Value Is the Word. Containment Is the Behavior.

The EY survey published Monday carries a headline about a "shift to value." The number underneath it tells a different story: 98% of leaders using token-based AI tools are reconsidering their strategies because of escalating token expenses.

Companies are showing signs of reckoning with setting priorities rather than merely driving adoption. — Dan Diasio

That is not a shift to value. That is a cost panic wearing a strategic adjective. On the same day, two other stories broke. Corporations are replacing thousands of customer-service staff with AI chatbots — Microsoft cut 10,000 support jobs, Brinks halved its call center ^[1]. And a cybersecurity firm called Reco launched a browser tool that lets security teams block unsanctioned AI prompts at the network edge, sold as data protection but functionally a kill switch for AI spending the enterprise never approved ^[2]. One story says "value." The other two say containment. The three mechanisms differ in how openly they admit what they are. The browser-level blocking is the most disguised. Reco's tool gives security teams real-time policy enforcement over AI prompts in the browser, framed as cybersecurity.

Posture tells a security team what an agent can reach, while runtime provides the control over what it's doing right now, without asking organizations to reroute their traffic through a gateway to get there. — Ofer Klein

But the thing being policed is not data exfiltration alone — it is consumption. Every blocked prompt is a token cost that does not appear on the bill. The security rationale is real, but the financial function is the one that explains why the product launched now, on the same day the EY survey made the cost crisis legible. The staff replacement is the most blunt. Microsoft eliminated 10,000 support positions, saving $750 million annually. Uber cut 10% of its support staff. Brinks halved its call center.

If something happened with little Johnny’s Xbox in the middle of the night, we can now solve that with AI. — Microsoft

This is not augmentation. This is substitution — the fastest way to make an AI deployment show a return is to delete the human cost it was supposed to offset, then count the savings as ROI. EY's Dan Diasio put the new standard plainly.

‘AI saves time’ is no longer a sufficient business case when the costs are mounting and difficult to ascertain over the long run. — Dan Diasio

When "AI saves time" is no longer a sufficient business case, the only remaining case is that AI replaces salary. That is what the numbers describe. Then there is the mechanism enterprises do not control: the providers themselves are throttling consumption. In March, Anthropic quietly reduced Claude session limits during peak hours, with analysts reading the move as an effort to push power users off fixed-price subscriptions and onto metered API billing that guarantees revenue per query ^[3].

To manage growing demand for Claude, we’re adjusting our 5 hour session limits for free/pro/max subscriptions during on-peak hours. — Anthropic

Even Microsoft, launching its enterprise agent platform, emphasized "scale-to-zero economics" as a core feature — the architecture itself now assumes that uncontrolled AI consumption is the primary threat, not a capability gap ^[4].

this refresh is a fundamentally different experience: secure per-session sandboxes with filesystem persistence, integrated identity, and scale-to-zero economics. — Microsoft

The sellers are building cost containment into the product because the unit economics strain under unrestricted volumes, and they know it. The pressure behind all three mechanisms is not speculative. Yann LeCun warned in June that AI labs face a reckoning because running costs exceed what customers pay.

Labs like OpenAI and Anthropic are going to have to increase prices, they're going to have to cut costs, or there's going to be a big bubble explosion. — Yann LeCun

The 95% of organizations reporting no measurable ROI from AI is a buyer-side figure — it says enterprises cannot find the returns they were promised ^[5].

The prices are going up of those AI services, but the cost of running them is going down, but not nearly fast enough. — Yann LeCun

The provider-side evidence — LeCun's warning, Anthropic's throttling, Microsoft's scale-to-zero architecture — suggests the sellers are also straining under the cost of serving AI at current prices. The two signals point in the same direction: the math does not close from either side. And yet the hyperscalers are accelerating in the opposite direction. The top nine cloud providers are projected to spend $830 billion on infrastructure in 2026, a 79% year-on-year increase, with AWS alone at $230 billion ^[6]. The divergence is the shape of the contradiction: infrastructure investment racing upward while enterprise consumption is being throttled, blocked, and redirected into labor substitution. The providers are building capacity for a demand curve their own customers are already bending downward. The strongest counterargument is that none of this is a contradiction at all — it is a J-curve. Stanford economist Erik Brynjolfsson has compared the current AI productivity lag to the PC revolution, where measurable benefits took a decade to appear.

We’re not going to know for some time what the productivity impacts of AI are, in part, because it’s going to be really hard to measure. — Erika McEntarfer

The ECB, meanwhile, found that AI-intensive firms in the euro zone tend to hire rather than fire, suggesting the US pattern of staff replacement may be a local choice rather than an inherent consequence of the technology ^[7].

In other words, AI-intensive firms tend, on average, to hire rather than fire. — European Central Bank

Both points may prove correct over time. But the thesis does not require the J-curve to fail. It requires only that "value pivot" does not describe what enterprises are doing right now. The behavior — blocking prompts, cutting headcount, accepting throttled service — is the record. The language is the packaging. Meta CTO Andrew Bosworth, reversing earlier internal encouragement of AI use, told staff what the "value pivot" actually means in practice.

Nobody should be using AI tools just for the sake of using them. — Andrew Bosworth

The Revenium CEO, launching an AI outcomes-tracking product in March, gave the condition a name: "agent debt" ^[8].

Every AI deployment without outcome tracking is accumulating agent debt. — Greg Rowell

Every deployment without outcome tracking is spending now against a return no one can measure. Whether the J-curve eventually resolves or not, the "value pivot" was the packaging for a unit-economics retreat, and the three mechanisms — blocking, replacing, throttling — are the receipts.