What does a worker
add, after the costs?
Useful output, minus compensation, coordination, and the value of available alternatives. Explore the range within a job—and what AI changes.
Software engineers
+0.28× net compensation
Modeled share with negative net value
Net value / annual compensation
Change the assumptions Explore sensitivity ↗
Why this shape
Unused features, rework and review burden create a negative tail; reusable systems create a long positive tail. The median is not empirically known.
What changes
Coding studies disagree across settings. Late-2026 agents expand possible output, but review, product judgment and demand remain constraints.
Evidence for this occupation 9 sources ↗
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Employment and annual mean wages; the 100 largest non-overlapping leaf occupations in this table define the sample.
Payroll employment, not the whole workforce. Excludes self-employed workers and most agricultural and household employment. Annualized wages are not actual annual earnings for part-time workers.
Task exposure is highest in clerical work. Physical and interpersonal tasks limit whole-job substitution.
Exposure is potential technical capability, not realized replacement or a measure of worker value. Our occupation exposure ratings are judgments, not copied ILO scores.
Exposure measures must be combined with adoption, feasibility and actual employment evidence.
An exposed task does not imply that a job is unnecessary or economically unproductive.
Across 4,867 developers, access to a coding assistant increased completed tasks by about 26%, with heterogeneous gains.
Task counts are not product impact, profits or current agentic capability. Industry-affiliated study.
Sixteen experienced developers completing 246 tasks took 19% longer with early-2025 AI tools.
Small sample in familiar mature repositories. Not a claim about all developers or late-2026 models.
Later results suggest more speedup, but selection and time-measurement problems prevent a reliable current estimate.
The authors explicitly warn against treating the follow-up as a clean productivity estimate.
Internal data associate agent adoption with more code, sales and contracts per person, while human attention becomes a bottleneck.
One unusually AI-intensive company; observational, vendor-authored evidence. Does not establish 10× or 100× worker-value prevalence.
An instrumented event study finds a lower junior employment share after adoption, mainly through senior employment growth.
Working paper; employment composition is not net output, and causal interpretation depends on the design assumptions.
Inspect this job’s parameters All are assumptions ↗
Official BLS occupation: Software developers. Parameters below are priors, not values estimated from wages or fitted to a worker dataset.
- Gross median
- 1.4×
- Spread parameter
- 1.35×
- Gross P99
- 10×
- Coordination cost
- 0.3×
- AI output gain
- 25%
- Alternative discount
- 0.15×
At this job’s average wage and a generic 1.43× benefits factor, one compensation unit is approximately $211,571. Wage data are context, not evidence of contribution. We do not estimate an individual worker’s pay.
The assumptions are part of the graphic.
A narrow definition of “utility”
The model asks what a worker adds for an employer or service recipient over one year, net of compensation, coordination burden, tooling, and a discount for available alternatives. Output means attributable value after nonlabor inputs, not sales revenue. Capital and team effects cannot be perfectly separated.
Salary is income to someone else, so this is not total social welfare. Health, dignity, long-run student outcomes, environmental effects and redistribution are not comprehensively priced. These cross-job comparisons depend on contestable valuations.
Research informs the shape, not the numbers
BLS establishes job prevalence, pay and duties. Experiments inform plausible AI mechanisms; observational 2026 studies add current context. The gross-output medians, spread, tails, coordination costs and AI coefficients are explicit editorial priors. No evidence establishes that the typical software engineer has zero net value.
The unit is a worker in a workplace, including role design, resources, demand and organizational fit—not fixed individual talent. Unused output is assumed to have no value.
One common scale
Net = gross contribution × output scenario × (1 + AI gain) − 1 compensation unit − coordination − alternative discount − AI tooling. Coordination does not include the worker’s own pay. The alternative discount is a sensitivity assumption, not a measured counterfactual replacement cost.
Index = 50 + 50 × asinh(4 × net) / asinh(400), clipped to 0–100. Zero net maps to 50. ±100 compensation units map to the endpoints. The index is nonlinear; a score of 80 is not twice a score of 40.
Population spread ≠ certainty
Ten quantile anchors define each assumed distribution; linear interpolation and 10,000 deterministic integration points produce the bars. Those points are not observed workers. Tails above +100× remain in the final bin and in percentile statistics.
The squeeze scenario uses 0.8× baseline output, half the AI gain and 1.6× the alternative discount. Upside uses 1.15× output, 1.6× the gain and half the discount. None is a forecast or confidence interval. Use comparisons to examine assumptions, not to rank human worth.
Research bibliography & coverage 17 studies and data releases + 17 BLS occupation groups ↗
The 100 largest leaf rows in the May 2025 OEWS national table are sorted by employment, with aggregates removed. They represent 106,118,330 payroll jobs. Data were released May 15, 2026; research was checked through November 16, 2026. This is a U.S. model, not a global estimate.
OEWS wages exclude benefits and stock compensation. Most annual wages are hourly rates × 2,080; for part-time jobs they are full-time equivalents. We retain the data’s date rather than inventing 2026 occupation counts or wage estimates.
Employment and annual mean wages; the 100 largest non-overlapping leaf occupations in this table define the sample.
Payroll employment, not the whole workforce. Excludes self-employed workers and most agricultural and household employment. Annualized wages are not actual annual earnings for part-time workers.
Private-industry wages are 70% of compensation; compensation / wages is approximately 1.43.
A national average, not an occupation-specific benefit rate; government and part-time benefit costs differ. Used only for the illustrative dollar conversion.
Task exposure is highest in clerical work. Physical and interpersonal tasks limit whole-job substitution.
Exposure is potential technical capability, not realized replacement or a measure of worker value. Our occupation exposure ratings are judgments, not copied ILO scores.
Exposure measures must be combined with adoption, feasibility and actual employment evidence.
An exposed task does not imply that a job is unnecessary or economically unproductive.
Observed use is much smaller than theoretical capability; no systematic unemployment rise was detected in highly exposed occupations in the study window.
Vendor-specific observational usage; neither a representative worker sample nor a causal productivity estimate.
Internal data associate agent adoption with more code, sales and contracts per person, while human attention becomes a bottleneck.
One unusually AI-intensive company; observational, vendor-authored evidence. Does not establish 10× or 100× worker-value prevalence.
An instrumented event study finds a lower junior employment share after adoption, mainly through senior employment growth.
Working paper; employment composition is not net output, and causal interpretation depends on the design assumptions.
Across 4,867 developers, access to a coding assistant increased completed tasks by about 26%, with heterogeneous gains.
Task counts are not product impact, profits or current agentic capability. Industry-affiliated study.
Sixteen experienced developers completing 246 tasks took 19% longer with early-2025 AI tools.
Small sample in familiar mature repositories. Not a claim about all developers or late-2026 models.
Later results suggest more speedup, but selection and time-measurement problems prevent a reliable current estimate.
The authors explicitly warn against treating the follow-up as a clean productivity estimate.
In the working-paper sample of 5,179 support agents, issues resolved per hour rose 14%; novice and lower-skill workers gained more.
One firm and an older assistant. The published version and sample differ; this page uses the cited working-paper result. Not an absolute net-value estimate.
In a 758-consultant experiment, AI helped on tasks within its capabilities and reduced accuracy on an out-of-frontier task.
Task-level treatment effects; applying the mechanism to other knowledge occupations is an extrapolation.
A randomized study across 66 firms and 7,137 workers found email time savings, but no detectable broad shift in task quantity or composition.
Time saved is not automatically economic surplus; organizations must put it to productive use.
Survey and administrative evidence shows modest time savings and little detectable change in earnings or recorded hours in the initial study period.
Different labor market and early models. A useful counterweight to extrapolating frontier-company results.
In one service company, supervisor quality materially changed team output and worker retention.
Supports including team spillovers, not any particular managerial median or a universal multiplier.
Teacher quality predicts later student outcomes, including earnings; benefits can accrue long after the work.
Does not identify total value relative to no teacher or a comparable 2026 annual employer surplus. Long-run benefits are not added to this narrow model.
Nurse staffing and team-specific experience affect patient outcomes in VA hospitals.
Supports team and safety mechanisms, not the numerical value distributions of all healthcare workers.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.
Occupational duties, physical requirements and work setting inform the qualitative task assessment.
Describes work. Does not measure net utility or validate our model parameters.