Future of Work with AI Agents: Auditing Automation and Augmentation Potential across the U.S. Workforce

TL;DR

The agent-building world picks what to automate based on what’s profitable or technically fun (coding, business analysis), not based on what workers want or where the impact is highest. This paper fixes the “want” half of that equation with real data: a survey framework that captures, task by task, how much each worker wants AI to do the task and how much human involvement they think it needs, cross-referenced against expert capability ratings. The headline artifact is WORKBank — 844 tasks, 1,500 workers, 52 experts — plus two reusable conceptual tools: a four-zone desire-capability landscape and the Human Agency Scale (HAS, H1-H5) that replaces “automate yes/no” with a spectrum of human involvement. The punchline result: 41% of Y Combinator startups are building in the “Red Light” (high capability, low desire) and “Low Priority” zones, while high-desire/high-feasibility tasks sit under-served. For anyone selling AI services, this is a free demand map of where the white space actually is.

Problem & Motivation

Everyone agrees AI agents will reshape work. Almost nobody has measured what workers want from them, task by task, before deciding what to build.

The concrete pain the authors call out:

• Coverage is narrow. Prior labor-impact studies fixate on a couple of domains — software engineering, customer support — because that’s where the data and the money are. That tells you nothing about the 100+ other occupations.
• The capital-only viewpoint. Studies emphasize tasks that are profitable to automate (coding) and skip whether the worker doing the task actually wants it automated. Worker values are missing from the optimization.
• Backward-looking data. Methods that mine “how people use ChatGPT” (e.g., Anthropic’s Claude.ai usage logs) only show what early adopters already do. That’s a lagging indicator — it can’t surface latent demand from people who haven’t been given the right tool yet.
• The binary trap. “Can this task be automated?” is the wrong question. Many tasks aren’t automate-or-not; they’re collaborate-at-some-level. There was no shared vocabulary to express “I want the AI to do 70% and check with me on the rest.”

The result: a giant misallocation. Builders point agents at tasks workers resist (and that may trigger backlash) while ignoring tasks workers are begging to offload. This paper is an attempt to put a measured demand signal in front of the people deciding what agents to build.

What’s New (Core Contribution)

Four concrete contributions. None is a new model — this is a measurement and framework paper, and that’s exactly its value.

1. The Human Agency Scale (HAS), H1-H5. Before: automation was a binary, or borrowed the SAE self-driving levels (L0-L5) which are “AI-first.” Now: a human-centered five-point scale — H1 (AI does it alone) to H5 (cannot function without continuous human involvement) — that gives workers, researchers, and builders a shared language for “how much human in the loop.” Crucially, higher isn’t better; H1-H2 tasks suit automation, H3-H5 suit augmentation.

2. A worker-centric auditing framework with an audio interface. Before: surveys ask binary or Likert questions and get shallow, speculative answers. Now: an audio-enhanced mini-interview lets workers talk through their reasoning, plus task-familiarity filtering (you only rate tasks you actually do) and guided consideration (you’re prompted to weigh job security, enjoyment, etc. before rating). This is a methodological contribution about eliciting calibrated preferences, not just collecting numbers.

3. The WORKBank database. Before: no dataset jointly held worker desire and expert capability at the task level. Now: 844 tasks across 104 occupations, sourced from the U.S. Dept of Labor’s O*NET, with worker desire A_w(t), worker-desired agency H_w(t), expert capability A_e(t), and expert-feasible agency H_e(t). First of its kind, and designed to be extensible as capabilities evolve.

4. The desire-capability landscape (four zones) + an investment audit. Before: no way to see supply (what’s being built) vs. demand (what workers want). Now: plot every task on desire × capability, get four zones (Green Light, Red Light, R&D Opportunity, Low Priority), then map YC companies and arXiv papers onto those zones to show where effort is actually going. This is the contribution with the most commercial bite.

How It Works (Technically)

There’s no neural architecture here. The “mechanism” is a measurement pipeline: how four numbers per task get collected, validated, and combined into the maps that produce the findings. Treat it like a data engineering + survey design system.

The four core quantities

For every task t, the framework produces four scores:

Everything downstream is a comparison of these four numbers. The intelligence is in (a) eliciting them honestly and (b) what you do when worker desire and expert capability disagree.

Step 1 — Source the tasks (and pick the right granularity)

Tasks come from O*NET, the U.S. Dept of Labor’s occupational database. The deliberate design choice: audit at the task level, not the occupation level. “Marketing Manager” is too coarse — within it, “compile product lists” and “negotiate contracts” have wildly different automation profiles. They filter to computer-compatible, multi-step tasks performed at least monthly. 2,131 tasks across 287 occupations survive the filter; after requiring ≥10 worker responses per occupation, the final set is 844 tasks / 104 occupations.

Step 2 — Elicit worker preferences without garbage-in

This is where the audio interface earns its place. Three design tricks fight the usual survey failure modes:

• Audio-enhanced reflection: the survey opens with a spoken mini-interview about the worker’s job and AI views. Talking is faster and richer than typing, so workers contextualize their numeric ratings against real experience instead of guessing.
• Task-familiarity filtering: you only get tasks relevant to your occupation, and you must confirm familiarity before rating. Kills speculation.
• Guided consideration: before rating, you’re explicitly prompted to weigh enjoyment and job-security concerns (for desire) and task characteristics like physical action, domain expertise, uncertainty, interpersonal elements (for HAS). This anchors ratings in the factors prior literature says actually drive them.

The audio transcripts are later mined with LLM-based topic modeling to extract qualitative themes (e.g., 28% of workers expressed fear/concern; top worries were trust 45%, job loss 23%, loss of human touch 16%).

Step 3 — Get the expert counterweight

Workers know their jobs but not the frontier of agent capability. So 52 AI experts (PhD researchers + practitioners) independently rate A_e(t) and H_e(t) using the same rubrics. Each task gets ≥2 expert ratings, with extra review until standard deviation ≤1. Agreement is reported via Krippendorff’s α (0.539 for capability, 0.511 for feasible agency) — moderate agreement, which is honest for subjective forecasting and the paper doesn’t oversell it.

Krippendorff’s α is an inter-annotator reliability metric: 1.0 = perfect agreement, 0 = chance. ~0.5 means raters agree well above chance but there’s real spread — expect noise in any single task’s expert score, trust the aggregate trends.

Step 4 — Build the desire-capability landscape

Plot every task with A_w(t) on one axis and A_e(t) on the other. Split each axis at its midpoint to get four quadrants:

• Green Light (high desire, high capability): build these now.
• Red Light (high capability, low desire): you can but workers don’t want it — proceed with caution, expect resistance.
• R&D Opportunity (high desire, low capability): workers want it, tech isn’t there — the research frontier.
• Low Priority (low desire, low capability): leave it.

Key statistic: desire and capability are barely correlated (Spearman ρ = 0.17). That low correlation is the finding — it means feasibility is a terrible proxy for demand, so building “what’s possible” systematically misses “what’s wanted.”

Step 5 — Audit where the money and research actually go

They take YC company descriptions and arXiv agent papers, then use gpt-4.1-mini to map each onto the O*NET tasks it addresses. Now you can count startups-per-zone and papers-per-zone. Result: YC effort is spread roughly evenly across zones — meaning 41% lands in Red Light + Low Priority (where workers don’t want it), while Green Light and Opportunity zones are under-built. Research papers skew slightly better (concentrated in R&D Opportunity) but still cluster on a handful of CS/engineering tasks.

Step 6 — Project the skill shift

Map each task to its O*NET “Generalized Work Activities” (skills). For each skill, compute (a) average wage (from BLS, a proxy for current economic value) and (b) average expert-feasible human agency H_e(t) (a proxy for how human-dependent it stays). Rank skills both ways and look at which move. Information-processing skills (analyzing data, updating knowledge) are high-wage today but low-agency (AI takes them) — they’re devalued. Interpersonal/organizational skills rise. That divergence is the “future of skills” signal.

Architecture & data flow

flowchart TD
  ONET[O*NET task database] -->|filter: computer-compatible, monthly, multi-step| TASKS[2,131 tasks / 287 occupations]
  TASKS --> WSURVEY[Audio-enhanced worker survey]
  TASKS --> ESURVEY[Expert assessment rubric]
  WSURVEY -->|1,500 workers, >=10 per occ| AW["A_w(t): automation desire"]
  WSURVEY --> HW["H_w(t): desired human agency"]
  WSURVEY --> AUDIO[Audio transcripts]
  ESURVEY -->|52 experts, >=2 per task| AE["A_e(t): capability"]
  ESURVEY --> HE["H_e(t): feasible agency"]
  AUDIO -->|LLM topic modeling| THEMES[Concerns & collaboration paradigms]
  AW --> LAND[Desire-Capability Landscape: 4 zones]
  AE --> LAND
  HW --> HAS[HAS spectrum analysis]
  HE --> HAS
  YC[YC companies] -->|gpt-4.1-mini mapping| LAND
  ARXIV[arXiv agent papers] -->|gpt-4.1-mini mapping| LAND
  AE --> SKILL[Skill shift: wage vs. human-agency]
  HE --> SKILL
  LAND --> WORKBANK[(WORKBank)]
  HAS --> WORKBANK
  SKILL --> WORKBANK

Interactive desire-capability landscape (schematic, illustrative points). Each dot is a task plotted by worker automation desire (x) vs. expert-assessed capability (y). Hover to see which of the four zones it lands in. This is the core lens of the paper — note how desire and capability barely track each other.

The algorithm, simplified

There’s no training loop; the “algorithm” is the audit pipeline that turns ratings into a zoned demand map. Here’s the core of it in runnable-looking Python:

# Build the desire-capability landscape and audit where supply is pointed.
# A_w, A_e in [1..5]; tasks come from O*NET.

def zone(desire, capability, mid=3.0):
    # Split each axis at the midpoint -> four quadrants
    hi_d, hi_c = desire >= mid, capability >= mid
    if hi_d and hi_c:   return "GREEN_LIGHT"      # want it + can do it -> build now
    if hi_c and not hi_d: return "RED_LIGHT"      # can do it, nobody wants it -> caution
    if hi_d and not hi_c: return "RD_OPPORTUNITY" # want it, can't yet -> research
    return "LOW_PRIORITY"                          # neither -> ignore

def audit_workbank(tasks, companies):
    # tasks[t] = {"A_w":..., "A_e":..., "desc":...}
    for t in tasks.values():
        t["zone"] = zone(t["A_w"], t["A_e"])

    # Map each startup to the task(s) it addresses, via an LLM judge
    supply = {z: 0 for z in ("GREEN_LIGHT","RED_LIGHT","RD_OPPORTUNITY","LOW_PRIORITY")}
    for c in companies:
        for t_id in llm_map_company_to_tasks(c["desc"], tasks):   # gpt-4.1-mini
            supply[tasks[t_id]["zone"]] += 1

    # The finding: is investment chasing demand, or just feasibility?
    misaligned = supply["RED_LIGHT"] + supply["LOW_PRIORITY"]
    total = sum(supply.values())
    return supply, misaligned / total      # paper: ~41% misaligned for YC

The conceptual core, as numbered steps:

1. Collect A_w(t) and A_e(t) per task (honest elicitation is the hard part, not the math).
2. Quadrant-split into four zones.
3. Use an LLM to map external supply (startups, papers) onto tasks → zones.
4. Compare the supply distribution against the demand zones to expose mismatch.

Built on Prior Work

Results & Evidence

What was tested: 844 tasks, 1,500 workers (filtered from 1,676 / 7,016 ratings), 52 experts, Jan-May 2025. Representativeness validated against BLS sector distributions.

Headline numbers:

• 46.1% of tasks get positive automation desire (A_w > 3) — even after workers were prompted to consider job loss and enjoyment. People want help.
• Top motivation: “free up time for high-value work” (69.4% of pro-automation responses). This is the augmentation story, not the replacement story.
• Desire ↔ capability correlation is only ρ = 0.17 — feasibility ≠ demand. The central empirical claim.
• 41.0% of YC company-task mappings land in Red Light + Low Priority zones. The investment-misallocation headline.
• Usage lags demand: the 10 highest-desire occupations = 1.26% of Claude.ai usage.
• H3 (equal partnership) is the dominant desired level in 47 of 104 occupations — workers want collaboration, not full handoff. Workers consistently want higher human agency than experts say is technically necessary (47.5% of tasks fall in the “lower-triangle” mismatch).
• Skill shift: information-processing skills (analyzing data) devalue; interpersonal/organizational skills rise in importance.

What the evidence does NOT establish — read this before you quote the stats:

• It’s a snapshot, early 2025. Capability ratings will be stale fast; A_e(t) from then may already be low today.
• Stated preference ≠ behavior. Asking “do you want this automated” is not the same as adoption. People misforecast their own future behavior.
• Self-report bias. Workers may withhold honest answers fearing surveillance or job loss (authors acknowledge this).
• Expert agreement is moderate (α ≈ 0.5). Per-task capability scores are noisy; only aggregate trends are reliable.
• 104 of 287 occupations, computer-compatible only — no physical work, no fully novel AI-era tasks (O*NET only has existing tasks).
• The YC/arXiv mapping leans on an LLM judge (gpt-4.1-mini) with no reported validation of mapping accuracy — the 41% figure inherits that uncertainty.
• Midpoint zone split is arbitrary. Where you cut the axes changes the zone counts; treat the four zones as directional, not precise.

The evidence strongly supports the qualitative claims (demand ≠ feasibility; workers want augmentation over replacement; investment is misaligned) and should be treated as directional, not precise, on the quantitative ones.

How You’d Use It

This paper is a demand map for an AI services business — that’s its highest-value use to you. Concrete applications:

• Lead-gen / niche selection. WORKBank is, effectively, a ranked list of “tasks people will pay to offload, by occupation.” Pull the Green Light zone for your target verticals and you have a pre-validated services menu. The under-served Green Light / R&D Opportunity zones are your white space — competitors (the 41% in Red/Low) are fighting over tasks workers don’t even want.
• Avoid the Red Light trap. Before pitching an automation, check whether the task sits in high-capability/low-desire territory. If it does, reframe the offering as augmentation (HAS H3) or you’ll hit adoption resistance no matter how good the tech is.
• Scope engagements with HAS as a contract term. “We’ll build this to H2 (you check at key points)” vs. “H3 (equal partnership)” is a crisp way to set client expectations about human-in-the-loop, and it maps directly to how much oversight tooling you need to build.
• Positioning and content. The “free up time for high-value work” framing (69% of workers) is the exact message that lands — sell time reclaimed, not jobs replaced. The skill-shift finding (interpersonal skills rising) is a ready-made thought-leadership angle for client execs worried about workforce planning.
• Discovery methodology. Steal the audio-mini-interview + task-familiarity-filter approach for your own client discovery. It’s a better way to find what to automate than asking a client’s manager what they think their team wants.

Build Your Own (Minimal Recipe)

You don’t need 1,500 workers. The valuable, reusable artifact is a mini desire-capability audit for one client or vertical. Smallest version that captures ~80% of the value:

Components, in build order:

1. Task list. Pull the client’s occupation(s) from O*NET (it’s free, downloadable). Filter to computer-compatible, multi-step tasks. Or just shadow/interview to list 30-60 real tasks.
2. Desire elicitation. A short survey to the team: per task, “if AI could do this completely, how much would you want it to? (1-5)” plus the HAS question “how much collaboration would you want? (H1-H5).” Add the guided prompts (enjoyment, job security) — they materially change answers.
3. Capability scoring. You (or your engineers) play the “expert” role: rate each task’s current agent feasibility 1-5. You already have this judgment.
4. The zoning function. Literally the zone() snippet above. Quadrant-split, label.
5. Output. A two-axis scatter and a ranked Green Light list. That’s a deliverable a client will pay for on its own.

The 1-2 genuinely hard parts:

• Honest elicitation. Self-report is biased by fear and politeness. The audio interview + familiarity filter is the real IP. If you skip it you get noise. Minimum viable mitigation: anonymize responses and prompt explicitly about job-security concerns.
• Capability scoring drift. “Can an agent do this?” is a moving target and subjective. Use ≥2 raters and record the date; treat scores as 6-month-perishable.

Reach for: O*NET bulk database (free), any survey tool (Typeform/Google Forms) or a small custom form for the audio step (Whisper for transcription), an LLM for topic-modeling the transcripts, pandas + matplotlib/plotly for the landscape. Total: a few days for a single-vertical pilot.

How to Improve It

Limitations are the leverage. Five concrete, testable directions:

1. Close the stated-vs-revealed-preference gap. Pair the survey with an actual agent pilot for a subset of tasks, then measure adoption. Hypothesis: revealed adoption correlates with A_w(t) but with a predictable discount — quantify the discount and you have a far more useful demand model.
2. Make capability scoring empirical, not expert-opinion. Replace A_e(t) expert Likert ratings with actual agent benchmark pass-rates on representative task instances (à la GAIA/AgentBench but mapped to O*NET tasks). This removes the α≈0.5 noise and makes the landscape auto-updatable as models improve.
3. Continuous / longitudinal WORKBank. The snapshot decays. Stand up a recurring quarterly audit so the zones move over time — that time-series is worth more than any single snapshot and would show R&D Opportunity tasks migrating into Green Light as capability lands.
4. Model emergent tasks. O*NET only has existing tasks. Mine the open-ended audio transcripts (they have them) to surface new AI-era tasks (“supervise the agent,” “audit AI outputs”) that the framework currently can’t see.
5. Validate the LLM mapping. The 41%-misaligned headline depends on gpt-4.1-mini mapping companies/papers to tasks with no reported accuracy check. Hand-label a sample, report precision/recall, and re-run — the number could move materially.

Glossary

• WORKBank — the paper’s database: worker desire + expert capability ratings for 844 occupational tasks across 104 occupations.
• O*NET — U.S. Dept of Labor’s standardized database of occupations, their tasks, and required skills; the task backbone here.
• BLS — U.S. Bureau of Labor Statistics; source of wage data used as a proxy for a skill’s economic value.
• A_w(t) / A_e(t) — automation desire (worker-rated) and automation capability (expert-rated) for task t, on a 1-5 Likert scale.
• H_w(t) / H_e(t) — worker-desired and expert-feasible human agency level for task t, on the HAS scale.
• Human Agency Scale (HAS) — five-level scale (H1 = AI alone → H5 = essential continuous human involvement) for how much human is in the loop; H1-H2 favor automation, H3-H5 favor augmentation.
• Automation vs. Augmentation — replacing human work (automation) vs. complementing/enhancing it (augmentation). The whole paper argues the future is more augmentation than the hype implies.
• Desire-capability landscape — scatter of tasks by desire (x) vs. capability (y), split into Green Light / Red Light / R&D Opportunity / Low Priority zones.
• Likert scale — an ordered rating scale (here 1-5) for subjective agreement/intensity.
• Spearman ρ — rank correlation coefficient; ρ = 0.17 here means desire and capability barely move together.
• Krippendorff’s α — inter-rater reliability metric (1 = perfect agreement, 0 = chance); ~0.5 = moderate agreement among experts.
• Jensen-Shannon Distance (JSD) — a symmetric measure of how different two probability distributions are; used to quantify worker vs. expert disagreement on HAS per occupation.
• LLM topic modeling — using an LLM to cluster/extract themes from free-text (here, audio transcripts of workers’ concerns).
• Generalized Work Activities — O*NET’s standardized list of skills/activities (e.g., “analyzing data,” “guiding subordinates”) that tasks map to.