The price tag you can see
xAI published a seat price for GrokBot this week, on its own official pricing page: Premium at $30, Super at $100, and Ultra at $200 per seat each month, with GrokBot included in all three tiers (vendor-reported, read on the page October 15, 2026), as our X Subscription pricing piece documented. The page even carries its own comparison line: the combined stack was “previously $90/month combined.” For procurement teams and IT operators, this is the first vendor price tag in this lane to land on a first-party page. It is also the smallest line in the bill.
The rest of the financial map stays hidden. A seat license is one line. Implementation, licensing at scale, operations overhead, and the long-term lifecycle of the system are the rest, and vendors almost never publish them. That leaves the people approving budgets to build the case on incomplete data.
The multiplier trap
The pricing page also shows how usage is metered, and this is the part worth staring at. The shared weekly pool is expressed as multipliers: 4x, 15x, and 50x by tier, against a base the page does not print (vendor-reported, as printed on the page). Without the base, the multiplier cannot become a per-action price. A procurement manager is being asked to approve a variable expense without knowing the starting point.
The wider market is no better. As our four money models map documented, vendor price tags were largely absent until now. Atlassian keeps deepening its OpenAI partnership with a significant spend commitment, and ServiceNow has confirmed it is moving toward Now Assist monetization via price uplifts (vendor-reported). The dollar figures for what those deployments actually cost stay opaque.
Building a case on seller-made numbers
What passes for cost evidence today is mostly seller-made. The only granular dollar benchmarks in circulation come from vendor-published TCO guides, which carry the commercial bias of whoever wrote them. Those guides put an enterprise deployment of roughly 250 to 500 users and 20 to 50 agents at $500,000 to $2,000,000 in annual licensing, plus $200,000 to $1,000,000 in one-time implementation (vendor-published figures). Operations add another 20 to 30 percent on top of licensing, and up to two-thirds of total system cost lands after deployment, in maintenance, integration, and support (also vendor-published).
The adoption numbers get quoted far more than the cost numbers, and they measure something different. Deloitte’s State of AI in the Enterprise 2026 survey, fielded in late 2025, finds 23 percent of companies using agentic AI to at least a moderate extent today, and 74 percent expecting to within two years. Those are intent and deployment-status figures. None of them say what a running agent costs. On the cost side, an IDC InfoBrief commissioned by DataRobot, published December 9, 2025, reports that 92 percent of organizations implementing agentic AI saw costs higher or much higher than expected, and 96 percent for generative AI deployments. The research was commissioned by a vendor in this market, which is itself worth naming.
What the forecast is actually forecasting
The number that circulates most in budget conversations is a prediction. In a Gartner press release from June 2025, the firm predicted that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. It is a forward-looking estimate, not observed data. Anushree Verma, a senior director analyst at Gartner, put the mechanism plainly: hype-driven projects “can blind organizations to the real cost and complexity of deploying AI agents at scale.”
When the procurement manager and the IT operator sit down to approve the budget, this is the material they have: a seat price, a vendor-published cost guide, a sponsor-commissioned survey, and a cancellation forecast. Recent named deployments at Kyndryl with Toyota Finance, HCLTech, and Workday published almost no dollar figures at all. DXC shared outcome percentages, but not what they cost to reach.
What would settle it is independent cost data from real deployments over time: what the budget actually did six, twelve, and eighteen months after agents went live. That evidence does not exist yet. Until it does, the approval decision runs on the most optimistic version of the math, and the bill arrives somewhere the dashboard never shows.