What Companies Actually Spend on AI Agents

Every figure in this post comes from a primary document: the original survey writeup, the analyst release, the academic paper, or the government data series. We opened every URL on August 24, 2026 and confirmed the number. Aggregator roundups, the pages titled 50 AI Statistics for 2026, were excluded as sources entirely, because most of their numbers cannot be traced past a chain of blogs citing each other.

Four rules governed what got in. Every number carries its population and field dates in the same sentence, because $186 million averaged across billion-dollar enterprises and a small-business budget are different facts wearing the same currency. Vendor figures about the vendor's own product are cited where they are honest, labeled every time, and never counted as independent evidence. Nothing is averaged: where two credible sources disagree, both numbers are here with the difference named. And an unknown stays unknown. Where a survey did not state its sample, we say so instead of guessing.

One mechanical thing worth knowing. Gartner, BCG, PwC, IBM, McKinsey, the OECD, and the Bureau of Labor Statistics all serve human readers fine and block automated fetches. Every figure from those publishers was confirmed against multiple independent quotations of the named document before it went in.

What would change our conclusion: a primary source publishing an actual distribution of per-project agent build costs, or a large-sample independent survey in which most organizations can demonstrate agent ROI. As of August 2026, neither exists. We looked.

The organization-level spend numbers are the best-documented part of this whole subject, and they are enormous.

KPMG's Q1 2026 global pulse, fielded February 17 to March 17, 2026 across 2,110 C-suite leaders at organizations with revenue above $100 million, put average planned AI investment for the next 12 months at $186 million. The 237 US respondents in the same survey came in higher at $207 million, and the technology-sector respondents projected $294 million. One survey, three cuts, and a $108 million gap between the global average and the tech sector, which is worth remembering the next time a single average gets quoted at you.

The share-of-budget numbers all point the same direction. BCG's AI Radar, fielded across 2,360 executives including 640 CEOs in 16 markets and published January 2026, has corporate AI spending roughly doubling from 0.8% of revenue in 2025 to about 1.7% in 2026, with the leading cohort of CEOs directing around 60% of their AI budget specifically to agentic AI. IBM's Institute for Business Value, surveying 2,000 senior technology executives across 33 geographies between January and April 2026, has AI at just under 15% of IT budget in 2025 and headed for nearly 25% by 2027. Deloitte's May-June 2025 survey of 548 US decision-makers found AI already taking 36% of the average digital budget, inside digital budgets that jumped from 7.5% to 13.7% of company revenue in a single year.

And the intent to keep spending is nearly unanimous: 88% of the 308 US executives PwC surveyed in April 2025 said their function plans to increase AI budgets in the next 12 months because of agentic AI, and more than a quarter of them plan increases of 26% or more.

One caution before you build a plan on any of this. Gartner forecast worldwide AI spending of $2.5 trillion for 2026 in January, then revised it to $2.59 trillion in May. When the best-resourced forecaster in the industry moves its own number by $90 billion in four months, treat every forecast in this post as a direction, not a coordinate.

Agent economics has one genuinely strange property. The markets are priced to the decimal, and the projects are not priced at all.

Gartner sized the enterprise AI coding-agent market at $9.8 to $11 billion annualized as of April 2026. Its forecast for AI agent software spending runs $86.4 billion in 2025, $206.5 billion in 2026, and $376.3 billion in 2027, and its AI services forecast reaches $609 billion by 2028 at a 21.4% five-year compound growth rate. The consulting market those services live inside hit $397 billion in 2024, up 4.5%, with AI demand named as the growth driver. BCG's February 2026 analysis, built on interviews with over 115 enterprise executives and 75 technology service provider executives, adds a $200 billion net uplift to the tech services market over five years from agentic AI alone, and notes agentic AI investment has grown more than 60% annually since 2023. On the capital side, the OECD's Preqin-based tally has 61% of all global venture capital, $258.7 billion, going to AI firms in 2025, up from a 30% share in 2022.

Not one of those documents tells you what your agent will cost. That absence is not an accident of our search. It is the finding: analysts price markets because vendors report revenue, and nobody reports per-project build costs because nobody is obliged to publish them.

What the market data does tell a buyer is where pricing is headed. Gartner's May 2026 pricing research notes agentic workloads burn 5 to 30 times the tokens of a standard chatbot interaction, and its June 2026 prediction is blunt. By 2028, per-developer spend on AI coding tools will exceed the average developer's salary as consumption-based licensing spreads. GitHub moved Copilot from seat pricing to an access-plus-consumption model on June 1, 2026, a shift Gartner's pricing research tracks. The era of predictable per-seat agent pricing is ending while the invoices are still small. Budget accordingly.

The run-cost data is where the published record gets genuinely counterintuitive, because both of these are true at once: tokens have never been cheaper, and running agents has never cost more.

The collapse first. OpenAI listed GPT-3 at $60 per million input tokens in June 2020; by February 2025 Google listed Gemini 2.0 Flash at $0.10, a 600-fold drop documented in an arXiv analysis of vendor list prices. Stanford's AI Index, cited by Deloitte's 2026 infrastructure work, has inference costs down 280-fold in two years. Gartner predicts that by 2030, running inference on a trillion-parameter model will cost providers over 90% less than it did in 2025. And an NBER working paper from December 2025, studying the LLM API market across OpenRouter and Azure, found open-source models priced about 90% below closed-source models of matched capability.

Now the other direction. Gartner's August 2026 forecast has inference cost per agentic workflow rising more than fivefold through 2028, because agents do not have conversations, they run loops, burning the same 5 to 30 times the tokens of a chatbot exchange that its pricing research flagged earlier in the year. An arXiv study tracking frontier models from April 2024 to November 2025 measured per-token prices falling 5 to 10 times a year while the total cost of running frontier-level models rose 3 to 18 times a year. McKinsey's cost-of-intelligence research found token usage varying up to 30x across runs of the same task, and IBM's Institute for Business Value had executives expecting an 89% rise in computing costs between 2023 and 2025, with every single executive surveyed reporting at least one generative AI initiative canceled or postponed over cost.

The volumes underneath are industrial. In Deloitte's November-December 2025 survey of 515 US decision-makers at organizations above $500 million revenue, 37% of enterprises were consuming 1 to 10 billion tokens a month and another 30% were above 10 billion, with 86% expecting AI infrastructure budgets to rise and typical budgets projected to more than triple over three years. Gartner expects worldwide inference spending on AI-optimized infrastructure to hit $23.3 billion in 2026, overtaking training spend at $19 billion. And the spend is lopsided: Accenture's tokenomics work found fewer than 10% of users and workflows driving the vast majority of an organization's AI bill.

So when a cost guide quotes you a tidy monthly run cost, remember what the primary record actually shows: a 63-incident catalog of LLM-agent budget overruns now exists on arXiv, compiled across 21 software projects between 2023 and 2026. Overrun is common enough to have its own literature.

This is the section the cost guides skip, and it is the one that should drive your decision.

Start with budget discipline: McKinsey's cost-of-intelligence research has 93% of organizations exceeding their AI budget, a figure it published without a stated sample, so weight it as a signal rather than a survey point. In KPMG's Q2 2026 pulse of 2,145 senior leaders across 20 markets, 42% admitted only partial visibility into their AI spending and 33% named poor understanding of usage costs as a key deployment challenge. Only 7% of that KPMG sample had established the ROI of their AI spend at all.

The realized-return numbers are worse than the budget numbers. BCG's October 2024 survey of 1,000 executives across 59 countries found 74% of companies had yet to show tangible value from AI. Deloitte's August-September 2025 survey of 1,854 senior executives across 14 European and Middle Eastern countries found 6% of organizations achieved AI payback within a year, 15% reporting significant measurable ROI from generative AI, and just 10% seeing significant returns from agentic AI specifically. MIT's Project NANDA, reviewing over 300 public enterprise generative AI initiatives in its July 2025 State of AI in Business report, put the share of pilots producing measurable P&L impact at 5%. IBM's February-April 2025 study of 2,000 CEOs found only 25% of AI initiatives had delivered their expected ROI, and only 16% had scaled enterprise-wide. We covered the ROI record in depth in how much ROI companies actually see from AI; the short version is that the spend curve and the proof curve have not met.

Even the activity-to-output ratio is now measured. An NBER working paper published May 2026, tracking over 100,000 GitHub developers, found AI coding tools raised coding activity 180% but the number of projects only 50% and actual shipped releases just 30%. Producing more code is not the same fact as shipping more product, and the gap between those numbers is where a lot of agent spend quietly dies. Gartner's January 2025 poll of 3,412 webinar attendees found only 19% making significant agentic investments, with 31% waiting, and its June 2025 prediction says over 40% of agentic AI projects will be canceled by the end of 2027 on escalating costs, unclear value, or inadequate controls.

One correlation in this pile is genuinely useful. In KPMG's Q2 2026 data, organizations with real-time visibility into AI costs reported established ROI at 15%, against 3% for those without: five times the rate. That correlates with paying attention; nobody has run the experiment that proves the tracking causes the return. What would have falsified it is visibility-rich organizations reporting the same 3%, and that is not what the data shows. Either way, the cheapest insurance in this entire post is a cost dashboard.

Every build-versus-buy decision on agents eventually collides with US wage data, so here it is, from the government series rather than a salary marketing page.

The Bureau of Labor Statistics puts the May 2024 median wage for software developers at $133,080 and for computer and information research scientists, the closest category to an ML engineer, at $140,910. The 2025 BLS wage data, republished with attribution by O*NET, has software developers at a $135,980 median, data scientists at $120,230, and research scientists at $140,300. Those are medians for the whole occupation, before the AI premium.

And the premium is real and growing. PwC's 2026 AI Jobs Barometer, built on a pool of more than a billion job advertisements, puts the average wage premium for AI-skilled workers at 62%, up from the 56% its 2025 edition reported, ranging from 118% in consumer markets down to 16% in government. Employers say they will pay it: in KPMG's April-May 2026 pulse of 204 US leaders at billion-dollar firms, 40% would offer a 6 to 10% salary premium for strong AI skills and another 38% would offer 11 to 15%. Deloitte's 2022-2023 tech talent research, older but still the best-documented scarcity figure, had nearly 90% of tech leaders calling recruiting and retention a moderate or major challenge.

So an in-house agent team means paying a $135,980 median plus a premium above it, for people nearly everyone is competing to hire. That is the alternative every agency quote and platform subscription is actually priced against, and we walked through that comparison properly in agency versus in-house hire. Renting the capacity for the length of the build, the way a forward deployed engineer engagement works, exists precisely because that search is slow and that premium is real. Worth holding beside it: the best independent evidence on what those skills produce, an NBER field experiment across 5,179 customer support agents, measured a 14% average productivity gain from AI assistance, 34% for novices and roughly nothing for the most experienced. The wage premium is priced for the promise. The measured gain is smaller and lands unevenly.

A chunk of the figures circulating about agent value come from the companies selling the agents. The honest way to handle those is to label each one and set an independent measurement beside it.

IBM reports its internal AskHR agent handled 11.5 million interactions in 2024 with HR transactions 75% faster and HR operating costs down 40%. Salesforce's customer story on Wiley reports a 213% return on investment and $230,000 saved. Databricks reports multi-agent workflow usage up 327% in four months on its own platform. None of those numbers is fake. Every one of them was produced by the organization with the strongest possible interest in the number being large, measured on a deployment it controlled.

The independent side of the ledger reads differently. The NBER experiment found a 14% average gain across 5,179 agents in 2020-2021. Deloitte's 1,854-executive survey from late 2025 found 15% with significant measurable generative AI ROI. BCG's December 2025 developer research found more than 85% of individuals stuck at shallow adoption stages and fewer than 10% reaching semiautonomous collaboration.

When a vendor's number sits 10x above the independent one, read it as the distance between a showcase deployment and the median buyer's outcome, and budget your project on the median.

A benchmark that only reports the good data is a brochure, so here is the bad data, named.

The build ranges, $5,000 to $180,000 on one page and $20,000 to $500,000 on the next. Every agency cost guide on the first page of this query publishes one, no two pages agree, and none cites a source, because there is no source. The most-cited page we reviewed carries over 40 cost figures with two citations between them. Those ranges exist to anchor your expectations. Nobody measured anything.

The claim that a build is 25 to 35% of three-year total cost of ownership. It circulates through the same cost guides, always without a traceable document. It might even be roughly right, but nobody has published the measurement, so treat it as folklore.

The claim that 59% of organizations are actively budgeting for AI agents in the next 12 months. Widely repeated, and we could not trace it past aggregator pages citing each other. Dropped.

Adoption statistics quoted without their question wording. Published agent adoption runs from 17% of 2,501 CIOs (Gartner, mid 2025) to 79% of 308 US executives (PwC, April 2025) depending entirely on who was asked and whether the verb was deployed or adopting. We took that spread apart figure by figure in how many companies have actually deployed AI agents, so this post does not relitigate it. The short of it is that the wording of the question predicts the number before you read it.

Even the government series carries a wording asterisk. The Census Bureau's BTOS survey measured US firms using AI at about 10% in late 2025 under its original question, then redesigned the question in November 2025 and measured 17% on the same population. When one wording change nearly doubles a federal statistic, quoting any adoption or spend number without its question wording amounts to laundering it.

The most useful thing this research pass produced is the list of numbers that do not exist. Seven questions, dodged by every primary source we opened:

What one production agent costs to build. Org-level totals, market sizes, and price predictions all exist in quantity. A published distribution of per-project build costs does not.

What implementation partners charge. No analyst, survey, or government series prices an agent engagement. The only public numbers are the ranges agencies invent about each other.

Run cost per agent per month. Token volumes and organization totals are published; a per-agent operating cost is not.

What small businesses spend. Every spend survey in this post samples organizations above $100 million, and usually above $1 billion, in revenue. The overwhelming majority of companies that exist are absent from all of it.

The build-run-maintain split. No source separates what it cost to ship from what it costs to keep alive, which is exactly the split Gartner's cancellation prediction says is killing projects.

The cost of the failures. Over 40% of agentic projects are predicted to be canceled by the end of 2027, and nobody has published what the canceled cohort had already spent.

How long a build takes. There is no published distribution of agent delivery time, and time is most of what the labor cost above actually prices.

We publish our own delivery numbers precisely because this gap is real: our pricing page commits to a fixed number in public, a starter build at $1,500 to $2,500 and a two-week production sprint at $5,000, because in a market with no published per-project data, a checkable price is the only honest kind.

Four moves fall straight out of the data.

Budget the run before the build. The unit price of tokens will keep falling and your bill will rise anyway, because agents multiply consumption faster than prices drop. Gartner's fivefold per-workflow forecast through 2028 and the 5-30x token multiplier are the two numbers to keep. Any quote that names a build price and waves at the running cost is quoting you half a system.

Instrument the spend from day one. The 7% who established ROI in KPMG's May 2026 pulse were disproportionately the ones who could see their costs in real time, at five times the rate of those who could not. A cost dashboard is the cheapest correlate of success in the entire published record, and an eval harness is how you measure the output side of the same question.

Read every range as an anchor somebody set for you. No primary source publishes per-project agent costs, so every range you are quoted is either the seller's invention or the seller's ai agent development cost commitment in writing. Only one of those can be checked afterward, so ask for the fixed number and the runtime assumptions behind it.

And size the bet to the failure data, not the vendor case study. With 6% of organizations seeing payback inside a year in Deloitte's 1,854-executive survey and a 40%+ predicted cancellation rate, the winning shape is a small scoped build that either proves itself in weeks or dies cheaply. That is the shape we sell, and it is also just what the numbers say.