Theoretical-vs-Observed Capability Gap
One-line
The share of work tasks LLMs could theoretically perform is several times larger than the share they are actually observed performing in professional settings. The gap is the diffusion process; how fast it closes determines when labor effects become visible.
The empirical artefact
Per massenkoff-mccrory-labor-market-impacts-2026 Figure 2:
| Occupational category | Theoretical (Eloundou β) | Observed (Claude usage) |
|---|---|---|
| Computer & Math | ~94% | ~33% |
| Office & Admin | ~90% | ~10–15% |
| Business & Finance | ~75% | ~15–20% |
| Legal | ~70% | ~5% |
| Education & library | ~80% | ~5% |
| Construction / Agriculture / Production / Installation | <10% | ~0% |
Even in the most-exposed category, actual deployment is roughly one-third of capability. In Legal and Education, capability is high but observed work-context use is near zero — usage exists but is mostly classified as personal/leisure or otherwise out of scope.
What drives the gap
The source lists five concrete drivers:
- Model limitations. Theoretically feasible tasks may still fail in practice (hallucination, context limits, weak tool use).
- Legal and compliance constraints. Healthcare (drug-refill authorisations, prescription information) and legal (representing clients in court) are gated even where the cognitive task is in scope.
- Software / integration requirements. Tasks require API surfaces, system access, identity verification that LLM products do not yet expose at scale.
- Human verification steps. Many workflows require a human in the loop even when the model could in principle proceed — see autonomy-slider.
- Workflow lock-in. Existing tools and habits are sticky; substitution requires retraining, re-tooling, and trust.
These are exactly the oecd-sme-enabler-quartet (skills, finance, connectivity, AI-enabling inputs) at the firm level — and at the worker level, ai-skill-shortage-as-diffusion-bottleneck is the dominant binding constraint.
What the gap closing looks like
The source’s framing: “As capabilities advance, adoption spreads, and deployment deepens, the red area will grow to cover the blue.”
Three interpretations of how:
- Linear catch-up. Deployment grows roughly proportionally with capability over 5–10 years; effects scale linearly with the share covered.
- Saturated catch-up. Deployment is governed by the binding constraint that varies by task (legal compliance vs API integration vs trust); some categories close fast, others never close. The eventual deployment shape is uneven across O*NET categories.
- Frontier-shifting. Capability itself moves; the blue area grows faster than the red can catch up. In this regime, the gap never closes — the measure of theoretical-vs-observed becomes a moving target.
The source assumes implicitly the second pattern (different binding constraints in different categories) and offers the methodology as a tool to track which categories are closing fastest.
Why this matters to Where Value Lands
- Operationalises diffusion at the work-task level. diffusion-vs-innovation is largely macro; this is the same idea at occupation granularity. The Ding diffusion-marathon thesis is supported empirically: capability has run ahead of diffusion by ~3× in the most exposed category.
- Frames the labor-market timing question. Labor effects are not currently detectable in unemployment data (see unemployment-did-exposed-workers). The gap explains why: deployment is still narrow. If/when the red area expands to ~70–80% of the blue, the labor effects implied by exposed-worker-demographics become much harder to avoid.
- Parallels scaling-gap at a different level. The 74% enterprise scaling gap is the same phenomenon measured at firm level. Both gaps are the diffusion process — and both are interpretation-flexible: bull frames see them as closing through better enablers; bear frames see them as evidence that AI does not deliver durable value at scale.
What would change the picture
- Sustained, broad-based growth in observed coverage without a parallel improvement in BLS growth projections. That would be evidence the diffusion is happening but the macroeconomic incidence is mild (compatible with task-based-framework’s <0.71% TFP cap).
- A widening gap. Capability advances (frontier-model jumps; agentic capabilities; world models per world-models-jepa) outpace deployment, meaning the measure of theoretical capability is itself unstable.
- An aggressive closing of the gap in specific categories (Computer & Math first; Legal or Healthcare second wave) accompanied by visible labor effects. This is the scenario the source’s framework is best positioned to detect.
Related
- massenkoff-mccrory-labor-market-impacts-2026 — source.
- observed-exposure-measure — the measurement instrument that quantifies the gap.
- eloundou-beta-exposure — the theoretical-capability score used as input.
- diffusion-vs-innovation — the broader frame this gap operationalises.
- scaling-gap — the enterprise-level twin.
- oecd-sme-enabler-quartet — the firm-level binding constraints.
- ai-skill-shortage-as-diffusion-bottleneck — the worker-level binding constraint.
- autonomy-slider — why human-in-the-loop requirements slow deployment.
- task-based-framework — macro envelope the gap-closing operates inside.
- H2_u-curve-of-value — the value-stack consequence: surplus accrues to capital owners of the deployed application layer as the gap closes.
- capital-labor-divergence — the labor-incidence consequence.
- ai-productivity-firm-level — productivity link contingent on the gap closing at firm level.