Mapping AI to real work
We broke all 923 US occupations into more than 270,000 steps and searched over 2,700 AI benchmarks for the ones that test them.
16 November 2026 · By the HeyMilo Labs team
Atlas exposes the gap between what benchmarks test and the work people actually do. It maps every occupation down to the steps of its real work and shows which of those steps any benchmark covers.
Other efforts go deep on a small set of jobs with human verifiers. Atlas goes wide first: every occupation, broken down to every step, with the evidence behind each judgment attached. Then it opens the map so experts can deepen it.
Benchmarks are our best way to measure what AI can do. But as they've turned into leaderboards to win, we've drifted from why we benchmark in the first place: to test whether a model understands, reasons, and can justify its work, rather than just pattern-matching its way to an answer.
Atlas is an open, public map shaped by the people who use it. It's built to help anyone understand where AI meets human work, and to turn its blind spots into benchmarks and environments with real economic value and impact. We go deeper into our findings below and explain what we’re building towards.
How do we know it’s real work? Every task comes from O*NET, the US government’s catalogue of what people do in each job, and every judgment cites the papers, datasets and code behind it. An AI judge makes each judgment and a second pass checks it (see how). On top of that, human experts are now reviewing Atlas domain by domain. That takes time, so if you know a field, help us review it or get in touch.
Benchmark coverage at launch
Atlas maps all 923 occupations, nearly 19,000 tasks and more than 270,000 steps. Here's how much of that work today's benchmarks test.
2,721 benchmarks analyzed
What we’re building towards
The goal is a score for every job that reflects the real work, with model performance weighted by each step, its share of working time and its wages. Atlas already links coverage, time and wages to every step, and nearly 20,000 steps have a published model score. What stands in the way:
- Most work isn’t measured yet. Only 9% of steps are known to have a benchmark testing them, even partly, and over half haven’t been judged yet. Unmeasured work can’t be scored, so we’ll turn the gaps into benchmarks and RL environments with HeyMilo Labs but we can’t do it alone.
- Scores are for whole benchmarks. A model’s result covers the whole benchmark, not the step. We’ll tie item-level results to the steps they test.
- Expert validation. Every judgment so far comes from AI. We’re bringing in domain experts and the people who do the work to check it, fix what’s wrong and add what’s missing, and we’ll publish how often they agree.
- Unknown steps. About 151,000 steps still need judging, mostly in the mapped occupations.
- Work O*NET doesn’t list. We’ll add new and changing tasks as real work evolves.
- Physical work. Hands-on steps are the least measured of all, and will need benchmarks of their own as robotics matures.
Working on any of these? Contribute or get in touch.
Methodology Brief
Atlas starts from O*NET's list of tasks for every US occupation. A single task can hide a lot of different work, so we break each one down into the individual steps a worker actually performs. We then judge every step against the benchmarks that might test it, reading each benchmark's own paper, code and task files rather than just its description, and a second, independent pass checks every coverage claim.
It's worth being clear about what coverage means. It tells you whether a benchmark tests a piece of work, not whether AI can do it, which is why model scores are shown separately. A few words in Atlas carry a specific meaning, explained on the Terminology tab, and the full method is on the Methodology tab.
Limitations
This is our initial release. Read the findings below with these in mind.
- One AI judge
- Every judgment comes from one model and is checked by an independent second pass. We’re having experts review judgments, tasks and more.
- Machine-written steps
- Tasks are split into steps by the same model, and experts are validating those steps.
- Two depths of search
We screened 153 occupations against every candidate benchmark. For the remaining 770 occupations, which account for 81% of all tasks, we used a benchmark-to-task map to select a narrower set of candidates. When tested against the fully screened occupations, the map recovered 80% of full matches and 47% of partial matches covering at least 30% of task steps.
Some apparent gaps in the second group therefore reflect missed matches rather than missing benchmarks. Its measured coverage is a lower bound, and comparisons between the two groups should be treated with caution.
- Model-shortlisted candidates
- Atlas draws on over 2,700 public benchmarks. For the 153 screened occupations, a language model shortlists the first batch of candidates by matching what each benchmark sets out to test against the occupation. A benchmark left off that shortlist is not judged there.
- Uneven evidence
- When only a benchmark's metadata can be read its judgments default to unknown. Unknown makes up 23.2% of all judgments.
- Scores skip unknowns
- A coverage score averages only the judged steps. A task with 5 covered and 5 unknown steps scores 1.0 while half of it is unassessed, so read scores alongside their evidence completeness.
- Benchmark-level scores
- A model's score is for the whole benchmark and not the single step it covers. A model can score 99% on a benchmark and still fail the one item that tests a step.
- Estimated economics
- Each task's share of working time is estimated from O*NET ratings and some occupations have none. Specialties that share one BLS row split its employment equally.
- Relative to O*NET
- Shares of working time count time on O*NET's tasks. Work that O*NET does not list is not measured at all.
Each of these shrinks as people contribute: expert workflows check the steps, new benchmarks and sources resolve unknowns, and reviewers check the judgments. See What we’re building towards and the Contribute tab.
Findings
Landscape
Most occupations are touched by at least one benchmark but coverage thins out quickly at the level of tasks and steps where the actual work happens. An occupation or task counts as covered here if a benchmark tests at least one of its steps, even partly.
613 of 923 occupations have at least one step a benchmark partly tests but only 143 of 923 occupations have a step that is fully tested.
3,808 of 18,829 tasks have a covered step but only 282 of 18,829 tasks have a fully covered one. None is covered end to end and more than half had no benchmark close enough to check at all.
23,954 of 271,399 steps are tested at all and only 733 of 271,399 steps are fully tested.
36% of US working time goes to tasks with a covered step and only 4.5% to tasks with a fully covered one. Nearly two thirds goes to tasks no benchmark touches.
2,721 benchmarks were analyzed and 1,930 were checked against at least one task.
One occupation task up close
Software Developers are among the best-covered jobs in Atlas, yet even their everyday work is only partly tested. Take the O*NET task “Modify existing software to correct errors, adapt it to new hardware, or upgrade interfaces and improve performance.” Atlas splits it into four paths of six steps each, and the coverage varies a lot between them. Fixing errors is fully covered and improving performance nearly so, but adapting software to new hardware and upgrading interfaces and integrations are only partly tested, and keeping older hardware supported has no benchmark at all.
If one of the most measured occupations still has evaluation gaps like these, there is a lot of work to be done for every other occupation. Every occupation in Atlas is broken down the same way into tasks, paths, subtasks and steps, with the benchmark behind each judgment.
Benchmarks
To see where benchmarks focus, we gave every task a dollar value: what the US pays for that work each year, based on how many people do it, their median wage and the share of their time it takes. Benchmark builders clearly lean toward the work that pays most. Among the most valuable quarter of tasks, 35% have at least one tested step, compared with just 9% of the least valuable.
Even so, most high-value work is still untested. 65% of the most valuable tasks have no tested step at all, and only 3.6% have a step that is fully tested. That thin slice of measured work is what claims about AI’s economic impact rest on today, and the gap around it is where a new benchmark or RL environment would pay off most.
For Atlas we chose breadth, reading over 2,700 benchmarks, so that “untested” means no benchmark we could find tests the work and not just that it is missing from a short list. That breadth also lays the foundation for Atlas to build on. The Benchmarks tab shows which of them cover the most steps.
Coverage by job family
Benchmarks are built around desk work. Computer, maths and legal jobs are the best covered yet even there fewer than three in ten steps are tested. Hands-on trades come last and most of their work has never been near a benchmark. This is not only a robotics gap though. Most work done on a computer is untested too and benchmarks tend to go where AI is already expected to help. The work we know least about is the work nobody expected AI to touch.
Model scores
Many benchmarks publish AI scores, such as a model’s pass rate on a coding benchmark. Atlas can link a published score to only about 20,000 of its more than 270,000 steps. For the other 93% there is no score at all so nobody knows how AI would do on that work.
What to build next
Atlas contains about 574,000 judgments: 1,520 found full coverage and about 61,000 found partial coverage. The fully covered steps come from 281 benchmarks and 22 of them account for half. They are led by a teaching-workflow benchmark, Terminal-Bench 4.0, ERP-Bench and DeepSWE 1.1. See the top 10 on the Benchmarks tab.
Every partial judgment names the missing part in a sentence and so becomes a specification for the next benchmark. Each one can be checked: 98.8% of judgments cite a passage from the benchmark verified word for word and every judgment, quote and missing part can be opened in Atlas.
- Extend
- 851 tasks could be covered by extending a benchmark that already exists.
- Rebuild or research
- Of the roughly 8,500 tasks checked against a benchmark, 34% need one rebuilt and 56% need more research. None can use a benchmark as it is.
- Where it pays
- The Build opportunity lens in Atlas ranks valuable work that benchmarks cover weakly. It is the starting point for new benchmarks and RL environments.
Mission
Our mission is to help the world understand where AI meets human work. A built out Atlas lets people in every industry see where they stand, support each other and move forward together, so that no one gets left behind.
Our methodology laid the first foundation of Atlas and we will keep building on it until the map reflects the work people really do and how AI compares to it. From there, we’ll continue to push and evolve Atlas to help the world understand the AI landscape.
Our background at HeyMilo Labs, understanding human workflows and skills across occupations, gives us a head start in mapping human work to AI. Read more on the HeyMilo Labs tab.
No single team can map work at this scale alone. We need researchers, engineers, domain experts and workers to help keep Atlas accurate for everyone. Your work won't go unnoticed, and we'll make sure it gets the credit it deserves. Learn more on the Contribute tab, or talk to us directly through the Contact us tab. We’re also open to collaborating across the RL development lifecycle, from data and benchmarks to environments and training.
How Atlas breaks down work
Atlas starts from O*NET, the US government’s catalogue of work, which lists the tasks of every US occupation. Atlas then splits each task into paths, subtasks and steps.
From O*NETEarlier benchmark maps go this deep and stop at the task
- Domain
A group of related occupations. The top level of the map.
Healthcare Diagnosing or Treating Practitioners - Occupation
One O*NET occupation, with its BLS employment and median wage.
Emergency Medicine Physicians - Task
One O*NET task statement for that occupation.
Analyze records, examination information, or test results to diagnose medical conditionsAdded by AtlasThe work inside each task, where coverage is judged
- Path
One version of the task with its own goal. Many task statements bundle several kinds of work; instead, Atlas maps each one separately.
Analyze patient medical records and historical clinical findings to establish or refine a diagnosis - Subtask
A stage within a path that groups related steps.
Synthesize medical record findings to formulate, narrow, or confirm a diagnosis - Step
The smallest unit of work. Benchmarks are judged against steps.
Record the working, differential, or confirmed diagnosis supported by medical record analysisTrigger what starts itEnd state what done looks likeDigital score 0 to 1, how much is done on a computer
Measures and lenses
- Benchmark coverage
- Whether a benchmark's items test a step the way a worker would have to do it. It measures the benchmarks, not AI ability.
Shown as0–100%Strong 67%+Moderate 33–67%Weak under 33%No evidence
- Coverage class
- The judgment a step gets. A partial judgment names the part the benchmark leaves out.
Shown asCoveredPartly coveredNot coveredNot enough evidence
- Coverage scorealso task or path score
- A score from 0 to 1 for how much of a task or path benchmarks test. Each judged step scores 1 if covered, 0.5 if partial and 0 if not covered, and the score is their average. Unknown steps are left out. For example, a task with 4 covered, 2 partial and 4 not covered steps scores (4 + 1) ÷ 10 = 0.50. A task counts as covered at 0.70 or above, partial from 0.35 and not covered below that. If fewer than half its steps are judged it shows as unknown.
- Evidence completeness
- The share of steps with a decided judgment. Below half, coverage shows as unknown, with hatching.
Shown as0–100% of stepsHigh 67%+Medium 33–67%Low under 33%None
- Model performance
- The best model score on the benchmarks that cover a step. Scores are for the whole benchmark, not the single step.
Shown as0 to 1HighMediumLowNo scores
- AI status
- Coverage and model scores read together. Models score well at 0.70 or more.
Shown asTested: models score wellTested: models score poorlyNot tested by any benchmarkUnknown
- Economic value
- Estimated annual value of the work, from BLS employment and median wages and the task's estimated share of working time.
Shown asUS dollars a yearTop, middle and bottom thirdNo economics
- Build opportunity
- Valuable work that benchmarks cover weakly: where a new benchmark or environment would add the most.
Shown asPercentile among tasks at the same level, 0–100%High, medium and low thirdsNeeds economics
- Build decision
- What to do with existing benchmarks for a task.
Shown asUse as-isExtend existingReconstructResearch furtherDo not use
- Readiness
- How far a task has got towards a ready-to-train RL environment, from discovery complete (R0) to training ready (R8). Each level has to be passed before the next. Every task is at R0 in this first release.
Shown as
- R0Discovery complete
- R1Workflow ready for expert review
- R2Benchmark analysis complete
- R3Build recommendation ready
- R4Human calibrated
- R5Prototype implemented
- R6Verifier calibrated
- R7Runtime validated and admitted
- R8Training ready
- Evidence grounding
- Where a conclusion comes from.
Shown asCode-confirmedSource-backedInferredSyntheticContradictedInsufficient evidence
Earlier benchmark maps count an O*NET task as covered once enough benchmark prompts are tagged to it. But a task is a chain of steps. Fixing a bug means reproducing it, finding the cause, editing the code and proving the fix works, and a benchmark can test the edit without ever testing the proof. So Atlas judges coverage step by step.
Process
- SourcesO*NET tasks and their frequency and relevance ratings, BLS OEWS employment and wages from May 2025, and 2,721 public benchmarks with their papers, code and published results.O*NET 31.0
- Read each benchmarkIts paper is read whole, with the README, the repository's grader, task and data files, and the dataset card. Every passage is checked word for word before it can be cited.2,721 read
- DecomposeEach task is split into paths, subtasks and steps. Every step gets a trigger, an end state, an effort weight and a digital score. Coverage is judged per step, not per task.271,399 steps
- Choose candidatesBenchmarks are tagged to the tasks they might test, but only to nominate candidates. Coverage is decided in the two passes. Screening is used for 153 occupations and mapping for the other 770, as set out below.1,930 checked
- Pass 1Each step is judged against each candidate, with a verdict, a confidence and a rationale citing verified passages. A step is covered when the benchmark tests all of it and partial when it leaves out a named part. A step can only be marked not covered if the benchmark’s paper was read in full. Otherwise it stays unknown.574,443 judgments
- Pass 2A blind, adversarial review of every coverage claim. When the passes disagree, the more conservative verdict is kept and marked for human review.88% agree · base run
- ScoresPublished model results, matched to a model registry and checked against each paper. Human and chance scores become baselines.54,941 results
Choosing candidates
This is the only stage that differs between occupations.
Screening153 occupations, 3,510 tasks
- Candidates
- A first batch of benchmarks shortlisted for the occupation by a language model, each then screened against the task's steps
- Checked per task
- Up to 10
Mapping770 occupations, 15,319 tasks
- Candidates
- Each benchmark tagged once to every O*NET task its items exercise, following Wang et al.
- Checked per task
- About 3.5 when a benchmark maps, none otherwise
How well the mapping finds benchmarks
Tested against the pairs the screened occupations had already judged.
| Pair found by screening | Pairs | Recall | |
|---|---|---|---|
| Fully covers a step | 102 | 80% | |
| Partly covers 30% or more of steps | 1,204 | 47% | |
| Partly covers a few steps | 2,553 | 24% | |
| All pairs | 3,859 | 32.5% |
Sources
- Wang et al., How Well Does Agent Development Reflect Real-World Work? (arXiv 2603.01203)
- Eloundou et al., GPTs are GPTs (2023)
- O*NET 31.0, US Department of Labor
- BLS Occupational Employment and Wage Statistics, May 2025
- Mercor, A benchmark map of the US economy
Includes information from the O*NET 31.0 Database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA), used under CC BY 4.0. O*NET® is a trademark of USDOL/ETA. This information has been modified and USDOL/ETA has not endorsed these changes.
Top 10 benchmarks by steps covered
Of the 1,930 benchmarks we checked against Atlas's more than 270,000 steps, only 281 fully cover even a single one. These are the ten that cover the most, ranked by the steps they fully cover and then by those they partly cover, with each benchmark listed once under its latest version. Steps are counted per occupation, so a benchmark that tests work shared by many similar occupations counts it once for each. TeachArena ranks first partly because O*NET lists college teaching as dozens of separate occupations, one per subject, with nearly the same tasks.
| # | Benchmark | Main domain | Occupations | Fully covered | Partly covered |
|---|---|---|---|---|---|
| 1 | TeachArenaHKUST and Alibaba Qwen | Postsecondary Teachers | 64 | 52 | 2,596 |
| 2 | Terminal-Bench 4.0Stanford and Laude Institute | Computer Occupations | 13 | 41 | 198 |
| 3 | ERP-BenchAgentic Labs | Business Operations Specialists | 111 | 30 | 1,226 |
| 4 | DeepSWE v1.1Datacurve | Computer Occupations | 8 | 30 | 116 |
| 5 | QFBench | Mathematical Science Occupations | 10 | 29 | 175 |
| 6 | Android BenchGoogle | Computer Occupations | 5 | 28 | 54 |
| 7 | PhysicianBenchStanford | Healthcare Diagnosing or Treating Practitioners | 33 | 27 | 687 |
| 8 | APEX-SWEMercor and Cognition | Computer Occupations | 14 | 26 | 247 |
| 9 | SpreadsheetBench 2Renmin University | Financial Specialists | 18 | 22 | 197 |
| 10 | DACompCAS and ByteDance Seed | Computer Occupations | 23 | 21 | 535 |
Software Developers
This is a sample of what Atlas holds for one occupation with a limited view. We’ll be doing a full launch soon so stay tuned! O*NET, the US government’s catalogue of work, lists 17 tasks for Software Developers. Atlas splits them into 46 paths and 280 steps for our initial release and then checks every step against 550 benchmarks. Three tasks are open below step by step.
Fixing and improving existing code is the best-tested work a software developer does, with 11 of its 24 steps fully covered. Setting performance standards is the opposite: no benchmark tests any of its 16 steps. That gap on work worth about $10B a year ranks it 4th of the nearly 18,000 tasks Atlas ranks for build opportunity.
- Workers in the US
- 1.69M
- Median annual wage
- $135,980
- Steps mapped
- 280
- Steps fully covered
- 18 of 280
Help map the work
Atlas is open. If you know how a job is really done, or you build benchmarks, you can add to the map. Every contribution goes to a review queue and is checked against its sources before it changes anything.
Who can help
- Researchers
- Add a benchmark, paper, dataset or evaluator we haven’t judged yet.
- People who do the work
- Show us how a task is really done, with your own write-up, a manual or a procedure, and add tasks or steps the map is missing.
- Teams and companies
- Map the work your team does and link the benchmarks that test it.
- Anyone
- Spot something wrong? Flag it, or suggest a fix with a source.
How it works
- Sign in to Atlas and press Contribute on the map.
- Pick the occupation, task or step it belongs to.
- Reviewers check it against its sources. Accepted work goes into a future release.
Review others' contributions
Contributions are reviewed by the community before the team decides. If you can review, a Review queue tab appears next to Contribute.
- Two queues
- For you holds contributions routed to you. General holds everything else.
- What you see
- The proposal and its sources, an AI pre-review that checks the links, and competing proposals for the same record side by side. Show on map takes you to the record.
- Your vote
- Accept, reject or needs changes, with a short reason. Add a source link or citation if you have one.
- What happens next
- After three votes the item goes to the Atlas team, who make the final call. Accepted changes arrive in a future release.
- Reviewer points
- When the team's decision agrees with your vote, you earn points, shown in the queue.
Advancing AI for real work
Atlas and HeyMilo Labs work together to keep the map true to how work is actually done, and to keep it current as that work changes.
Building HeyMilo shows us the roles businesses hire for, the skills they need and how the work gets done, across staffing agencies, expert networks and enterprise teams.
Understanding human workflows and skills across occupations gives us a head start on building AI that models real work. We use that insight to find gaps in AI evaluation and guide the data, benchmarks and RL environments we build.
What recruiting adds to Atlas
- Grow the map
- Add the roles, tasks and steps we see in real hiring, starting where O*NET is thin.
- Verify the steps
- Check machine-written steps against how the people we recruit describe and do the work.
- Find new work
- Catch tasks as workflows evolve, including work that only exists because of AI.
Talk to the Atlas team
Ask about the method, flag a benchmark we missed or a judgment you think is wrong, or get in touch about working together.
Message sent
We'll get back to you as soon as possible!



