No Last Mile: A Theory of the Human Data Market

Abstract: The standard framing treats structured human-data work as transitional, a bridge between today's imperfect models and a future state where automation is complete. We challenge this view by modeling structured human data as a persistent production input: evaluation, rubric-based judgment, auditing, exception handling, and continual updates that convert raw model capability into dependable, deployable performance. These activities accumulate into a reusable AI capability stock that raises productivity by improving reliability on existing tasks and by expanding the frontier of task families for which AI can be used at high confidence. Crucially, this capability stock depreciates as tasks and contexts drift, standards evolve, and new edge cases emerge. In a tractable baseline model, an interior steady state implies a closed-form, strictly positive long-run labor share devoted to structured human-data work whenever depreciation is positive, a "no last mile" result in which maintenance demand persists even as models improve. We then microfound aggregate capability with a portfolio of task families featuring diminishing returns, frontier entry, and complementarity, generating reallocation toward low-maturity and bottleneck families and a Roy-style mechanism for within-structured wage dispersion. Finally, we map model objects to observable proxies using standard data layers, and provide a conservative calibration suggesting a 5-7% steady-state structured labor share in the long run.

• The paper develops a macroeconomic model showing that structured human-data work—evaluation, auditing, exception handling, and model updates—remains permanently necessary because AI capability depreciates through technological, environmental, and organizational drift.
• The model derives a positive steady-state structured labor share that rises with capability depreciation and elasticity, with conservative calibration estimating roughly 5.85% of labor, or a 3.13%–8.53% 10th–90th percentile range.
• The paper predicts that new task families and capability bottlenecks redirect workers toward low-maturity areas, producing persistent maintenance demand and greater wage dispersion that can be tested using job postings, labor-market data, adoption surveys, and firm telemetry.

Overview and motivation

This paper develops a macroeconomic theory of structured human-data work—evaluation, rubric-based judgment, auditing, exception handling, and continual model updates—and argues that this work is a permanent production input rather than a transitional phase preceding full automation. The central claim is that structured human data accumulates a reusable "AI capability stock" that both improves reliability on existing tasks and expands the set of task families where AI can be deployed with high confidence. Because this stock depreciates through technological, environmental, and organizational drift, equilibrium labor demand for structured human-data work never vanishes: the paper's titular "no last mile" result.

The motivation is an empirical tension in the recent literature. Adoption surveys show rapid uptake of generative AI (nearly 40% of U.S. adults aged 18–64 using it by late 2024), and firm-level deployments show substantial productivity gains (roughly 15% more issues resolved per hour in a large customer-support setting) (2603.00932). Yet representative Danish evidence finds precisely estimated near-zero effects on earnings and hours over two years, ruling out changes larger than about 2%. The paper reconciles these facts by locating the near-term adjustment margin in task reorganization—quality management, exception handling, workflow updating—rather than headline displacement.

Baseline model and the closed-form steady state

The baseline economy allocates a fixed labor endowment $\bar L$ between unstructured production ($L_U$) and a structured human-data sector ($L_S$). Output is Cobb–Douglas in capital and unstructured labor, scaled by $A(k)=\bar A k^\gamma$, where $k$ is the capability stock accumulated by structured labor according to $k_{t+1}=(1-\delta_k)k_t+\eta L_{S,t}$. Competitive labor mobility equalizes sectoral wages, with the structured wage equal to $\eta v$, where $v$ is the shadow value of capability.

The main theoretical result is existence and uniqueness of an interior steady state with a closed-form structured labor share:

$$s^* = \frac{\gamma\,\delta_k}{\gamma\,\delta_k + (1-\alpha)(r+\delta_k)}.$$

Two comparative statics carry the paper's falsifiable content. First, $s^*$ is strictly positive whenever $L_U$0 and increasing in $L_U$1: faster drift implies a larger permanent maintenance share. Second, $L_U$2 is increasing in the capability elasticity $L_U$3 and decreasing in the discount rate $L_U$4. The implication is direct: as long as deployed AI systems lose alignment with a changing world, some strictly positive fraction of labor must be devoted to maintaining that alignment, regardless of how capable foundation models become.

The authors also offer a nonrivalry reading of the mechanism. Because codified artifacts (rubrics, evaluation harnesses, audit protocols) are reusable across downstream interactions, the capability stock resembles partially nonrival knowledge in the Romer sense. Since private platforms need not internalize all spillovers across firms and task families, the competitive allocation plausibly underinvests relative to the social optimum; the baseline $L_U$5 is therefore best read as a conservative lower bound on the long-run maintenance share.

A notable interpretive move is the decomposition of depreciation into three channels—technological (model/tooling updates breaking rubrics), environmental (distribution shift creating new edge cases), and organizational (changing standards and policies)—with a proposed event-study design estimating $L_U$6 as an event-driven hazard rate of maturity loss.

Task families, reallocation, and wage dispersion

The extended model microfounds aggregate capability with a portfolio of task families, each with its own maturity stock evolving under concave codification technology $L_U$7 with family-specific drift $L_U$8. Aggregation is additive or CES, with $L_U$9 capturing complementarity and bottlenecks. New families arrive at state-dependent intensity $L_S$0, which rises around major model releases, in digitally intensive sectors, and after governance shocks.

Five propositions structure the reallocation results:

• Marginal equalization: interior allocations equate weighted marginal codification returns across families, shifting effort toward high-weight and low-maturity families.
• Frontier reallocation: increases in entry intensity $L_S$1 pull structured labor toward newly arrived, low-maturity families.
• Saturation: diminishing returns mean mature families require less incremental labor per unit of capability improvement.
• Maintenance demand: even with zero entry, positive drift requires a strictly positive long-run flow of structured labor; with entry, the requirement is bounded away from zero by both maintenance and frontier investment.
• Bottlenecks: under CES complementarity, low-maturity families acquire high shadow value and attract persistent investment even as others saturate.

The paper then introduces a Roy-style assignment model in which workers with family-specific comparative advantage $L_S$2 sort into families maximizing $L_S$3, where family shadow prices $L_S$4. High-judgment families combine high shadow value with thin supply of qualified workers, generating wage premia, while standardized families price near the general outside option. A further prediction is that dispersion widens following frontier entry or drift shocks, steepening the cross-family distribution of shadow values. This delivers a right-skewed within-structured wage distribution driven by frontier and bottleneck families—a testable claim distinct from the view of structured workers as interchangeable "training replacements."

Measurement and calibration

The measurement section maps model objects to five standard data layers: O*NET-based occupation–task exposure maps, job postings and career-graph data, online labor market panels, business adoption surveys, and within-firm deployment telemetry. Structured labor share is proxied by postings-based taxonomies of human-data roles; task-family entry by cluster births classified by trigger type; maturity by rubric stability, inter-rater agreement, rework intensity, and failure rates on fixed audit sets; and drift by degradation-threshold crossing hazards decomposed over event windows.

The calibration is deliberately conservative. With $L_S$5, $L_S$6, $L_S$7, and a low capability elasticity $L_S$8, a Monte Carlo exercise over 200,000 draws yields a steady-state structured labor share centered near 6% (mean and median 5.85%, SD 1.98 percentage points), with a 10th–90th percentile interval of [3.13%, 8.53%]. Under these priors, $L_S$9 and $A(k)=\bar A k^\gamma$0. Both human-data parameters are chosen to bias the estimate downward; extending the upper bound on $A(k)=\bar A k^\gamma$1 to 0.12 shifts the mean toward roughly 8%. The calibration therefore supports a mid-single-digit long-run share as a lower-bound envelope—one consistent with muted short-run aggregate labor-market effects coexisting with an active, persistent reorganization margin.

The empirical roadmap proposes three identification designs: event studies around model releases, policy changes, and content-regime shifts; cross-environment comparisons linking structured shares to sectoral digitization; and sorting tests relating rate premia to bottleneck and thinness-of-supply proxies.

Policy implications

The no last mile result carries two governance claims. For workforce policy, treating structured human-data roles as transitional risks misallocating training investment and undervaluing workers performing a structurally permanent function; the framework argues for treating them as a distinct occupational category with career pathways and credentialing. For regulation, current frameworks concentrate accountability at model deployment, whereas the model locates alignment in ongoing maintenance labor. This motivates transparency requirements around maintenance activity, standards for evaluation design, and accountability when capability stocks depreciate through underinvestment—relevant to multi-year procurement horizons for governments and institutions.

Limitations and open questions

The authors concede several constraints explicitly. The model compresses heterogeneous deployment settings into few parameters and treats the capability-to-productivity mapping in reduced form; it abstracts from platform market power, contracting frictions, and general-equilibrium feedbacks that could amplify or dampen structured labor demand. On measurement, the key latent objects—capability maturity, drift rates, task-family entry, shadow values—require combining data layers that are not jointly observable, and cleanest validation depends on within-firm telemetry that is difficult to access at scale. The calibration's priors on $A(k)=\bar A k^\gamma$2 are asserted rather than estimated, so the 5–7% figure inherits their conservatism. Open questions include whether the three drift channels can be empirically separated with the proposed hazard design, and whether the Roy-sorting predictions survive once platform monopsony power is introduced.

Conclusion

The paper formalizes structured human-data work as a persistent input whose accumulation and depreciation jointly determine a strictly positive long-run labor share, delivering a closed-form steady-state share calibrated conservatively at roughly 5–7%. Its portfolio extension adds reallocation toward low-maturity families, complementarity-driven bottlenecks, and Roy-style wage dispersion within structured work. The contribution is primarily theoretical, but it specifies measurable proxies and identification designs that make its propositions falsifiable with existing data layers.