AI Reshaping the Value of Time: A Silent Revolution from Duration to 'Critical Moments'

Classic time-based billing systems, such as hourly wages, labor charges, and the "billable hour," derive their rationale from a fundamental insight in information economics: when output is difficult to observe or attribute, the principal can only pay for observable inputs—time. This paradigm implicitly rests on three core assumptions: execution duration is a scarce constraint, meaning an extra hour yields an extra hour's output; human time is homogeneous, divisible, and interchangeable; and duration serves as a reliable proxy for effort and value, especially when the agent's true effort is unobservable.

In professional services, this paradigm has been most thoroughly institutionalized. However, signs of change are already apparent: the proportion of billable hour revenue for U.S. accounting firms is projected to decrease from 70.0% in 2020 to 67.1% in 2025 [Verified, IPA]. Although this is only about a 3 percentage point drop over five years, it sufficiently illustrates that the strong inertia of the old paradigm is being slowly but genuinely eroded, foreshadowing a profound structural adjustment in the making.

The core theory of the new paradigm can be summarized in one sentence: when a scarce resource becomes abundant, the constraint does not disappear; it migrates to the immediately adjacent next stage [Verified, Balland/Duke CE]. The advent of generative AI and intelligent agents has dramatically driven down the marginal production cost of executive tasks such as coding, copywriting, analysis, design, and customer service, approaching the price of computing power. For instance, a controlled experiment with GitHub Copilot showed approximately a 55.8% increase in task completion speed [Verified, Peng et al. 2023]; a field experiment by BCG involving 758 consultants also indicated that AI users improved their task completion, speed, and quality by 12.2%, 25.1%, and 40% respectively [Verified, Dell'Acqua et al., Organization Science 2025].

However, the collapse of execution costs has not eliminated bottlenecks; instead, it has shifted them to the verification stage. As Catalini, Hui, and Wu point out in "Some Simple Economics of AGI," there is a "measurability gap": the price of generating a seemingly correct answer is collapsing, while the price of judging whether it is correct is not [Verified, Catalini]. Execution can be infinitely parallelized (e.g., deploying hundreds of agents), but verification cannot be equivalently parallelized. Human attention cannot be branched; reading the 100th seemingly reasonable draft still consumes the same limited human attention as reading the first [Verified, Balland]. Consequently, a rift has emerged between raw output capacity and trustworthy output, which is widening as execution costs decrease.

This asymmetry drives the value anchor to shift from "continuous time" to "discrete moments." When execution approaches zero, true value concentrates on a few discrete, irreversible, and high-leverage nodes: defining objectives, setting acceptance criteria, signing off on reputational/legal/safety risks, and using judgment to supersede data in slow-feedback domains. These "critical moments" constitute a minuscule fraction of the total duration but determine the vast majority of outcomes. The value of time thus exhibits convexity: the marginal value of most hours approaches zero, while a few minutes can dictate success or failure. This is the fundamental reason why "duration-based pricing" fails, as it linearizes convex value across time, obscuring the true distribution of value. The unit of billing also shifts accordingly: clients no longer pay for hours invested but for delivered results and assumed responsibility. Catalini calls this "Rent cognition, own trust": models can be rented, but responsibility, chains of evidence, failure libraries, and balance sheets cannot [Verified].

A true "paradigm shift" requires falsifiable criteria, not merely pricing adjustments. This report proposes three core criteria: C1 (billing unit switch, i.e., at least one mainstream category adopts results/responsibility as the default billing unit), C2 (constraint migration, where execution approaches zero and verification becomes the growth constraint), and C3 (revenue-hours decoupling, where the elasticity of unit revenue to hours invested is significantly <1). Current assessments indicate that the revolution is indeed underway but only halfway complete: C1 is clearly occurring in categories like software and customer service, C2 is evident in fast-feedback domains, but C3 still lacks systematic evidence. This precisely defines the "layered and asynchronous" nature, revealing the complexity and non-uniformity of the transformation.

Evidence supporting the revolution continues to emerge. For example, Intercom's Fin AI charges based on successful outcomes—solving problems ($0.99), process handoffs ($0.99), etc., and "unsuccessful attempts are not charged" [Verified, Intercom], indicating that "by outcome" can now serve as a billable unit. Salesforce Agentforce also launched conversation-based pricing ($2) in October 2024, restructuring the SaaS cost model [Verified]. In the accounting industry, the proportion of billable hour revenue continues to decline, with alternative pricing models penetrating complex service lines [Verified, IPA]. More intuitively, quantitative data from China's OPC (One-Person Company) scenario shows that every 1 RMB invested in AI costs can replace approximately 72 RMB in development labor expenditure under China's baseline, and up to 91 times in Hangzhou [Verified, Honghuhui "2026 One-Person Company Insight Report"], directly confirming the trend of "execution tending to zero." A Duke CE case also vividly illustrates that "critical moments," such as "red-circling" two key parameters in a high-stakes negotiation, are far more valuable than all the information fully recorded by AI [Verified, Balland].

However, the revolution is not occurring unconditionally; significant counter-examples and boundary conditions exist. For instance, a survey of U.S. law firms revealed that despite 72% offering alternative fee arrangements, 58% believed AI "had not changed billing practices," and law firm revenue growth was primarily driven by approximately a 9% rate increase [Verified]. This suggests that in domains characterized by institutional lock-in, slow feedback, and high responsibility, AI has improved efficiency but failed to alter billing units, with "efficiency dividends" absorbed by rate increases. Furthermore, Klarna rolled back its AI customer service strategy after over-prioritizing costs and compromising service quality [Verified], highlighting that the value of relational moments cannot be overlooked. Catalini warns that when executors and inspectors share training data and incentives, it can lead to "synthetic confidence" contaminating verification, meaning "a reasonable error by one model becomes an approved answer by another model" [Verified]. This implies that the true boundary of the new paradigm is not "routine vs. non-routine," but "measurable vs. immeasurable": any outcome that can be captured by data and optimized through feedback, regardless of prestige, will be exposed to automation [Verified, Catalini]; human advantage is temporary and requires continuous migration.

For businesses, pricing power is shifting from selling "access" to selling "verified/indemnified outcomes." In critical scenarios, clients are not concerned with which model completed the task, but rather who guarantees the final result. Thus, competitive barriers will pivot from model capability to failure libraries, regulatory approvals, chains of evidence, and underwriting capacity, with leading companies potentially aligning more with the valuation logic of insurance companies [Verified, Catalini]. Companies should cease measuring AI success by "output volume" and instead focus on "verified throughput," because output that cannot be confidently used, sold, and indemnified by the company is not productivity, but a hidden liability.

At the organizational level, the most profound impact is not merely layoffs, but the disruption of apprenticeships. The execution layer has long served as a training ground for developing judgment; no one learns auditing simply by reading financial statements but must accumulate experience by manually working through hundreds of models. With AI absorbing execution, if organizations do not deliberately rebuild "supervised high-density decision cycles," they risk discounting future verification capabilities for current profits, akin to "eating the seed corn" [Verified, Catalini]. Simultaneously, each time an expert corrects AI, they risk converting their own judgment into reproducible software, intensifying the "coder's curse." Future organizational structures may evolve into an "AI sandwich" model of "director-agent-underwriter": the top layer defines intentions and boundaries, the middle layer executes at scale, and the bottom layer independently challenges and assumes responsibility. Organizational time management will also shift from "scheduling work hours" to "attention budgeting," reserving in-depth human review for irreversible, reputation-sensitive, or client-facing matters.

For individuals, the strategy is no longer about "doing cheaper things more efficiently," but about building capabilities at both ends of execution. Three roles will become increasingly valuable: the director (who translates vague intentions into actionable constraints and can identify "seemingly successful but actually erroneous" outcomes), the underwriter (who backs results with licenses, reputation, or capital), and the meaning-maker (taste, legitimacy, belonging, narrative, and moral judgment, whose value stems from social consensus rather than functional testing) [Verified, Catalini]. The convexity of time value implies that leverage lies not in "working one more hour," but in increasing the density of judgment during critical moments: making decisions, documenting reasoning, post-mortem reviews, accumulating a "decision history" rather than a "portfolio of works." In a market saturated with synthetic capabilities, a credible record of judgment itself becomes a scarce form of capital.

The One-Person Company (OPC) can be considered the purest experimental ground for this paradigm shift: it compresses "human + AI infrastructure" into a single entity, thereby maximally exposing the constraint of "non-scalable attention." Existing data on China's OPC ecosystem corroborates this trend: non-technical founders now constitute 75%, with operational backgrounds (26%) almost equaling technical backgrounds (25%) [Verified, Honghuhui "2026 One-Person Company Insight Report"]. This implies that when code can be generated, technology is no longer a prerequisite barrier; judgment capabilities such as industry knowledge and user insights become the new moats, which is fully isomorphic with the mechanism of value migrating from execution to judgment.

For OPC research, four core propositions can be put forward. First, the "verification bandwidth bottleneck proposition": an OPC's production capacity ceiling is not determined by computing power or tools, but by the founder's tolerable verification load. Second, the "pricing decoupling proposition": for OPCs pricing by results, the correlation between unit revenue and hours worked should be significantly lower than for peers pricing by hours/subscription; a decrease in hours does not equate to a decrease in revenue. Third, the "judgment as asset proposition": OPCs possessing vertical domain knowledge, failure libraries, and decision records should have higher survival rates and pricing power than OPCs relying solely on tool proficiency. Finally, the "trust structure proposition": the greatest vulnerability of a single entity lies in "single-point verification" and "lack of underwriting capacity," thus OPCs require external trust infrastructure (such as insurance, guarantees, peer review, community reputation) to convert individual judgment into tradable, trustworthy results. Of course, existing OPC data largely comes from community questionnaires and media profiles; the 72x/91x multiplier in HACR from the Honghuhui report, sample representativeness, and the effectiveness of "special policies in over 20 cities" [Verified, Hangzhou Daily] all urgently require independent verification.