and 20,244 total employees.
Hype Cycle placements need a date check. The page at gartner.com/en/articles/hype-cycle-for-artificial-intelligence is still serving the July 2025 edition as of 21 August 2026, and that is the source of the GenAI Trough placement, the AI-ready data and AI agents peak placements, the ModelOps plateau expectation, the $1.9 million average 2024 GenAI spend, the under-30% CEO satisfaction figure, the 57% not-AI-ready figure and the 130 Hype Cycles over 1,900 innovations counts. The 2026 material is on the Agentic AI Hype Cycle article, 15 April 2026 and the GenAI Hype Cycle article, 22 July 2026. The full placement maps are behind Gartner's client gate; everything cited here comes from the ungated summaries.
Gartner's financials are from the Form 10-K for fiscal year 2025, filed 12 February 2026 and the Form 10-Q for the quarter ended 30 June 2026, filed 4 August 2026. Revenue, contract value, client and wallet retention, the "lower levels of spending by existing clients" explanation, the free-information and AI risk factor, the $126.0 million federal contract value with less than half retained, the Digital Markets sale at approximately $110.0 million, the 20,244 headcount, the 53 conferences and 83,000 attendees, the 77% multi-year contract share and the paid-in-advance description are all in those two filings. The roughly $406,000 per client enterprise is my arithmetic on contract value divided by the disclosed "over 13,000" client enterprises; because that count is a floor, the real average is lower, and Gartner does not publish a per-client average.
BTOS figures are from the Census Bureau's own downloads for cycle 202616, published 13 August 2026: National.xlsx, Employment Size Class.xlsx and URR.xlsx for the 12% response rate. The AI supplement figures come from AI_Supplement_Table_2026.xlsx, which pooled more than 117,000 distinct firms between 17 November 2025 and 8 February 2026. The size-class letter codes are defined in the data dictionary sheet of the size class file. This is a probability sample of businesses, which is why I lead with it, but the 12% unit response rate is a genuine limitation and the estimates are non-response adjusted rather than raw.
The November 2025 question-wording change and the pre-revision 3.5% to 10% trend are documented in Census's own America Counts article, 26 May 2026 and in CES Working Paper 26-25, "The Microstructure of AI Diffusion", which also lists the three simultaneous causes of the late-2025 jump and the employment-weighted figures. Charting BTOS across that date without a break marker overstates the trend badly.
The survey comparison table: the Fed staff note is Jeffrey S. Allen, "Monitoring AI Adoption in the U.S. Economy", 3 April 2026, which carries the Survey of Business Uncertainty figures, the 5% to 40% spread across sixteen earlier surveys, and the assessment of BTOS as the best available firm-level estimate. The 46% small-business figure and the not-a-random-sample statement are both from the 2026 Report on Employer Firms, n=6,525, fielded September to November 2025. McKinsey's sample size, fielding window and weighting, and the 88% and 79% adoption figures, are as reproduced in the Stanford HAI 2026 AI Index: the 88% and 79% adoption figures from its economy chapter, and the sample size, fielding window, weighting and the "directional rather than comprehensive" caveat from the full report PDF. McKinsey's own page would not open for me, so I have not quoted any McKinsey figure the AI Index does not reproduce, and I have deliberately left out the widely cited EBIT-impact numbers because I could not verify them at source.
The MIT NANDA 95% figure: its methodology and its own limitations disclaimer are quoted from the report itself, read from a third-party PDF mirror because MIT no longer serves the report at its original address, which now returns the NANDA group's overview page. That is a weak citation chain and it is one reason I would not build an argument on the 95%.
The stronger "no impact yet" evidence is Atlanta Fed Working Paper 2026-3, "Firm Data on AI", and the productivity paradox is documented in Atlanta Fed Working Paper 2026-4. The Wharton figures and the 1,000+ employee screening criterion are from Accountable Acceleration, October 2025.
BLS: the statement that AI is captured only implicitly is on the BLS productivity and AI page, and the software investment and total factor productivity figures are from AI and the rise of software investment, Monthly Labor Review, May 2026. The per-employee AI spending concentration is from the Atlanta Fed macroblog, 6 May 2026.
Worker-side figures: the 55% of workers using AI on the job is from the Census Bureau's March 2026 Household Trends and Outlook Pulse Survey coverage, 11 August 2026.
The efficient-growth finding for firms under $3 billion in revenue is from Gartner's 29 May 2026 CFO release, based on an analysis of 101 efficient growth leaders. No Gartner press release covers AI spending by small or midsize businesses specifically; the only research document that would is behind the client gate, and I have not read it or quoted any figure from it.

Four Forecasts for One Year in Nine Months. What the Census Bureau Tells You for Free.