Zapier's Heaviest AI Users Spend $30,000 a Month. The Company Still Can't Measure the Return.

Zapier CEO Wade Foster did something most executives avoid this week. He told a reporter what his company spends on AI, then said he couldn't tell you what it earns back.

Speaking with Tim Keary for a Forbes piece published August 18 on the gap between token spend and measurable value, Foster laid out Zapier's internal usage in three tiers. Most employees run a few hundred dollars a month. Most developers run into the low thousands. The company's heaviest users, the ones Foster calls its top AI builders, clear as much as $30,000 a month on their own.

Foster contextualized that top number almost immediately. In a LinkedIn post accompanying the interview, he described those developers as "looping coding agents at scale on greenfield projects, fixing bugs and building new functionality." That work is autonomous agents running in loops against real codebases, and the bill scales with the number of loops.

He also went out of his way to call that work some of the company's most intentional and scrutinized, with protections and guardrails in place. The framing reads like someone who has defended the line item in a budget meeting.

Where that puts Zapier

Almost nobody publishes these numbers, which is what makes the disclosure notable. The aggregate data suggests Zapier is sitting comfortably inside a fast-moving pack rather than out ahead of it.

The Ramp AI Index, built on card and bill-pay data from more than 70,000 U.S. businesses, found that the top 1% of firms spent a median of $7,400 per employee on AI in July 2026. In January that same cohort spent $2,590. The top 10% moved from roughly $281 to $650 over the same stretch. The median American company spent $11.95 per employee, which puts more than 600x between the heaviest spenders and everyone else.

Goldman Sachs, cited in the same Forbes piece, expects token consumption to multiply roughly 24-fold to 120 quadrillion tokens a month between 2026 and 2030, even as per-token inference costs drop 60% to 70% a year. Prices per unit are falling. Total consumption is rising faster. The bill goes up anyway.

So the spend curve is well documented. The value curve is where things fall apart.

OpenAI made the case against its own customers

The most damaging data point in this conversation came from OpenAI.

In a 69-page working paper published August 11, OpenAI researchers analyzed more than 17 million ChatGPT Enterprise messages across over 1,500 organizations, comparing usage against financial performance while controlling for company size and industry. The paper's headline argument is a "frontier gap," the idea that firms adopting AI, particularly agents they can delegate work to, are pulling away from firms that aren't.

Then, on page 35, a small table. As Fortune first flagged, the researchers report that revenue per employee is "not meaningfully associated with output tokens per employee" once controls are applied. Neither messages sent nor tokens generated tracked with the financial outcome everyone is implicitly buying.

Foster cited that finding directly, which is an odd thing for a customer to amplify. It also explains why ROI has stopped being a nice-to-have metric. Keary's reporting notes the fallout is already visible in cost overruns at Uber and Amazon, and in the backlash against "tokenmaxxing," the brief 2026 fashion for treating heavy consumption as a proxy for engineering ambition. Keary covered the turn against that mindset in a July piece as well. Some firms, G2 among them, have responded by building efficiency indices that treat AI as a metered utility.

Zapier's own data cuts against the tokenmaxxers

Zapier's product telemetry argues against the spending behavior its own engineers exhibit.

The company's AI Workflow Index started with a panel of 1,500 mid-market and enterprise customers, then narrowed to the top 25%, or 375 companies, ranked by AI workflow adoption. Among those leading adopters, AI accounts for only about 18% of the steps in an AI workflow. It is the largest single category of step and still less than a fifth of the total. Rules, logic, and conventional code carry the rest.

Foster's read on that: you can be a top AI builder without shredding through tokens, especially when you use AI to write the code and build the automations, which then run cheaper than the AI would.

It is a self-serving conclusion for an orchestration company to reach, and it is probably also correct. The distinction is between AI as the labor and AI as the tooling that produces the labor. A model that writes a deterministic workflow charges you once. A model that is the workflow charges you on every run. At volume, those two architectures produce very different invoices. Zapier has been making this argument in its own marketing for months.

Foster stops short of turning it into a rule. You can also be a top AI builder who uses a ton of tokens, he wrote, because more of what Zapier ships now gets iterated on by agents running in loops, and that does not come cheap.

The finding underneath the numbers

The most useful line in Foster's post is the last one. Everyone Keary talked to could say what they spent. Not one of them measured ROI the same way.

That is the condition enterprise AI is in as of August 2026. Spend is trivially observable, because it arrives monthly and itemized on a corporate card. Value is not, and the industry has not agreed on a definition, let alone a methodology. Ticket deflection, engineering hours saved, features shipped, and revenue per head are all in circulation, and they are not measuring the same thing.

Until that settles, a $30,000-a-month developer means whatever your CFO already believed. One reads it as a power user compounding leverage. Another reads it as an unexamined line item. Both are working from the same data, because there isn't any other kind yet.