Talvio Methodology | How the Training Priority score works

How the score is built

In two stages. The first happens once for every occupation. The second happens each time you upload a roster. No AI model runs when a score is produced, and the same inputs always give the same score.

Built once for every occupation

Current version: July 2026. Rebuilt only when the method changes.

What each job involves
O*NET rates 41 kinds of work activity for each occupation: how important each one is to the job, and at what level it is done.
How much AI can help with each activity
Each activity gets a value from 0 to 1 for how much of it AI can help with today. It uses the one AI skill that fits the activity best, such as drafting text or pulling information together, and a reviewed rating of how good AI is at that skill.
Add it up for the job
Activities that matter more to the job count for more. The result is the share of the job’s work that AI can help with. For every occupation it falls between 27% and 41%.
Lower it for physical work
The share is cut by how many of the job’s O*NET abilities are physical or hands-on. The cut is squared, so very physical jobs are held back more. This step separates jobs more than any other.
Put it on the 0.5 to 10 scale
A fixed rule stretches the results so the lowest 1% of occupations sit near 0.5 and the highest 1% near 9.5. The order does not change, and no job scores zero.

Applied to your roster at each upload

Your file changes which occupations are counted and how many people are in each.

Match each job title to an occupation
Titles are cleaned and compared with O*NET’s occupation titles and alternate titles. Matches under 90% confidence are flagged for review. For a flagged title, an optional AI assistant suggests three likely occupations and a person picks one. The assistant never sets a score.
Add up by department
Each role’s score is weighted by its headcount to give department and organization figures, and to place people in three training groups: 7 and above, 4 to 6.9, and under 4.

Show the math for these stepsHide the math

The score is deterministic: the same inputs always give the same score, and the whole calculation is arithmetic. No AI model runs when a score is produced. The score combines what an occupation does, from O*NET, with reviewed ratings of how good AI is at each skill today, based on public tests and published evidence.

1. How much each activity matters For each occupation and work activity, Talvio multiplies the O*NET Importance by the Level, then scales the results within the occupation so they add up to 1.

w_o,j = IM_o,j × LV_o,j ∑_j(IM_o,j × LV_o,j)

2. Which AI skill helps which activity A reviewed 41 x 12 table holds, for each work activity and AI skill, the share of that activity AI can help with, from 0 to 1. The values are the medians of five AI raters, each given a different professional role and none shown the scores, reviewed and adopted by a person on 2026-07-07.

0 ≤ M_j,k ≤ 1

3. How good AI is at each skill today Each AI skill carries a reviewed score from 0 to 10, used here as a fraction.

m_k = C_k10

4. How much AI can help with each activity Each activity is scored by the one AI skill best suited to it. Using the best single skill, rather than adding up all twelve, stops unrelated skills from inflating desk-work activities.

A_j = max_k(M_j,k · m_k)

5. Reduce it for physical work The desk-work total is reduced by the share of the job’s O*NET abilities that are physical or hands-on, squared. AI helps with thinking and communicating, not lifting and touching, and squaring the share means very physical jobs are held back more.

base_o = ∑_jw_o,jA_j raw_o = base_o · (1 − G_o)²

6. Put it on the 0.5 to 10 scale Raw scores are stretched onto the shown scale by a fixed rule that keeps the order the same. It is re-fitted only when the method version changes. The floor is 0.5: no occupation scores zero, because every job has some duties AI can help with.

s_o = 9 · (raw_o − p₁) / (p₉₉ − p₁) + 0.5 TAP_o = clip_{[0.5, 10]}(s_o)

The score measures the share of an occupation’s weighted work that AI can help with. It is not a rank, and it is not a claim about what AI can do in general. Two occupations with the same score have a similar share of work AI can help with; they do not necessarily do similar work. Everywhere the product explains a score, it names the O*NET work activities behind it, because that is what the score is built from. Method version: TPS v2.2, July 2026.

How US jobs compare

Counting each scored job by how many people hold it, the average US worker scores 4.2. Half of all workers (50%) are in jobs scoring under 3.5, mostly hands-on work. About a third (31%) score 6.5 or higher, mostly office and professional work.

10M

15M

20M

7 and above

12345678910

Score (0.5 to 10)

US workers by score, in half-point bands. Worker counts from the Bureau of Labor Statistics, May 2025. 96% of US workers hold a scored occupation; the rest are left out.

Average score by industry

Each industry’s average covers every job in it, office staff included. The 20 industry sectors range from 1.5 (agriculture, forestry, fishing) to 7.4 (finance and insurance). The dashed line marks all US workers.

All US workers 4.2

Finance and insurance7.4Professional, scientific, technical7.0Management of companies6.8Information6.5Educational services6.0Government (other than schools and hospitals)4.8Real estate4.8Utilities4.4Wholesale trade4.3Health care and social assistance4.1Other services3.9Administrative support and waste3.6Arts, entertainment, recreation3.5Retail trade3.4Manufacturing3.3Mining, oil and gas3.0Construction2.8Transportation and warehousing2.6Accommodation and food2.5Agriculture, forestry, fishing1.5

Show occupation groups and smaller industriesHide occupation groups and smaller industries

Average score by occupation group

All US workers 4.2

Legal8.5Business and Financial Operations8.2Computer and Mathematical7.4Management6.8Architecture and Engineering6.7Community and Social Service6.4Office and Administrative Support6.3Life, Physical, and Social Science6.2Educational Instruction and Library6.2Arts, Design, Entertainment and Media5.7Sales4.3Healthcare Practitioners and Technical3.6Personal Care and Service3.1Protective Service2.8Healthcare Support2.8Food Preparation and Serving Related2.2Installation, Maintenance, and Repair1.9Production1.9Transportation and Material Moving1.6Construction and Extraction1.6Building and Grounds Cleaning and Maintenance1.1Farming, Fishing, and Forestry1.0

Industries in more detail

Industry	Workers	Average score	Share scoring 7 or above
Accommodation and Food Services	14,276,590	2.5	1%
Accommodation	2,007,720	3.0	4%
Food Services and Drinking Places	12,268,870	2.4	1%
Administrative and Support and Waste Management and Remediation Services	9,114,350	3.6	14%
Administrative and Support Services	8,599,170	3.6	14%
Waste Management and Remediation Services	515,180	2.7	6%
Agriculture, Forestry, Fishing and Hunting	413,350	1.5	3%
Forestry and Logging	41,620	2.1	4%
Support Activities for Agriculture and Forestry	371,730	1.5	2%
Arts, Entertainment, and Recreation	2,765,320	3.5	7%
Museums, Historical Sites, and Similar Institutions	181,800	4.8	14%
Performing Arts, Spectator Sports, and Related Industries	603,790	4.4	16%
Amusement, Gambling, and Recreation Industries	1,979,740	3.1	4%
Construction	8,298,380	2.8	7%
Construction of Buildings	1,861,040	3.5	12%
Specialty Trade Contractors	5,243,000	2.6	6%
Heavy and Civil Engineering Construction	1,194,340	2.6	7%
Educational Services	13,852,050	6.0	36%
Educational Services	13,852,050	6.0	36%
Finance and Insurance	6,281,650	7.4	63%
Securities, Commodity Contracts, and Other Financial Investments and Related Activities	1,099,930	8.0	74%
Monetary Authorities-Central Bank	20,490	7.5	71%
Funds, Trusts, and Other Financial Vehicles	33,650	7.4	61%
Insurance Carriers and Related Activities	2,617,090	7.4	61%
Credit Intermediation and Related Activities	2,510,490	7.3	62%
Federal, State, and Local Government, excluding State and Local Government Schools and Hospitals and the U.S. Postal Service (OEWS Designation)	10,242,890	4.8	24%
Federal, State, and Local Government, excluding State and Local Government Schools and Hospitals and the U.S. Postal Service (OEWS Designation)	10,242,890	4.8	24%
Health Care and Social Assistance	24,031,380	4.1	9%
Ambulatory Health Care Services	8,947,200	4.6	11%
Hospitals	6,712,760	4.0	11%
Social Assistance	4,976,250	3.9	6%
Nursing and Residential Care Facilities	3,395,180	3.2	5%
Information	2,879,620	6.5	49%
Publishing Industries	910,660	7.3	67%
Computing Infrastructure Providers, Data Processing, Web Hosting, and Related Services	485,550	7.3	62%
Web Search Portals, Libraries, Archives, and Other Information Services	174,790	7.3	66%
Broadcasting and Content Providers	343,200	6.7	44%
Telecommunications	591,190	5.1	28%
Motion Picture and Sound Recording Industries	374,240	4.9	16%
Management of Companies and Enterprises	2,828,130	6.8	54%
Management of Companies and Enterprises	2,828,130	6.8	54%
Manufacturing	12,654,310	3.3	12%
Computer and Electronic Product Manufacturing	1,002,320	5.2	35%
Chemical Manufacturing	889,150	3.9	14%
Miscellaneous Manufacturing	616,780	3.9	13%
Machinery Manufacturing	1,092,170	3.7	17%
Electrical Equipment, Appliance, and Component Manufacturing	425,210	3.7	15%
Petroleum and Coal Products Manufacturing	112,470	3.5	12%
Printing and Related Support Activities	355,560	3.5	8%
Transportation Equipment Manufacturing	1,753,760	3.5	16%
Apparel Manufacturing	79,960	3.0	7%
Fabricated Metal Product Manufacturing	1,436,510	3.0	8%
Beverage and Tobacco Product Manufacturing	329,940	2.9	6%
Furniture and Related Product Manufacturing	339,750	2.7	7%
Paper Manufacturing	353,210	2.7	6%
Textile Product Mills	94,410	2.7	5%
Plastics and Rubber Products Manufacturing	705,590	2.7	7%
Primary Metal Manufacturing	364,190	2.6	7%
Nonmetallic Mineral Product Manufacturing	414,630	2.6	5%
Leather and Allied Product Manufacturing	23,760	2.5	4%
Food Manufacturing	1,778,490	2.5	4%
Textile Mills	79,900	2.4	5%
Wood Product Manufacturing	406,580	2.3	4%
Mining, Quarrying, and Oil and Gas Extraction	570,220	3.0	10%
Oil and Gas Extraction	113,950	4.6	26%
Support Activities for Mining	268,510	2.7	7%
Mining (except Oil and Gas)	187,760	2.5	5%
Other Services (except Public Administration)	4,484,430	3.9	13%
Religious, Grantmaking, Civic, Professional, and Similar Organizations	1,429,400	5.7	33%
Personal and Laundry Services	1,585,390	3.3	3%
Repair and Maintenance	1,469,640	2.9	6%
Professional, Scientific, and Technical Services	10,800,470	7.0	58%
Professional, Scientific, and Technical Services	10,800,470	7.0	58%
Real Estate and Rental and Leasing	2,427,950	4.8	25%
Lessors of Nonfinancial Intangible Assets (except Copyrighted Works)	21,250	7.4	57%
Real Estate	1,837,320	5.2	30%
Rental and Leasing Services	569,390	3.4	8%
Retail Trade	15,503,410	3.4	4%
Furniture, Home Furnishings, Electronics, and Appliance Retailers	764,510	4.1	10%
Health and Personal Care Retailers	1,076,240	4.1	5%
Clothing, Clothing Accessories, Shoe, and Jewelry Retailers	1,157,300	3.9	3%
Sporting Goods, Hobby, Musical Instrument, Book, and Miscellaneous Retailers	1,504,440	3.8	5%
Motor Vehicle and Parts Dealers	2,043,570	3.5	8%
Building Material and Garden Equipment and Supplies Dealers	1,392,800	3.4	3%
Gasoline Stations and Fuel Dealers	1,042,980	3.1	1%
General Merchandise Retailers	3,275,200	3.1	2%
Food and Beverage Retailers	3,246,380	2.8	1%
Transportation and Warehousing	7,448,640	2.6	5%
Support Activities for Transportation	812,950	3.8	10%
Pipeline Transportation	56,210	3.7	14%
Water Transportation	69,600	3.7	12%
Scenic and Sightseeing Transportation	30,850	3.6	4%
Transit and Ground Passenger Transportation	532,510	3.4	6%
Air Transportation	563,220	3.3	4%
Rail Transportation	190,930	2.8	6%
Truck Transportation	1,499,660	2.5	4%
Warehousing and Storage	1,963,680	2.3	4%
Postal Service (Federal Government)	618,900	1.9	2%
Couriers and Messengers	1,110,140	1.9	2%
Utilities	598,000	4.4	21%
Utilities	598,000	4.4	21%
Wholesale Trade	6,024,600	4.3	20%
Wholesale Trade Agents and Brokers	443,730	5.3	26%
Merchant Wholesalers, Durable Goods	3,392,640	4.5	21%
Merchant Wholesalers, Nondurable Goods	2,188,230	4.0	16%

Show how the comparison is builtHide how the comparison is built

Worker counts

From the Bureau of Labor Statistics’ Occupational Employment and Wage Statistics, May 2025. BLS surveys employers and publishes how many people work in each occupation, for the whole country (155.5 million workers) and for each industry. Talvio does not estimate which jobs an industry has; it uses these published counts. An industry’s average is each occupation’s score weighted by how many people in that industry hold it: 600 nurses at 3.1 and 400 billing clerks at 7.2 average 4.7.

Matching BLS occupations to scores

Each BLS occupation takes the score of the matching O*NET occupation. Where O*NET splits a BLS occupation into specialties, the general occupation’s score is used: registered nurses count at 3.1, not at the average of the nurse specialties. Where BLS merges several O*NET occupations into one, it takes their simple average.

What is left out

96% of US workers hold a scored occupation. The largest groups without a score: Sales Representatives of Services, Except Advertising, Insurance, Financial Services, and Travel (1,256,010); Project Management Specialists (1,066,670); Financial and Investment Analysts (361,980); Production Workers, All Other (251,700); Medical Records Specialists (194,720); Office and Administrative Support Workers, All Other (192,260); Emergency Medical Technicians (180,510); Postsecondary Teachers, All Other (149,840). BLS also withholds some counts in small industries to protect employers’ privacy; those workers drop out of that industry’s average. The survey does not cover the self-employed or farms, so the agriculture average reflects forestry, logging and farm support services, not farmworkers. The counts are jobs, not people: someone with two jobs counts twice.

How these averages compare with published measures is under Checks against outside research.

Limits

The score rates how much of a job’s work current AI can help with. It does not measure whether an industry uses AI or how much it gains.
Compare industries by order, not by ratio. An industry averaging 6 is not twice as suited to AI as one averaging 3.
The share at 7 and above depends on where the line sits: 31% of US workers score 6.5 or higher, against 19% at 7 or higher, because many office jobs score just under 7.

Download every industry and occupation group (CSV)

Checks against outside research

Each number is rank agreement: 1 means the two put occupations in the same order, 0 means no relation. No published measure tests exactly what this score claims, so these are the closest comparisons.

0.858

Expert ratings894 occupations

Agrees closely with a study in which experts rated where AI cuts the time a task takes in half.

Source: Eloundou, Manning, Mishkin and Rock, Science, 2024

0.458

Where AI is really used479 occupations

Agrees moderately with where people use Claude at work. Real use agrees with the score at least as well as it agrees with the expert ratings (0.458 against 0.363).

Source: Anthropic Economic Index, March 2026 report · data

0.972

Built from the same data781 occupations

Agrees very closely with the Felten AI exposure index. That index is built from the same O*NET data, so this checks consistency and adds no independent evidence.

Source: Felten, Raj and Seamans, Strategic Management Journal, 2021 · data

Finished professional work

Barely agrees with OpenAI’s GDPval, which tests how well AI completes whole professional tasks. That measures a different thing: how well AI does the work, not how much of a job it can help with.

Source: OpenAI, GDPval, 2025

Industry averages

The industry averages under “How US jobs compare” were also built from three published measures, using the same BLS worker counts. Talvio puts industries in nearly the same order as the expert ratings and the Felten index, and in a similar order to Microsoft’s measure of Copilot use.

Industries compared	Expert ratings (Eloundou et al.)	Felten index, language-model version	Copilot use (Microsoft)
Sectors (20)	0.94	0.95	0.77
Industries (3-digit) (85)	0.92	0.94	0.78
Industries (4-digit) (246)	0.93	0.95	0.78

Averages agree more closely than single occupations do, because each one covers thousands of workers. Most of the agreement comes from a pattern every measure shares: industries with more office and desk work score higher. For single occupations, agreement with Microsoft’s measure is 0.71 (751 occupations). The Felten index is built from the same O*NET data, so its agreement checks consistency more than it confirms. Microsoft’s measure comes from how people use Copilot, which trails what AI could do in some industries.

Sources: Eloundou et al., Science, 2024 · Felten, Raj and Seamans, 2023 · Tomlinson et al., 2025

Thirteen versions of the method were tried. Eleven moved the typical occupation fewer than 20 places out of 894, the bar set before the test. The two that failed both change which abilities count as physical, and both made agreement with the expert ratings worse.

Show the detail behind these checksHide the detail

The score measures how much of an occupation's work AI can help with today. The ranking agrees closely with expert ratings of where AI helps (Eloundou et al. beta, rank agreement rho=0.858, n=894 occupations) and with records of where people really use AI (Anthropic Economic Index, rho=0.458, n=479 occupations, March 2026 release). The agreement with real use is moderate and covers only the occupations that have usage records; even so, real use agrees with the score at least as well as it agrees with the expert ratings (rho=0.458 vs 0.363 on the same occupations). The claim is that the score lines up with where AI is used, not that it predicts use closely.

The expert ratings are Eloundou et al. beta, the share of an occupation’s tasks where AI cuts the time in half: every scored occupation, 95% confidence interval [0.839, 0.873]. The usage records are the Anthropic Economic Index release of March 2026.

The Anthropic Economic Index does not come from O*NET: it records where people really use AI, and it agrees with the score at least as well as it agrees with the expert ratings (rho=0.458 vs 0.363 on the same occupations). That answers the worry that the score only agrees with things built from the same data.

The score barely agrees with GDPval because the two measure different things: the score agrees closely with expert ratings of where AI helps (rho=0.858, n=894) and moderately with real use (rho=0.458, n=479), while GDPval measures how well AI does finished professional work. That is the expected pattern for a measure of where AI can help.

The claim is deliberately narrow: on the occupations with usage records, the score lines up with real AI use at least as well as the expert ratings do (score vs use rho=0.458; expert ratings vs use rho=0.363). It does not claim to predict AI use in any particular occupation or organization.

Read the external validation paper

For readers who want the full comparisons, the limits, the references, and the detail behind the checks.

Physical work

Talvio also shows how much of each job is physical. It is measured separately and never added to the Training Priority score.

How much of the workday is physical

From O*NET data on how much of the time a job is spent sitting, and how often it is spent in a vehicle.

What kinds of physical work

Every O*NET task statement for the job is labeled as one of eight kinds of physical work, such as hands-on care of people or precise handwork, or as not physical.

How far machines have come

Each kind is marked today, testing, or not yet. Machines that do this work are still rare in most workplaces, and a single physical score did no better than the plain share of physical work in testing, so Talvio shows the share and the kinds, not a score.

The share of time not seated agrees with the BLS Occupational Requirements Survey at 0.93 across 437 occupations, and with an expert’s blind forecast at 0.92 across 33 occupations.

Show the physical work method in fullHide the physical work method

The Training Priority score describes the desk side of a job: the work AI tools can help a trained person with. Most workforces also hold a lot of work done on your feet, with your hands, or behind a wheel. The physical work layer shows how much of each role that is, and what kinds of physical work it holds, beside the Training Priority score and separate from it. Two rules hold everywhere it appears. The physical numbers are never added to, averaged with, or ranked against the Training Priority score. And no single physical score is shown, because the evidence supports the breakdown, not one number.

Every number comes from public O*NET 30.3 records for the matched occupation, and the whole calculation is arithmetic. No AI model runs when a number is produced. Physical work method: September 2026.

1. The share of work time that is physical From O*NET Work Context, on its 1 to 5 scale: how much of the time the job is spent sitting, and how often it is spent in a vehicle. The overall figure is the larger of time not seated and time in a vehicle, rounded to the nearest 5 and read as “about.” O*NET records how often the job is in a vehicle, not how many hours, so the two parts are always shown separately.

notSeated_o = 5 − sit_o4 inVehicle_o = veh_o − 14 headline_o = 100 · max(notSeated_o, inVehicle_o)

2. Kinds of physical work Every one of the 16,830 O*NET task statements carries one label from a fixed rubric: one of eight kinds of physical work, or not physical. Labels came from exact word rules plus three independent readings of the rubric, settled by majority. On a held-out sample, 95% of the tasks labeled physical were physical, the labels found 94% of the physical tasks, and they named the exact kind right for 85% of physical tasks. A role’s share for a kind is the importance-weighted share of its physical tasks with that label.

kind_o,k = ∑_{t ∈ k} IM_o,t ∑_{t physical} IM_o,t

3. Three stages Each kind sits at one stage, from ratings of what machines can do today by five AI raters working blind, approved by a person and refreshed twice a year. The evidence file includes machines that were tried and abandoned. The percentages beside the stages are shares of the physical work, that is, the kinds at that stage added together, and the bar is split in the same proportions.

stage_o,s = ∑_{k ∈ s} kind_o,k bar_o,s = headline_o · stage_o,s

Below a headline of 20 the breakdown is replaced by “Little physical work in this job.” An occupation with no physical task in its O*NET task list (180 of 894) shows the bar without a breakdown. O*NET task lists rarely mention walking or standing, so that time counts in the overall figure and does not appear as a kind of physical work.

The kinds of physical work and their stage

Kind of physical work	Stage today
hands-on work with people	Not yet
repair and troubleshooting	Not yet
routine upkeep and cleaning	Testing
running fixed machines	Today
lifting, carrying, and moving objects	Testing
moving around on foot	Testing
precise handwork	Testing
driving and operating equipment	Testing

Today machines do this in fixed, controlled settings today

Testing machines do some of this in controlled settings; not dependable in open ones

Not yet no machine does this on its own yet

A machine doing this in a factory does not mean one could do it in your building. These numbers describe the work, not your building, and they never tell anyone to buy or try robots.

How much to trust it

Two checks were written down before any number was computed and then run once.

rho=0.92

Blind test against an expert An expert wrote down what to expect for 40 randomly drawn occupations, seeing only the job titles. The file was locked before any number existed. On the 33 that have time-use data: rank agreement 0.92 against a bar of 0.75; 22 of 33 inside the expert’s range (67%, against a bar of 60%); typical error 10.2 points; the biggest kind matched in 56% of cases, and was in our top two in 75%.

rho=0.93

Outside check: BLS Occupational Requirements Survey A survey of employers, run separately from O*NET. Our time-not-seated share against the survey’s percent of the day spent standing: rank agreement 0.93 across 437 occupations. At the middle of the range O*NET reads 4 points lower. Where the two disagree most, O*NET reads low for hands-on trades, so our top of the scale is squeezed even though the order agrees closely.

±15

What we promise on screen Usually within 15 points of an expert's estimate. Two known limits, and how each is handled: the vehicle part counts jobs that ride in a vehicle every day even when driving is not the job, so every breakdown shows both parts as numbers; and walking and standing are under-counted in the list of kinds, so every list says so.

One rule in the blind test was never exercised: every role the expert called mostly hands-on care should show at least half its physical tasks as “not yet.” No such role came up in the random draw, so that rule is recorded as untested. On the roles used while choosing the method, nursing assistants show 77% hands-on work with people, home health aides 84%, and registered nurses 78%.

We built a physical score and decided not to show it

A 0.5 to 10 physical score was built and tested against two outside references. It agreed with the things it should agree with, differed from the things it should differ from, and did not simply repeat its own inputs. It still failed the test we had set in advance: it was no better than a plain share of physical work at predicting either reference (it was worse by 0.087 and 0.090), and versions built with random weights did just as well. Two blind reviews of the method agreed the extra structure added little. The score stays in the record, with this written down, and appears nowhere in the product.

The full method

The formulas are under “Show the math” above. These are the rest of the detail. Method version TPS v2.2, July 2026. Physical work method, September 2026.

How good AI is at each skillThe twelve AI skills, their ratings, and the public test behind each one.

These reviewed scores shape what to teach. They answer the question “which AI skills matter for this work?” The Training Priority ranking itself is built from the work activities, not from these scores.

AI skill	How good AI is at it today	Measured by
Written content generation and editing	7.5/10	HELM Capabilities
Information synthesis and research	6.0/10	Artificial Analysis Intelligence Index
Structured data analysis and quantitative reasoning	5.5/10	Artificial Analysis Intelligence Benchmarking
Coding and software engineering	6.5/10	SWE-bench
Conversational support and customer interaction	6.5/10	Arena Text Leaderboard
Translation and cross-language work	7.0/10	Artificial Analysis Multilingual Index
Speech and audio processing	6.5/10	Artificial Analysis Speech to Text Leaderboard
Image and document understanding	7.5/10	MMMU-Pro
Image, video, and design generation	7.0/10	Artificial Analysis Image Model Leaderboard
Planning, scheduling, and structured decision support	5.0/10	GDPval
Tool use and autonomous agents	4.0/10	METR Time Horizons
Domain-specialist reasoning	6.5/10	GPQA

Occupations we do not scoreThe 122 left out, grouped by reason.

122 occupations are left out on purpose: catch-all codes, codes that O*NET splits, and military jobs. They stay visible as excluded rows with a reason, rather than being scored quietly.

The reasons are A catch-all codes, B codes O*NET splits into others, and C military jobs. Filling the gaps from similar occupations was tested and is not used.

894

Occupations we score Occupations that have O*NET 30.2 Work Activities data.

A residual Left out because these catch-all occupation codes have no work activities of their own in O*NET.

B split-code Left out because O*NET 30.2 has no Work Activities data for this detailed code.

C military Out of scope by design for Talvio's civilian occupation scope; not a coverage failure.

Read the external validation paper

How Talvio decides which jobs gain most from AI training

The short version

Built from public data

Lowered for physical work

A place to start training

Read it by group

How the score is built

Built once for every occupation

What each job involves

How much AI can help with each activity

Add it up for the job

Lower it for physical work

Put it on the 0.5 to 10 scale

Applied to your roster at each upload

Match each job title to an occupation

Add up by department

One job, start to finish: registered nurses

What moves a score most

What the AI skill ratings do

How US jobs compare

Average score by industry

Average score by occupation group

Industries in more detail

Worker counts

Matching BLS occupations to scores

What is left out

Limits

Checks against outside research

Industry averages

Limits

Physical work

How much of the workday is physical

What kinds of physical work

How far machines have come

The kinds of physical work and their stage

How much to trust it

We built a physical score and decided not to show it

The full method