Where AI Value Lands

Task-Based Economic Framework (Hulten's theorem, easy vs hard-to-learn)

4 min read

Task-Based Economic Framework

One-line

A formal labor-economics framework for bounding aggregate AI productivity gains from below by decomposing output into tasks, then asking which tasks are exposed, which are profitably automatable, and how easy each is to learn statistically.

The framework

Aggregate output is the execution of a sequence of tasks, each allocated either to human labor or to automated capital. Under standard assumptions, Hulten’s theorem says that aggregate productivity gains are roughly:

(fraction of tasks impacted) × (average task-level cost saving)

That is — even a technology that is impressive at the task level only moves the macro needle in proportion to how much of the economy it touches and how much cost it removes when it does.

The numbers (per acemoglu-simple-macroeconomics as cited in bear-case-synthesis)

QuantityValue
US labor tasks theoretically exposed to gen AI~19.9%
Of those, share profitably automatable in 10 years~23%
→ All US tasks profitably automatable~4.6%
Average task-level cost saving on those tasks~15.4%
Resulting 10-year US TFP gain (upper bound)0.55–0.71% cumulative
10-year US GDP gain (upper bound)0.9–1.6% cumulative

For comparison, optimistic forecasts (McKinsey, GS Economists) put 10-year US productivity gains at ~9% cumulative — >10× higher.

Easy-to-learn vs. hard-to-learn tasks

A critical refinement. Tasks differ in how learnable they are by current AI architectures:

  • Easy-to-learn: objective performance metric, clear input → output mapping, abundant training data. Examples: code autocomplete, document summarization, ad copy drafting.
  • Hard-to-learn: context-dependent judgment, multi-variate decisions, physical-world intuition, reliance on tacit/historical experience. Examples: novel project planning, cross-functional negotiation, complex medical diagnosis, M&A judgment.

Empirical productivity gains from gen AI are derived almost exclusively from easy-to-learn tasks. Hard-to-learn tasks are where most enterprise value lives, and they are precisely the tasks the scaling-wall makes hardest to automate. Adjusting for this friction pulls the TFP ceiling down from ~0.71% to <0.55% cumulative over 10 years.

Two corrosive corollaries

  1. GDP can rise while welfare falls. Gen AI lowers the cost of socially-negative tasks (SEO spam, manipulation, deepfakes, exploitative ads). These show up as productive activity in GDP while degrading welfare. See capital-labor-divergence.
  2. Productivity gains for low-skill workers accelerate their displacement. When AI most helps the entry-level worker, it also makes their role most replaceable — and the displaced labor floods the unautomated remainder, driving wages down across the lower distribution.

Why this might be wrong

  • The 19.9% / 23% / 15.4% numbers are model-dependent estimates; sensitivity to assumptions about cost-of-deployment and verification overhead is high.
  • The framework holds aggregate output structure fixed. If AI enables new tasks (genuinely human-complementary), the bound is conservative.
  • Hard vs easy may be the wrong axis. Test-time compute, agent harnesses, and specialized scaffolding can move tasks across the boundary over time.

Data

Graph View

Backlinks