A program of EffectiveSC
How We Use AISources checked October 24 to 28, 2026
Where things stand / Keeping up with the Joneses
Most companies using AI are catching up with their rivals or keeping pace, not pulling ahead, and in this site's judgment much of their token spend works like a tax unless it builds something they own and reuse.
Robb Wilson, who leads this research, works at OneReach.ai, which sells software for orchestrating AI agents, and this page's estimate has agents taking over more of the work. EffectiveSC, which publishes How We Use AI, plans to offer memory and routing services for AI, and this page's judgment, that token spend builds an edge when it leaves behind shared memory and when routine work goes to cheaper models, favors those services. Claude, an Anthropic model, helps research this site, and Anthropic appears in the figures below as a maker of agents and a seller of tokens. The layers, the judgment and the estimate are this site's own reading; the cards are what others measured or said.
Two consultancies sort companies by how far along they are with AI, and they land on a similar split: a small group setting the pace, a larger group keeping up and a majority still catching up. Both sort firms by what people say about their own organizations, not by company accounts. A third survey shows why firms keep spending before they see a return: they fear falling behind.
5%
BCG classes 5% of companies as future-built, 35% as scaling and 60% as laggards; future-built firms show 1.7x the revenue growth and 3.6x the three-year shareholder return of laggards.
BCG's tiers come from executives rating their own companies' maturity. The revenue and shareholder-return gaps are associations, not effects.
Disclosure: BCG sells consulting on AI.
about 6%
In McKinsey's survey of May and June 2026, nearly 9 in 10 respondents (89%) reported regular AI use in at least one business function and 80% said AI had improved their own productivity, but only 37% said AI had contributed to their organization's EBIT, about the same as the 39% of 2025; about 6% are high performers, who attribute 5% or more of EBIT to AI and report significant value, and nearly three quarters of them redesigned workflows against about a quarter of other respondents.
Who was asked: 1,719 participants at all levels in 97 nations
Respondents at every level, from executives to individual contributors, reporting on their own organizations; not company accounts. The share saying AI added to EBIT covers the whole sample, not one layer of firms.
Disclosure: McKinsey sells consulting on AI.
64%
64% of CEOs acknowledge that fear of falling behind drives investment in some technologies before they clearly understand the value.
Who was asked: 2,000 CEOs in 33 countries and 24 industries
Chief executives agreeing with a statement in a survey; it doesn't say how much of their spending the fear drives.
Disclosure: IBM, which ran the study with Oxford Economics, sells AI software and consulting.
See also: The companies out front, Are companies getting a return on AI?
These are the project's names for the split above. The behaviors are the project's summary of how the same surveys describe each group; they aren't measured separately, and nobody has yet tested whether the layer a firm is in predicts its results. Firms catching up hold AI as table stakes: licenses, approved tools and a policy, AI handed to middle managers, pilots that rarely reach production, value seldom tracked and almost no agents. No one there can name a workflow AI changed. Firms keeping up have rolled seats out widely and scaled a few uses. Their people feel more productive, but the gains are the ones every rival also gets, so usage rises while unit costs and margins don't. This is doing enough not to lose, and it is where a firm can most easily believe it is pulling ahead. The survey this site is preparing will set what organizations do against how they rate themselves; no published survey measures that gap yet. The Joneses run a program their chief executive sponsors, with value targets. They redesign core workflows end to end rather than add AI to old ones, track value, train most of their people, use agents more than anyone else and spend more on AI, not less. Their gains show up in margins or growth and rest on things the firm owns. The layers describe firms, not how much of the work AI does on its own: a firm can lead on licenses while a person still does every step.
See also: How many run agents, and how much they let them do
This is the project's own reading, not a measurement. Token spend works like a tax when it buys output any rival can buy at the same price: drafts, summaries and searches that every competitor's staff also get. The gains then show up as parity, or as lower prices for customers, rather than as margin. It builds an edge only when it leaves behind something the firm owns and reuses and a rival can't rent: workflows redesigned around agents, memory of what the firm has learned and corrected, and its own tests of what good work looks like. The cards show the tax side. The price of a token keeps falling, yet bills rise because each task uses more tokens. One company's coding agents ran through a year's budget in months, staff ran needless work to climb an internal leaderboard, and most of the custom assistants people built and shared drew little use. So the cut worth making is not total spend. It is the burn that leaves nothing behind: tokens spent for show, frontier models doing routine work a smaller model could do, and agents rediscovering what the firm already knows. Budget by accepted outcome, not by token or seat. The edge side is thinner: firms that redesigned their work report better results, but that is an association, and no study yet measures whether a firm's own assets lower the tokens it needs per accepted result.
9x to 900x a year
At a fixed capability, the price per token fell 9 to 900 times a year depending on the benchmark (March 2025); Epoch's newer per-task estimate is about 13x a year (September 2026).
The price of the cheapest model that reaches a given benchmark score. How fast it falls depends heavily on the benchmark.
4.5x
Per-token cost halved from December 2024 to December 2025 while tokens consumed grew 4.5x; separately, Bain finds cost per task often flat because tasks use more tokens and firms stay on frontier models.
Bain cites a bank's data and gives no base for the count of tokens. Flat cost per task is its separate observation.
Disclosure: Bain sells consulting on AI.
about four months
Uber used up its full-year 2026 AI budget within about four months, driven by AI coding tools such as Claude Code, and now caps each employee at $1,500 a month per AI coding tool.
Reported by The Information and Bloomberg, from Uber's chief technology officer and a spokesperson; both reports are behind paywalls.
Disclosure: Claude Code is Anthropic's product, and Claude helps research this site.
An employee-built Meta dashboard ranked the top 250 token users among more than 85,000 staff; its builder took it down in April 2026 after its data leaked, and Meta said it didn't ask.
A news report of one internal dashboard. It shows token counts treated as a measure of standing, not what the tokens produced.
Amazon switched off KiroRank, an informal employee-built beta that ranked staff by AI use, after some staff had agents run needless tasks to climb it, adding compute cost.
A news report with Amazon's confirmation. The leaderboard was an informal test version a group of employees built, never formally approved.
28.5%
Of about 334,000 public custom GPTs in early 2024, 28.5% had more than 10 conversations; the store's top-listed GPTs drew about a third of conversations.
Public GPTs in OpenAI's store, counted early in its life. It says nothing about how often companies reuse what their own staff build.
See also: Judgment: What this means for a company, Spend and tokens
These are forecasts and a model, not measurements. AI is taking on longer tasks, more of the work runs without a person typing each step, and an agent burns many times the tokens of a chat, so forecasters expect token use to grow far faster than the number of people using AI. Their forecasts disagree widely, and each comes from a company that sells computing, models or advice. This site's own projection, an internal model built in October, has a base case inside their spread. Split by the three layers, it has the firms keeping up, not the Joneses, paying most of the token bill by the end of the decade, because many workers adopting agents adds more volume than a few going deeper. The Joneses' share falls even though each of their workers still burns several times what a worker at a firm catching up does, so their edge moves from how much they spend to how little they need per accepted result. Checking agents' work becomes a real share of the bill as well. The layer shares are the model's assumptions. They aren't shown as figures here, because the model isn't published yet and nobody has tested it against a company's accounts.
about 3 hours
At 80% reliability, frontier AI agents completed software tasks of about 12 minutes in early 2025 (Claude 3.7 Sonnet) and about 3 hours by April 2026 (Claude Mythos Preview); at 50% reliability the latest models exceed 16 hours, beyond what METR's tasks can measure reliably.
Software tasks in METR's own suite, timed by how long they take skilled people. METR re-fits its numbers as it adds models, and its longest estimates are the least reliable.
77%
77% of business API transcripts on Claude show automation patterns, against 12% augmentation; directive Claude.ai chats rose from 27% to 39%.
Claude classified the transcripts itself, and the API is where businesses build automation, so this isn't a share of all AI use.
Disclosure: Anthropic's own data about its own product; Claude, Anthropic's AI, helps research this site.
A lead model with subagents beat a single Claude Opus 4 by 90.2% on Anthropic's internal research eval; Anthropic says agents use about 4x the tokens of chat and multi-agent systems about 15x.
General statements in one maker's engineering post, not a measured average. Other sources give other multiples, such as Gartner's range on the next card.
Disclosure: Anthropic's own figures about its own product; Claude, Anthropic's AI, helps research this site.
5x to 30x
Gartner predicts providers' cost of inference on a 1-trillion-parameter model will fall over 90% from 2025 to 2030, while agentic tasks need 5 to 30 times a chatbot's tokens, so total inference spend still rises.
A forecast. The cost that falls is what providers pay to run a model, not the price buyers pay.
24x
Goldman Sachs expects token consumption to grow 24x between 2026 and 2030, to about 120 quadrillion tokens a month, implying about 5 quadrillion a month in 2026.
A bank's forecast of token use worldwide, driven by consumer and business agents.
87x
Dell projected token consumption to grow 3,400% (35x) by 2030 (May 2026), then said inference-driven tokens will grow 87x, to 3,600 quadrillion, by 2030 (September 2026). No base year is given.
A seller of AI servers forecasting demand for them, with no base year.
Disclosure: Dell sells the servers that run AI.
See also: How much will companies spend checking AI agents' work?
These questions decide whether the token bill is a tax or an investment, and none has a measured answer yet. Does the layer a firm is in predict its results, or do the leaders simply report more? How many firms believe they are pulling ahead while only keeping up? Where do the gains go: if they reach customers as lower prices rather than staying as margin, the spend worked as a tax. What share of tokens produce nothing anyone reuses, and does memory a team shares lower the tokens needed per accepted result, as the project expects? Does the tax fund US leadership in AI? Most private AI investment goes to US companies. The split of companies' model spending by provider that the project relies on comes from Menlo Ventures, an investor in Anthropic, which it ranks first, and the data on which providers businesses pay comes from Ramp, which sells software for managing token spend. Against that, much of the hardware is made abroad, open models, some of them Chinese, carry a large share of tokens on at least one developer platform, and nobody traces a dollar of token spend through the labs, clouds and chipmakers it reaches. Claude, an Anthropic model, helped compile this research, one more reason the project treats the question as open. And what happens when agents, not people, direct the work? The project's reading so far is that the best results per dollar come from agents doing the work with people at the decision points, and that the check, not the model, limits how far that can go.
The AI at Work Benchmark
Score your organization's practices, measure your people with an anonymous pulse, and see both against the published evidence.