Half of America Uses AI. Fifteen Percent Actually Works With It.

Ramp, Gallup and Teradata measure the gap between bought and used, OpenAI opens the Agents API, Washington buys 23 million ChatGPT seats, and Anthropic counts 200 million distillation exchanges.

I was talking to a friend who does AI consulting last week, and he told me something that's been rattling around in my head since. His clients aren't using AI. I mean, they've bought it. They've got the licenses and the pilot and the slide with the rocket ship on it. But when he sits down with the actual people doing the actual work, most of them touched it once, got a mediocre answer, and went back to how they did things before. And this is a guy whose whole business is getting companies to use the stuff. So I went and looked at the numbers, and it turns out he's not describing his clients. He's describing the country.

Three data sets, the newest of them published Wednesday, put figures on it, and the rest of the week's news reads differently once you've seen them. That's the Big Story. Below it: OpenAI turned its agent plumbing into an API call, the federal government bought ChatGPT for 23 million people, Anthropic put a number on how much of its model has been copied, and the card networks agreed on how a shopping agent proves who it is.

Ramp's September AI Index, published Wednesday under the headline "Cracks in the AI thesis, part 2," counts money rather than opinions: 43.8% of U.S. businesses paid Anthropic for subscriptions or tokens in August and 39.8% paid OpenAI, and both figures barely moved, up 0.34 and 0.09 points month over month. The firms that spend the most spent less. Among the top 1% of AI spenders, median spend per employee fell 9.7%, from $7,976 to $7,205 a month, and the going rate in Ramp's data is $0.68 per million tokens, down 41% from the March peak of $1.15. Ramp economist Ara Kharazian allows that August is when engineers take vacation, then makes the harder point: companies are setting defaults that route work away from frontier models because the cheaper ones are good enough.

That's the buying side. Gallup's second-quarter read of 22,573 employed adults, fielded in May, is the using side: 52% of U.S. workers now use AI in their role at least a few times a year, 30% use it a few times a week or more, and 15% use it daily. Forty-seven percent say their organization has integrated AI tools, up six points in a single quarter, which is the fastest-moving number in the survey and the one with the least to do with what any given employee does on a Tuesday morning. Teradata's "Arrested Automation" survey of 1,000 technology and data leaders at companies with 500 or more employees, run by Wakefield Research in late March and early April, supplies the executive side: 90% plan to spend more on agentic AI in the next 12 months, 37% report measurable business impact, 68% describe their companies as still experimenting or developing, and 40% say more than four in ten of their pilots pause before production.

Put the three together and the shape is the same at every altitude. The access numbers are at or past half. The habit number is 15%. The impact number is 37%, and it comes from the people whose job is to report impact. The press counts the first kind, has your company bought it, have you ever used it, because that's the number that goes up every quarter. The work happens in the second kind. What to watch this fall isn't whether "have used" crosses 60%. It's whether the daily number moves at all, because that's where the return is hiding, and the vendors selling into that gap have started to notice. Thursday's Deep Dive takes the three data sets apart. Monday's job is to say the gap is real, it's measured, and it's yours to close before your competitor does.

Read the full story →

22%

The share of U.S. job cuts announced so far in 2026 that employers attributed to AI, 116,175 of 529,914, per Challenger, Gray & Christmas's August report, which also notes AI slipped to the fourth-most-cited reason in August after leading for five straight months. Fifteen percent of workers use AI every day, and it's the stated reason for one layoff in five. Both numbers can be true at once. One of them is easier to put in a press release.

Source: Challenger, Gray & Christmas

OpenAI opened the Agents API to every developer in public beta on Thursday. It's the managed harness that already runs Codex and ChatGPT for Work: sessions, orchestration, context compaction and recovery happen on OpenAI's side, and you supply the tools and pick where the code runs. Sandboxes come from OpenAI or from partners including Cloudflare, Modal, Vercel and E2B, subagent delegation is built in, and there's no fee beyond tokens and container time. Two lines in the fine print matter for regulated buyers: data residency is U.S.-only for now and there's no zero-data-retention option. For everyone else, the six weeks your team spent building an agent loop just became a paragraph of code, which moves the hard part back to the question this whole issue is about: whether anyone uses the thing once it's built.

Anthropic's September threat-intelligence report, covering December through August, puts a number on distillation: nearly 200 million Claude exchanges across five campaigns run by labs training their own models on Claude's reasoning, TechCrunch reports. The largest, attributed to operators linked to Alibaba, generated 151 million exchanges between May and July, peaking near three million a day across 3,500 accounts, to improve the Qwen models. A Moonshot AI campaign of roughly 300,000 requests through 5,000 accounts focused on surveillance-footage analysis, with traffic Anthropic says came from the Chinese military. The NSA, FBI and CISA had already named six China-based companies in a joint advisory days earlier. If you standardized on Qwen or Kimi open weights because they were cheap and good, provenance just became a due-diligence line, and your counsel is going to ask where the capability came from.

The General Services Administration and OpenAI signed a 27-month OneGov agreement on Thursday that runs through December 31, 2028, and makes ChatGPT, including Astra and the Daybreak cyber platform, available to roughly 23 million federal, state, local and tribal employees. The $15-per-user monthly license fee is waived, usage is discounted 50%, and there's no minimum spend, replacing the $1-per-agency deals that expire September 30. FedScoop's K. Sophie Will notes the competing Gemini and Claude agreements haven't been renewed yet. Twenty-three million seats is the biggest "has access" number of the year, and it's consumption-priced, so the government is about to learn what every enterprise CFO already knows: the bill tracks the habit, not the headcount.

Announced in São Paulo on Thursday, the Know-Your-Agent framework stitches together Visa's Trusted Agent Protocol, Mastercard's Verifiable Intent and Ant International's Agentic Mobile Protocol so an AI agent that registers with one network is recognized by the others. Ant CIO Jiang-Ming Yang put it plainly: register with Ant, and you don't register again with Visa or Mastercard. There's no technical spec, governance body or rollout date yet, and Forkast's Tessa Vaughn is right that the hard part starts now. McKinsey pegs agent-driven consumer commerce at $3 trillion to $5 trillion by 2030. If you sell anything online, agent traffic is going to arrive carrying credentials, and the question for your payments team is whether your checkout accepts them or bounces them.

Measuring AI Adoption among Firms: How You Ask Matters

Alexander Bick, Adam Blandin, David Deming, Nicola Fuchs-Schündeln and Jonas Jessen · Federal Reserve Bank of St. Louis · Essay · 6 min

Why worker surveys said 35% to 40% of Americans used AI while firm surveys said 5% to 7%: the Census asked about AI "in producing goods or services," and when it changed the question to "any business function," the answer nearly doubled. Read it before Thursday's Deep Dive, which starts where this piece stops.

Read it →