The $6 Trillion AI Bill: Why 'Always-On' Agents Will Break Your Budget

Are AI agent subscriptions worth the cost? For most companies right now, no. Not because the technology doesn't work. The pricing model is changing under your feet, the ROI is still unproven, and the vendors setting those prices have a trillion-dollar reason to squeeze you.

The $6 trillion gap nobody in procurement is talking about

Start with the number that explains everything else here. Bain & Company's 7th Global Technology Report finds the AI industry needs to generate $6 trillion in annual revenue by 2031 just to justify the data center spending already committed.¹ Existing consumer and enterprise AI services might get there for $1.8 trillion of that.² The other $4.2 trillion has to come from somewhere, and nobody can point to it yet. Robotics and drug discovery get mentioned a lot. So does agentic AI.

Independent analyst Ed Zitron puts a sharper number on the two companies most exposed: Anthropic and OpenAI together need to make or raise more than $1.25 trillion over the next four years just to cover their compute obligations and cloud commitments.³ Anthropic alone is on the hook for roughly $330 billion in cloud spend to Amazon, Google, and Microsoft.³ These aren't companies with a growth problem. They're companies with a solvency problem, and solvency problems get solved by raising prices on whoever is already paying.

That's the context every enterprise buyer needs before signing another agent contract. The vendor across the table isn't pricing based on what your workflow is worth. They're pricing based on what they owe.

Why is AI agent pricing moving from seats to meters?

The flat, predictable per-seat SaaS fee is disappearing from agentic AI, replaced by usage-based billing that runs the meter on every action an agent takes.

• Salesforce Agentforce has layered consumption pricing on top of existing per-user licensing: $2 per customer conversation, or $500 per 100,000 Flex Credits.⁴ A mid-size company handling 50,000 service interactions a month would spend $100,000 on conversations alone, before touching its Service Cloud seat fees.⁵
• Microsoft's Copilot Studio stacks a per-message charge on top of the $30/user/month M365 Copilot seat. The base plan includes 25,000 messages for $200/month, with overage running about $0.008 per message, and custom agents built on Azure AI Foundry add a further token-based charge on top of that.⁵
• Anthropic has moved its largest enterprise Claude customers off the flat $200/user/month plan entirely, onto per-token billing with mandatory monthly spend commitments, eliminating the 10-15% volume discounts that used to soften the blow.⁶⁷ Licensing experts expect heavy users to see bills double or triple.⁷

This isn't one vendor experimenting with pricing. It's the whole category moving the same direction at once, because the economics forcing the move are the same for everyone.

Why does 'always-on' cost more than it looks?

Vendors market always-on agents as a capability upgrade. Under metered pricing, it's a cost multiplier.

OpenAI's newly unveiled agents, called dots, chase user goals across Slack and Teams continuously, with minimal supervision, competing directly with Meta's autonomous agents.⁸ Weekly users of Codex and ChatGPT Work have already climbed past 35 million, up from about 5 million Codex users in June.⁸ Microsoft has been pushing its own always-on agent, Scout, in the same direction.

The pitch is that an agent working continuously in the background gets more done than one you have to prompt. The billing reality is that continuous operation means many more micro-actions per task, and every one of those actions is a metered event under the new pricing structures. A task that used to be one flat-fee session now generates dozens or hundreds of billable calls. Vendors call this capability. Finance teams should call it usage amplification: the same underlying work now costs meaningfully more, and nobody modeled that shift in advance.

That unpredictability is the real problem. A per-seat fee lets a CFO forecast next year's software line to the dollar. A per-token, per-conversation, per-message fee attached to an agent that runs continuously does not. You find out what it costs after you've spent it.

Can corporate IT budgets absorb the cost of AI agent subscriptions?

Meanwhile, the money to absorb any of this isn't materializing at the pace vendors need it to.

Gartner forecasts global IT spending will cross $6.37 trillion in 2026, up 14.2% year over year, with software spend rising roughly 15.5% on the back of GenAI model development and AI-embedded SaaS like Copilot and Agentforce.⁹ But that top-line growth doesn't match what CIOs are seeing in their own budgets. In Evanta and Gartner's 2026 CIO survey of more than 1,100 CIOs, only 43% reported an increase in operating budget, and 38% reported flat budgets, even as 67% said they plan to invest in AI and machine learning.¹⁰ Most of the organizations being asked to adopt metered agent pricing don't have new money to fund it. They're expected to fund it out of budgets that barely moved.

The anecdotes from companies that already adopted these tools show what happens when the meter runs faster than the planning cycle.

• ServiceNow. Its CIO met with the company's CFO to figure out how to contain Claude Enterprise costs for the rest of the year, only months into the budget cycle.³
• Stripe. Its engineering staff were reported to be burning roughly $2.8 million a month in AI tokens.³
• Zillow. Its AI spending is on pace to hit $7 to 10 million in 2026, more than 20% of its 2025 net income, while internal data showed a 50% increase in engineer time spent reviewing AI-generated output.³

That last point matters as much as the dollar figure. Zillow didn't just spend more on AI. It spent more on AI and got more human review work in return. If you're modeling agent spend against a headcount-reduction story, that's the number that should worry you most.

Do AI agents actually deliver ROI?

Here's what should end the conversation for most buyers: the value case for agentic AI is still mostly unproven, even before you factor in what the new pricing does to the bill.

MIT's GenAI Divide report, based on 150 leadership interviews, a 350-employee survey, and analysis of 300 public AI deployments, found that about 95% of generative AI pilots fail to deliver measurable P&L impact.¹¹ The same research found internally built AI tools succeed roughly a third as often as tools bought from vendors.¹¹ Gartner goes further on agentic AI specifically, predicting that more than 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls, and estimating that only about 130 of the thousands of self-described agentic AI vendors offer genuine agentic capability.¹² The rest, in Gartner's phrase, are agent-washing.

Put those two findings next to the pricing shift above and the picture gets uncomfortable. Companies are being asked to move onto metered billing for a category where 95% of pilots don't show up in the P&L, and where four in ten projects are expected to get canceled before they get the chance. You would not sign a five-year lease on a car you had a 5% chance of actually driving to work.

Shadow AI: the budget line nobody approved

While procurement negotiates the official agent contract, employees have already moved on without waiting for permission. UpGuard's State of Shadow AI research found that 81% of employees and 88% of security leaders report using unapproved AI tools at work.¹³ That means the real AI spend and risk exposure inside most companies is already bigger and messier than whatever line item IT signed off on. Every one of those unapproved tools is a metered API key running on someone's personal card or a team's no-code automation, quietly accumulating usage the finance team has never seen. If you haven't looked at how a single employee's agent can blow through a usage budget, the shadow version of this problem is usually worse than the sanctioned one, because nobody set a cap.

The Remy verdict: rent less, own more

Here's the honest framework. An agent subscription is worth the cost only when three things are true at once: the task has clear, measurable value, the metered cost of running it is bounded and known in advance, and you have actual governance over what the agent does and how much it spends doing it. Most agentic AI contracts being signed right now fail at least one of those three.

The vendors pushing always-on agents aren't wrong that the technology can do real work. They're under too much financial pressure to have priced it in a way that protects your budget while they figure out how to hit $6 trillion. That pressure is their problem to solve, not yours to fund.

The alternative isn't rejecting AI agents. It's refusing to rent your agent infrastructure from someone whose pricing model exists to close their own revenue gap. Building and owning the workflows that actually run your business on infrastructure you control means you set the ceiling on cost, not a vendor chasing a trillion-dollar target. That's the case for treating agents as a piece of infrastructure you own and operate on your own terms, in a tool like Remy, rather than a metered service you rent and hope stays affordable. Own the agent, own the ceiling. Rent it, and someone else decides where the ceiling is, and they have every incentive to keep raising it.