The Cost AI's New Metric Leaves Out

When OpenAI launched its enterprise agent product this spring, the pitch was autonomy: AI that could execute workflows across an organization without a human driving every step. The product shipped with a Compliance API and mandatory human-approval requirements for sensitive tasks ^[1].

Teams can now create shared agents that handle complex tasks and long-running workflows, all while operating within the permissions and controls set by their organization. — OpenAI

A task that cannot proceed until a person signs off was not completed by the AI. The human's time — the review, the judgment, the sign-off — is not in the task price. That single requirement is the whole contradiction in miniature. The contradiction runs through the industry's two big narrative shifts of 2026. The first is the move from "cost per token" to "cost per task." OpenAI CFO Sarah Friar has been the most visible advocate, arguing that cheaper models can actually cost more because they require "more attempts, more time, or more human review" to produce usable output ^[2]. The logic is that a task — a completed, useful unit of work — is what customers should pay for, not the raw compute that went into it.

A lower-cost model may have cheaper tokens, but getting great results may require more attempts, more time, or more human review. — Sarah Friar

The second shift is the rebranding of autonomous agents as "human-led AI." Across the industry, the language has softened from agents that act to agents that advise. DHL, SAP, and BMC Helix have all deployed agentic AI in the past year, and in every case the framing is the same: the AI "suggests" alternate routes, "proposes" cost reductions, operates as a "workforce multiplier" — never as a replacement ^[3]. Executive teams require "clear guardrails and consistent human oversight" for these deployments.

AI is a workforce multiplier: Now’s your chance for pie-in-the-sky projects that were too expensive or unrealistic before. — Ryan Manning

The two pivots cannot both be true. A task metric implies the task is done — that the unit of value is the completed work, not the machine's attempt at it. But the governance frameworks now emerging across the industry concede that agents cannot complete tasks reliably without human oversight whose labor cost the task metric does not count. The evidence for this concession is not hidden. It is in the product specs. A new corporate AI adoption framework called "Humalogical Balance" explicitly positions human judgment as the lead and AI as the servant, requiring employees to vet AI output and challenge errors ^[4].

AI is different: It does not just ask people to learn a tool; it asks them to rethink how they work, and most companies still treat it like a software rollout. — Annette White-Klososky

KnowBe4's security strategist warns that excessive agent autonomy triggers "error propagation, false positives, and automated behaviors that conflict with broader business goals," and prescribes "mandatory human review for all high-impact decisions" ^[5]. GoodData's field CTO, a builder inside the agent ecosystem, concedes that while agents technically work, "once they're live, there's not enough control over configuration and governance" ^[1].

the agent works, but getting it into production takes too long, and once they're live, there's not enough control over configuration and governance — Peter Fedorocko

Even GitHub's Agent HQ, which integrates Claude and Codex agents directly into developer workflows, includes the caveat that "agents can still make mistakes," and the agents are deployed for code commits and pull request comments that remain subject to human review ^[6]. The pattern is uniform: every deployment that works concedes the same thing the task metric denies. The human supervisor's labor is not merely unpriced. It is unmeasurable under either metric. A Coupang AI director has identified what he calls the "Request-to-Silicon Gap" — the inability to trace a single business request through agentic loops down to actual compute costs. A single agentic request triggers numerous downstream operations with unpredictable cost variances ^[7].

If you can't trace a request to the silicon that served it, you aren't operating AI infrastructure. You're guessing. — Srikanta Datta

The Linux Foundation launched a Tokenomics Foundation this month — with JPMorgan, Accenture, IBM, Oracle, SAP, and ServiceNow as founding members — specifically because companies cannot link token usage to ROI. Its executive director puts the problem plainly: "a flat limit tells you nothing about whether the spend earned anything" ^[8].

A CFO should collaborate with their teams to build visibility into consumption by model, workload, team, and project. — J.R. Storment

Revenium built an entire product, AI Outcomes, to solve the measurement gap ^[9].

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

The human supervisor's labor sits outside the pricing model because it sits outside the machine — a cost borne by the organization's payroll, not the API bill, and therefore invisible to both the token metric and the task metric that replaced it. Wayne Liu of Perfect Corp calls the result an "AI hangover." The real bottleneck, he argues, is not the cost of compute but the human judgment required to validate machine output. Generative AI has shifted the professional burden from creation to supervision, and the supervisor's wage — the reviewer's time, the approver's attention — is the one cost neither metric captures ^[10].

Cheap intelligence creates value only when it improves judgment. — Liu Ruilin

We are entering the AI hangover, when we start to realize its true cost. — Liu Ruilin

The task metric captures the model's cost. It has no column for the supervisor's. And the supervisor, every enterprise deployment now confirms, is not optional.