They budgeted for seats and got tokens
7 Nov 2026 · News
Enterprise AI budgets for 2026 assumed more seats across more teams. Agents turned those plans into token bills.
Companies planned to nearly double their AI spend this year. On average they nearly quintupled it. Menlo Ventures' new enterprise survey finds buyers budgeted for about 1.9x last year's bill and hit about 4.7x, roughly 2.5 times what they planned, because agents made spend scale with work instead of headcount. The seat model is still how many firms buy AI. The meter is how they pay for it.
The plan said double. The bill said five.
Menlo's 2026 State of Enterprise AI draws on a Morning Consult survey of nearly 400 U.S. enterprise decision-makers, fielded 8 to 26 September, covering more than $4 billion in AI spend those buyers oversee. The average enterprise reported AI spending at 4.7 times its 2025 level against a plan of 1.9 times. That is about 2.5 times what was budgeted.
The median overshot too, but less: 2.2 times actual against 1.8 times planned. Menlo calls the gap between average and median a "fat head." A smaller group of heavy users pulls the mean up. Breadth barely moved. The share of enterprises paying for AI rose only from 55% to 65%. Most of the money came from deeper use among people already buying, not a wave of new logos.
Agents broke the seat math
Menlo's read on the miss is blunt. Agents work on their own, run in parallel, and keep going after people log off, so spending scales with work rather than headcount. That is the same break we wrote about on the consumer side on 4 November, when flat AI plans met agents that never log off. There the open-ended monthly allowance cracked. Here the annual seat-shaped budget cracked.
The labs already know the meter. Anthropic's Enterprise plan charges a seat fee for access only, with no included token allowance, then bills every token in Chat, Claude Code, and Cowork at API rates. OpenAI's token-based Enterprise billing keeps seat fees and usage charges as separate lines, and seats still bill even if a user sits quiet for a month. Google's Gemini Enterprise docs sell licenses with pooled quotas, then optional pay-as-you-go when the pool runs out.
What buyers still lack is a clean public split they can plan against. One line for people. One named line for agent or background work that keeps running. Seat fees plus a lump of "usage" is progress. It is not yet two numbers a finance team can put in a forecast when agents, not chat windows, eat most of the tokens.
Then and now
On 4 November we said the consumer flat rate was built for someone who logs off. This survey is the enterprise twin of that break. On 24 October we noted that about 1% of firms still pay most of the OpenAI and Anthropic bill. Menlo's fat-head average versus a calmer median points the same way: a short list of power users can blow past any seat plan while the long tail is still early.
Yesterday we wrote that a slower data-center buildout is cover for hyperscaler CapEx. That was the supply side. This is the demand side. Software and API spend can race ahead of a seat forecast even while campuses slip.
Our call
By 31 March 2027, at least two of OpenAI, Anthropic, or Google will publish an enterprise pricing page or help doc that lists agent or background work as its own named allowance or line item, separate from the human seat fee and from interactive chat usage. We're wrong if fewer than two do that by the date, or if they keep agent spend buried inside a single undifferentiated usage pool with no named agent line.