AI Skill Premiums Engine.
The dollar premium each AI skill commands on top of base comp, by role and by city. 411 skills. Quarterly refits. 95% confidence intervals on every estimate.
What skill premiums are.
A skill premium is the extra dollar amount a worker earns for holding a specific skill, controlling for role, seniority, city, and company stage. The premium is what is left after you subtract everything else the market is paying for. If two Senior ML Engineers in San Francisco are otherwise comparable but one knows CUDA kernel optimization at production depth, the CUDA premium is the difference in their total compensation.
Why this matters: comp negotiation rarely happens at the headline number. It happens at the skill level. The candidate says, "I bring distributed LLM training and CUDA kernel optimization," and the offer team mentally calibrates against the marginal value of those skills. Without a quantified premium, the conversation is anchored by feel. With one, both sides have a defensible number.
Skill premiums also matter for comp design. When a company is building a band for role:senior-ai-engineer, the band has to account for the fact that two candidates with the same title can earn $80K apart based on the skill bundle they bring. The premium catalog lets comp teams build bands as base + skill bundle rather than as a flat range that has to absorb all the variance internally.
For agents, skill premiums are the explanation layer. When analyze_compensation returns a salary estimate, the agent can explain what is driving the number: the role contributes X, the city contributes Y, the funding stage contributes Z, and these four skills contribute W. That explanation is what turns a number into a decision-ready answer.
Methodology.
We fit a mixed-effects regression on every quarter's comp dataset. The model is structured so the coefficient on each skill indicator can be interpreted as the marginal dollar contribution of that skill, holding everything else constant.
Each βᵢis one skill's premium. The fixed effects on role, seniority, city, and funding stage absorb the structural variance: we are not double-counting a city premium as a skill premium. The random effects on company and year absorb the company-specific and time-specific noise so the skill coefficients reflect cross-market averages, not a single company's pay philosophy.
The dataset has three layers. The base layer is H-1B LCA filings, which carry employer, role, wage, and city: but no skill tags. We join skill tags from the job-posting layer, where postings explicitly list required and preferred skills. The community submissions layer carries self-reported skill bundles paired with offer numbers. The three layers triangulate so a skill is only kept in the catalog when it appears with consistent directionality across at least two layers.
The model is refit quarterly. The methodology version increments on every refit (2026.1 → 2026.2, etc.). The previous version's coefficients remain available via ?as_of=2026-Q1 so customers can reproduce historical answers and audit the rate of change.
Top current premiums.
Sampled from the May 2026 refit. Premiums are quoted as the additional total comp (base + equity) the skill adds, with the 95% confidence interval. Live data lives in the API.
Sample shown. Full catalog of 411 skills available via GET /api/v1/intelligence/salaries/skills. Methodology version 2026.2.
How buyers use it.
The most common single-customer pattern is the negotiation builder. The candidate lists the four to seven skills they bring, multiplies each by the role-specific and city-specific premium, and sums into a defensible floor over base. Because every premium ships with a confidence interval, candidates can quote both the point estimate and the range when they negotiate, signaling both numeracy and honesty.
Comp teams replace flat ranges with skill-bundle bands. A Senior ML Engineer at the 75th percentile becomes "base 75th + CUDA bundle + distributed-training bundle," giving comp teams the granularity to pay for what the candidate brings without artificially expanding the headline range. The bands are easier to defend in committee.
Recruiters write JDs that match the comp band they can pay. When a role's posted range is $220K-$280K, the recruiter knows what skill bundle the comp band assumes, and the JD lists those skills accurately. Mismatch between posted skills and actual offer becomes the leading indicator we use to flag JDs that will under-pay candidates.
Engineers and learning teams use premiums to sequence curriculum. If CUDA optimization carries a $58K premium but takes 200 hours of focused practice to reach production competence, the engineer can compute the ROI per hour. Combined with the Skill Half-Life engine, the calculation also accounts for how long the premium will hold before decay.
Confidence intervals and sample sizes.
Every premium ships with three transparency fields: sample_size (number of distinct comp observations that informed this estimate), confidence_low / confidence_high (the 95% CI), and confidence_level (which is currently always 0.95: we will expose customer-tunable CIs in 2026.4).
Wider CIs mean less data or more variance. CUDA kernel optimization has tens of thousands of observations across three years and a tight ±$8K window. A new skill like llm-judge-eval has only a few thousand observations and a wider ±$15K window. We expose both numbers so customers know exactly how much weight to put on each.
Premiums whose 95% CI crosses zero are flagged with disagreement_flag: true. That means the regression cannot confidently distinguish the premium from zero: the skill may not yet have separated from noise. The MCP analyze_compensationtool surfaces this flag in the Decision-Ready Response's limitations array so agents can warn users about low-confidence premiums.
Beyond CIs, the /lineage/[data_point_id] endpoint returns the full provenance trail for any premium: which observations went into the regression, which sources they came from, how the residuals are distributed. That is Bloomberg-grade data provenance applied to skill-level comp.
API integration.
Three integration paths. All three return the same envelope.
See the full reference at /orbyt-intelligence/api-docs. The MCP server's six tools are documented at /orbyt-intelligence/mcp.
Pricing and access.
Skill premium queries require the Pro tier and the skills:read scope. The substrate-level Role Taxonomy Engine ships on Build ($99/mo); the regressions and refits behind premiums require Pro because they cost meaningfully more to compute and serve.
Full pricing detail at /orbyt-intelligence/pricing.
See also.
Methodology version 2026.2. Last updated May 2026. Premium estimates refit quarterly; methodology versions visible in every response. Premiums whose 95% CI crosses zero carry disagreement_flag: true.