Enterprise AI agents at work

Menlo Ventures’ fourth annual State of Enterprise AI report, released November 3, contains one number that should reorder how everyone in this industry thinks about their customer: agents now generate 80% of all tokens on OpenRouter, roughly seven times human-driven usage. Agent traffic has grown 37-fold since it overtook human traffic in February.

The headline figure — enterprise AI deployment spending of $176.8 billion by September 2026, up 5.4 times from $33 billion a year earlier — is impressive. But the spending pattern inside it matters more than the total, and it tells a stranger story than “AI is booming.”

The money didn’t spread. It concentrated.

Adoption rose only modestly, from 55% to 65% of companies. The growth came from existing adopters spending far past plan: average spending reached 4.7 times last year’s level against a planned 1.9 times, and surveyed companies spent 2.5 times what they had budgeted for the year. This is not a broad diffusion curve. It is a smaller set of power users going deep — and what they are buying is agent work that continues after people step away.

The vendor concentration is just as stark. Anthropic, OpenAI, and Google together take 69% of deployment dollars, up from 50%. Inside enterprise LLM API spending, Anthropic rose from 40% to 55%, OpenAI from 27% to 31%, and Google fell from 21% to 9%. Coding is the largest application category at $34.0 billion — the first place agents proved they could do useful work unsupervised, and still the place where the money pools.

Two caveats belong in any honest reading. The report draws on a Morning Consult survey of nearly 400 US enterprise decision-makers overseeing more than $4 billion in AI spending, fielded September 8–26, plus a bottom-up market model — so the $176.8 billion is a modeled estimate, not an audited total. And Menlo discloses that it invests in Anthropic and OpenRouter, two of the report’s biggest winners. Neither caveat kills the story; both should temper how precisely anyone quotes it.

The customer changed before the sales pitch did

Here is the part most vendors are still missing. If agents consume 80% of tokens, then the “user” of an AI product is increasingly a machine doing a task: reading sources, comparing options, and citing what it can verify. A page written only for a human skimming a results list misses the customer who now consumes most of the tokens.

That flips several comfortable assumptions at once:

  • Discovery is now machine-readable or invisible. When an agent does the shopping, the winning source is the one whose facts are exact, structured, and retrievable — not the one with the prettiest landing page. OpenAI’s own claim this week, that ChatGPT’s search-like queries run at roughly one fifth of Google’s volume, points the same way: a second search market at scale, with its own citation habits and far fewer citation slots per answer.
  • Budgets follow workloads, not headcount. Companies blew past AI budgets 2.5-fold not because they hired more AI users, but because agents run while humans sleep. Spending scales with autonomous work, which means the old per-seat pricing logic is quietly breaking.
  • Control becomes the product. The same week Menlo’s numbers landed, CrowdStrike shipped an AI Kill Switch — a human hand on a switch outside the agent’s reach — and acquired two agent-monitoring firms for a reported ~$495 million combined. The market is already pricing the governance layer for a world where the actor is not a person.

What to watch

Three questions decide whether this is a durable shift or an overshoot. First, does the spending concentration unwind — do the 35% of companies still not paying for AI join, or does the gap between power users and everyone else widen? Second, does agent token share keep climbing past 80%, or does it plateau as agents get more efficient per task? Third, do the controls arrive before the first expensive incident, or after it?

The deepest irony of the Menlo report is that it was written for human executives, in a market where the reader that matters most may soon be an agent summarizing it for them. The companies that understand that — that build for the machine customer, price for autonomous workloads, and ship the off-switch alongside the agent — are the ones this $176.8 billion is actually flowing toward.

Sources: Menlo Ventures, 2026 State of Enterprise AI (November 3, 2026), via its published summary; survey of nearly 400 US enterprise decision-makers (Morning Consult, September 8–26, 2026) plus Menlo’s market model; Menlo discloses investments in Anthropic and OpenRouter. Dollar totals are modeled estimates, not audited figures.