Enterprise software vendors are currently locked in a race to the bottom on model costs. Notion, for instance, recently touted its gen-6 agent harness, claiming a 90% reduction in average costs compared to previous models. These vendor-reported benchmarks are compelling, promising that credits will go ten times further as agents move toward 24-hour durable execution. Yet, this focus on compute efficiency obscures a fundamental shift in the enterprise operating model: the emergence of the “reviewer-at-zero” phenomenon.
While vendors market the cost savings of their model routers, they are effectively offloading the agent’s margin of error onto internal human review queues. This labor is entirely unpriced in the vendor’s subscription model, creating a hidden, unmeasured tax on the enterprise. We see this architecture codified in new product features. Notion’s “Review for me” safety mode, which uses a separate model to check whether an action is safe before the agent takes it, and Microsoft’s Guardian Agents, which require provable human approval before consequential actions proceed, are not just security features; they are admissions that the underlying models cannot be trusted to operate autonomously at scale. As Satya Nadella noted, the most trustworthy system is one designed to trust the model the least.
The data suggests this “review tax” is already eroding the productivity gains promised by agentic workflows. A Harvard working paper found that while AI coding agents increased code volume by 30%, they also triggered a 49% increase in the average time to review a pull request and raised the share of workers performing code review by 14%. Similarly, research from the BCG Henderson Institute indicates that 47% of employees now spend more time managing AI than performing their core tasks. These figures suggest that enterprises are trading predictable software licensing costs for unpredictable, unpriced human labor costs.
This reliance on human oversight is becoming a market and regulatory mandate. A 2026 report from ChannelEngine found that 76% of consumers demand a human review step in automated processes, with only 17% expressing comfort in fully autonomous purchasing. Furthermore, the regulatory landscape is hardening. Colorado’s new ADMT act gives consumers the right to request meaningful human review of an adverse automated decision, and the attorney general’s proposed rules under the act would require that review to be completed, with a response to the consumer, within 45 days of the request, before an independent reviewer with actual authority to approve, modify, or override the decision. Those rules are still in draft, written to take effect January 1, 2027. In this environment, the human is no longer an optional safety net; they are a load-bearing component of the agent harness.
The implications for enterprise buyers are significant. When you deploy an agent, you are not just buying software; you are assuming the full liability for its margin of error. Because the human reviewer remains the ultimate point of accountability, the enterprise is paying for the agent while simultaneously staffing the oversight layer required to keep it from failing. This creates a false sense of profitability. If the bottleneck for agent deployment shifts from model latency to human review capacity, the scalability of these systems will be capped by the availability of skilled human labor, not by compute power.
Ultimately, the “reviewer-at-zero” model forces a choice between speed and safety that most enterprises are not yet accounting for in their budgets. Whether you are navigating the agent governance choice or managing the supervisor agent shift, the cost of human oversight must be factored into the total cost of ownership. Until vendors include the cost of human-in-the-loop in their value propositions, enterprises will continue to pay a hidden premium for the privilege of supervising their own software.
