Agent Delegability Index: Delegation as a Change of Owner

Sources#

• The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries

Summary#

Espinal Maya (Universidad Nacional de Colombia) & Jiménez Londoño (EAFIT), arXiv 2610.30752 (2026-10-22, 32pp, empirical, OSF-preregistered) ask a different question from the instruments on Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated. Those instruments ask whether a model can speed up a task. This one asks whether a worker can hand the task over. In the authors' words, exposure "measures the intensity of assistance on a task whose owner does not change. Delegation measures a change of owner." The paper's examples run in both directions. A radiology report is highly exposed but stays with the radiologist, because a licensed person has to sign it. Reconciling a bank statement against a ledger was barely exposed to a chat model, but an agent with access to both systems can finish it.

The headline numbers:

• 35.2% of task time in world employment is at least partly delegable (95% CI 33.9–36.7). 2.5% is fully delegable (1.9–3.2).
• The delegation divide runs from 22.5% in low-income countries to 28.3% in lower-middle, 35.0% in upper-middle and 48.7% in high-income countries. The slope is 7.23 pp per log point of GDP per capita (CI 6.52–7.83, R²=0.71, 146 countries).
• Occupation-level ADI correlates at r=0.93 with Eloundou et al. and 0.89 with Felten et al. Both values sit above the preregistered upper bound of 0.80, which the authors had said in advance would mean "ADI adds little beyond exposure." They accept that reading.
• The preregistered adoption gap, the idea that middle-income countries under-use AI relative to their delegable base, is not supported. Across 101 countries, delegability and AEI usage per working-age person correlate at r=0.86, and the middle-income gap has one-sided p=0.270.

The authors' own conclusion is the right way to read the paper: "ADI's contribution is not a new ranking of occupations. It is delegation levels and barriers on top of exposure, for the world."

The ladder and the barriers#

Each task gets one of four levels. A frozen rubric (SHA-256 published, with a script that refuses to run against a modified file) defines them against a fixed reference technology: "publicly available AI agents as of October 2026," meaning tool use, code execution, browsing, computer use and files, but no body.

The rubric records four barriers separately from the level: physical, accountability (law, licence or liability needs a human to sign), trust (in-person care or persuasion) and access (systems or data an agent would not normally reach). The rater is told to judge technical feasibility, "not what employers do and not what the law allows."

Task-level distribution (unweighted). 25.9% of tasks are L0, 24.1% L1, 47.6% L2 and 2.4% L3. Partial delegation is the most common rating. Agents can run large parts of many tasks, but in the models' judgment they finish very few without a human step.

Barriers by major group (§5.3, exploratory). The physical barrier flags 72–94% of task weight in the five manual and service groups (94.1% for elementary occupations). In the four office groups the most common flag is access, at 64–74%. Accountability is highest for professionals (44.8%) and trust for managers (51.6%). Clerical work has the fewest accountability and trust flags (10.8%, 18.0%), and so it is the one group that reaches L3 in volume. Managers rank first at L2+ (81.8%) but have almost no L3 weight: "the models leave the decision with a human."

Occupations. 13 ISCO unit groups reach 100% at L2+, led by coding and proof-reading clerks (4413, 67.7% at L3), contact-centre salespersons and miscellaneous clerical workers. 16 groups sit at 0%: hand launderers, civil-engineering and mining labourers, stonemasons, sewing workers. ADI correlates −0.74 with Webb's robot exposure. In the authors' words, "the work agents can take over is the opposite of the work robots were expected to take."

Full delegation peaks at twelve-minute tasks#

H4 is supported. The probability that a task is L3 follows an inverted U in log human-minutes (coefficients 0.87 and −0.18, both p<0.001), with a peak at ~11.8 minutes. Very short tasks tend to be moments inside physical or interpersonal work. Long tasks bundle in a judgment, coordination or sign-off step, and one such step is enough to hold the whole task at L2.

This bounds the result from a different side than Task Time-Horizon Scaling. METR's horizon measures how long a task an agent can complete on a software basket, and that figure is now hours. ADI says the ceiling on whole-task handover across the economy comes from how tasks are bundled, not from capability length: a longer task is more likely to contain a human-owned step. Two caveats apply. The minutes are the rater model's own estimate, made in the same call as the level, so the curve shows "the models' view of task length rather than a measured duration." And both are October 2026 snapshots of a moving target.

The divide is composition, by construction#

Every country gets the same occupational index, built from the US O*NET task mix. A country's ADI therefore differs only in which occupations employ its workers, plus within-group weights borrowed from US OEWS for high-income countries and Colombia's GEIH for everyone else. The authors are explicit that the divide "does not say that an agent performs the same task worse in a poorer country; it says that the work poorer countries do is work agents cannot reach." The extremes: Singapore 64.2%, Luxembourg 63.5%, UK 54.9%, US 51.9%. At the bottom are Burundi 14.7%, Mozambique 17.9% and Madagascar 19.7%, with India at 27.8%.

That construction matters when ADI is set next to the vault's other cross-country gradients:

• Against the AEI Cadences survey (Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated § cross-sectional gradients): survey respondents in high-income countries report a share of their own tasks AI can do that is about 10pp lower, while ADI's delegable share is 26pp higher. The two do not contradict each other. ADI measures the composition term only, and its design makes the within-occupation term zero. The survey reads within-respondent perception, and it suggests the within-occupation term may run the other way. So the size of the divide is an upper estimate of the composition effect, not a net cross-country gap.
• Against Google ATLAS (Google AI & Economy ATLAS): ATLAS finds the share of conversations that are work-related is highest in Africa. ADI's H2 regression independently finds that low-income countries use Claude more than their delegable base predicts (gap coefficient −0.81, two-sided p<0.001, the opposite of the hypothesis's direction). Two labs' telemetry, read through two different designs, point the same way.
• The physical gate is the same variable as Task Saturation: Broad but Shallow AI Diffusion's extensive margin (ATLAS: occupations with zero AI saturation are the manual ones) and as the end-effector premise of Post-Scarcity Macroeconomics. ADI is the first instrument here that measures that gate employment-weighted for the whole world and versions it (v2026.1), so a later release can show whether it moves.

Transfer risk. O*NET has only 193 task statements for ISCO major group 6 (skilled agriculture), the largest employer in low-income countries. With ESCO skills instead of O*NET tasks the slope falls from 7.23 to 4.66, or 6.92 when only group 6 is swapped. The divide survives but shrinks by about a third, and the authors cite the teleworkability precedent: Dingel–Neiman overstates feasibility in poorer countries (Gottlieb et al.), and ADI's access barrier plays the role that connectivity played there. The divide is also partly the complements argument seen from the capability side. ADI is a ceiling, and the authors call the comparison with usage "a ceiling with a floor."

Convergence: the exposure family again#

The H3 result fits the pattern Steele & Cruz found on Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated: instruments agree when they share a data source or construct family. ADI is LLM-rated against a theoretical-capability question, the same family as Eloundou (GPT-4 rater) and Felten. It lands at 0.93 and 0.89 with them and 0.88 with Eisfeldt. It correlates 0.69 with observed Claude usage by occupation (Handa et al.), and that cross-family figure is the more informative one. It shows near-zero correlation with Webb's AI-patent measure (0.11), in line with Steele & Cruz's finding that older instruments correspond to little. One caution the paper does not raise: these correlations are computed over 433 ISCO unit groups, each an average over a median of 48 tasks, sometimes pooled across several SOC codes. Aggregation raises correlations, so they are not comparable one-for-one with Steele & Cruz's O*NET-occupation-level figures (n≈700–870).

The rater is a measured variable#

ADI's ratings come from Claude Sonnet 5.5 (primary) and Claude Haiku 5.5, called through claude -p (v2.1.306) with no tools, no MCP, no session persistence and a neutral one-line system prompt, one task per call. Three of the method findings belong on the harness pages as much as here:

• The harness frame moves the rating. Re-rating the 1,500-task validation sample under Claude Code's default system prompt ("a coding agent with tools") raises Sonnet's mean level from 1.04 to 1.53 and its L2+ share from 37.6% to 56.8% (κ=0.60 between frames). Haiku barely moves (1.36 → 1.40). Agreement with the human rating drops from κ=0.85 to 0.61 for Sonnet. A model told it is an agent rates more work as delegable to agents. This is Harness-Induced Belief Divergence measured on an elicited judgment about the world rather than on task state: same task, same model, different system prompt, different answer.
• Batching breaks the rating. In the pilot, 25 tasks per call agreed with one-task calls at only κ=0.60 (Sonnet) and 0.56 (Haiku), below the 0.80 floor, so the main pass rates one task per call.
• Model choice matters more than any weighting choice. The world headline is 35.3% with Sonnet alone and 51.4% with Haiku alone. The consensus takes the lower level when the models disagree (32.5% of tasks), so the headline follows Sonnet. Test–retest is κ=0.97 (Sonnet) and 0.88 (Haiku), which makes the gap between them systematic rather than noise.

Validation, and what bounds it. The registered design was three independent humans rating 1,500 tasks. It was replaced before any results were opened (deviation D2) with 300 tasks rated by a group led by a coauthor with student raters. That group was blind to model output, but individual raters were not recorded, so no human–human agreement exists. Human–consensus agreement is α=0.88 and κ=0.85, with 71% exact agreement. The human rating is the more generous one: of the 50 tasks humans put at L3, the consensus puts 42 at L2. The disagreement is about "whether a task an agent can carry out end to end still needs a human to check the result." An agentic panel (Claude Opus 5.5 with web search) agrees at κ=0.87. Raters, panel and adoption data all come from one developer, and the authors name the circularity themselves. It is the same structure as Annotator as Subject: Can Gemini 3 Flash Bias Its Own Capability Profile?, where a model co-annotates the battery it is scored on.

How much to believe#

• Strong on design discipline. The rubric hash is frozen, the registration predates the main pass, every number in the paper traces through a claims.jsonl, the outputs are released, and deviations are dated. A post-hoc exclusion of four countries (D3) moves the headline from 35.20% to 35.17%.
• The preregistration cannot be checked yet. It was registered 2026-10-21 on OSF and is embargoed until 2027-01-31, so the claim that hypotheses, bounds and seeds were fixed in advance rests on the authors' word until then. The main pass of 37,592+ model calls ran between registration and the 2026-10-22 posting.
• Every number is a model's judgment. The authors' own line: "read every number here as a model's judgement, not a measurement of the world." The human check that would bound rater bias is weaker than the one registered.
• Two failed preregistered hypotheses are reported as failures. H2 and H3 both failed and are reported as such, which counts in the paper's favour.

Connections#

• Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated — the instrument family ADI joins and partly collapses into: theoretical-capability raters correlate with each other at ≥0.88 here, as Steele & Cruz's data-source clustering predicts. ADI adds the level scale, barriers and ISCO reach, not a new ranking
• Task Saturation: Broad but Shallow AI Diffusion — ATLAS's physicality gate on the extensive margin, which ADI's physical barrier measures employment-weighted for 146 countries
• Harness-Induced Belief Divergence — the wrapper sensitivity: Claude Code's default system prompt lifts Sonnet's delegability ratings by half a level on the same tasks
• Task Time-Horizon Scaling — METR's capability horizon on one side, ADI's inverted-U on the other: full handover peaks at ~12-minute tasks because longer tasks bundle a human-owned step
• Conversation-to-Delegation Shift — the usage side of the same asking→doing distinction: Codex telemetry measures delegation as it happens, ADI rates where delegation is technically possible
• Post-Scarcity Macroeconomics — the end-effector premise, quantified: the physical barrier flags 72–94% of task weight in the manual groups that dominate low-income employment
• Organizational Complements to AI — why ADI is a ceiling: the authors call the usage comparison "a ceiling with a floor," and the gap between them is where complements, prices and institutions act
• Google AI & Economy ATLAS — the cross-lab check on the H2 null: ATLAS's work-share inversion and ADI's low-income usage surplus point the same way
• Anthropic Economic Index — the adoption data ADI validates against (country usage per working-age person, r=0.86; occupation usage, r=0.69), and the same developer as its raters
• Annotator as Subject: Can Gemini 3 Flash Bias Its Own Capability Profile? — the same-developer circularity in a different instrument: the rater also helped build what it is checked against

Open Questions#

• Does the delegation divide survive a country-specific task mix? ADI holds the within-occupation term at zero by construction, and the AEI survey suggests that term runs the other way. Settler: ADI re-rated on country-specific task data (STEP/PIAAC skills-use surveys or national O*NET analogues for a low- and a middle-income country), compared on the same occupations.
• Is the r=0.86 delegability–usage correlation partly a shared-developer artifact? Raters and usage are both Anthropic. Settler: re-run the H2 regression with non-Anthropic usage (penetration-adjusted ATLAS country usage) and with a non-Anthropic rater on the 1,500-task validation sample.

Sources#

• The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries — Cristian Espinal Maya & Santiago Jiménez Londoño, The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries, arXiv 2610.30752, 2026-10-22 (title page dated October 23), 32pp, empirical, OSF preregistration 9zntf (embargoed to 2027-01-31). Cited for the abstract, §2 (the ladder, barriers, exposure-vs-delegation argument), §4.2 (rater configuration, batching and wrapper pilot tests), §5.1–5.4 (task distribution, occupations, barriers, H4), §6 (headline, H1, countries), §7 (agreement, human and agentic validation, H3, H2, wrapper sensitivity) and §8 (limitations). Parse notes. Tables 1–3 read clean, and Table 3's top and bottom rows match the §5.2 prose. Table 4 (country regressions) is row-shifted: standard errors and the constant slide down a row in columns 2 and 4. All its coefficients are quoted from the §6.2 and §7.4 prose, never from the table. Figures 3–4 were not opened, because every barrier and major-group figure used here is stated in the prose. Source-internal inconsistencies: the introduction lists an MCP server with "three tools" while Appendix C.5 lists four (adding top_occupations). §8.2 says O*NET has 193 major-group-6 task statements "and 193 of them fall in the human-validation sample," which cannot hold for a 300-task sample stratified at ≥20 per major group, so it is probably a typo for the 1,500-task sample or a smaller count. Neither touches a cited number.