Job postings that mention AI skills advertise salaries 28 percent higher — nearly $18,000 a year — than otherwise comparable ones (Lightcast, 2025). When economists put real hiring decision-makers in front of real trade-offs and made them choose, prompting skill was worth 2 percent above the average salary of a skilled worker in their firm (IZA, 2026).
Both numbers are correctly measured. They are measuring different things, and the gap between them is where mid-market screening criteria are quietly going wrong.
The second number comes from a discrete choice experiment run inside the 2025 wave of Germany's BIBB Training Panel: 992 firms, 15,660 forced choices between applicant profiles, roughly 81 percent of the sample under 250 employees. And the finding that should change a Q4 req is not the price. It is what the price refused to buy.
What 992 Firms Chose When Forced to Trade
The design is the reason to take this seriously. Respondents were owners, managing directors, branch managers and HR managers — people who actually sign off on hires — each shown pairs of applicant profiles that varied on five attributes: prompting skills, occupation-specific skills gap, social skills, gender, and salary expectation relative to a skilled worker in their own firm. Education and relevant experience were held constant. They picked one, repeatedly.
That structure converts preferences into a currency. Here is the full price list (IZA, 2026):
| Attribute | Effect on hiring probability | Willingness to pay |
|---|---|---|
| Intermediate prompting skills | +4.3 points | +2.0% |
| High prompting skills | +5.2 points | +2.5% |
| Intermediate social skills | +19.9 points | +9.0% |
| High social skills | +33.8 points | +14.6% |
| Considerable occupation-specific skills gap | −31.5 points | −13.6% |
Every attribute was significant at the one percent level. High social skills were worth roughly six times what high prompting skill was worth. A considerable gap in occupation-specific skills cost about what high social skills earned, in the opposite direction.
One more line from the same model makes the magnitude concrete. An applicant asking for 3 percent above the firm's average skilled-worker salary lost 4.8 points of hiring probability; asking for 6 percent above cost 13.8 points. High prompting skill gains 5.2 points. Prompting fluency buys back the hiring penalty of asking for about 2.5 percent more money, and then it is spent.
The Postings Premium Is Not the Hiring Premium
It is tempting to read the Lightcast figure and the IZA figure as a contradiction. They are not, and understanding why is the difference between a defensible req and an expensive one.
Lightcast analysed 1.3 billion job postings and compared advertised salaries on postings that list AI skills against postings that do not (Lightcast, 2025). That is a between-role comparison. Postings that ask for AI skills are systematically different postings — more senior, more technical, more concentrated in higher-paying functions and firms. The 28 percent is real, and much of it is the price of the job, not the price of the skill.
The IZA experiment holds the job constant and moves only the candidate. That is a within-role comparison, and within the role, the marginal value of prompting fluency is about 2 percent.
If you benchmark a req against posting-level data, you will conclude that AI fluency commands a large premium and build your screen accordingly — weighting it heavily, paying up for it, trading other things away to get it. The revealed preference of 992 hiring decision-makers says the market is paying a small premium for it inside a given role, and that it is not substituting for anything.
Prompting Skill Did Not Buy Its Way Past a Skills Gap
This is the result that carries the argument, and it survives in the direction nobody wants.
The authors tested explicitly for a compensatory effect — whether strong prompting skills could offset an occupation-specific skills gap. They could not. High prompting skills showed no significant interaction with either a slight or a considerable skills gap. Worse, the interaction between intermediate prompting skills and a slight skills gap was marginally significant and negative: about 3 percent lower hiring probability than an applicant with low prompting skills and a full skills match. Intermediate prompting fluency made a small domain gap read worse, not better.
The authors' conclusion is the sentence to carry into a hiring meeting: the demand for prompting skills "seems to upgrade skills requirements in jobs" rather than substitute for them (IZA, 2026).
AI fluency is an added condition. It is not a trade.
Where It Did Have Leverage
One interaction was positive. Applicants with intermediate prompting and intermediate social skills were about 4 percent more likely to be hired than the low-low reference group. High social skills showed no significant interaction with prompting at any level — there is nothing to add to a candidate who already has the thing employers pay 14.6 percent for.
So prompting fluency has leverage in exactly one place: helping a candidate with low-to-intermediate interpersonal capability. It has none where it is most often invoked — covering for thin domain depth.
That direction matters because the human half of the requirement is getting heavier, not lighter. PwC's analysis of over a billion job ads found that skills demanded in the most AI-exposed roles are changing more than twice as fast as in the least exposed, and that newly emerging tasks are 2.5 times more likely to rely on empathy, judgement and creativity (PwC, 2026).
Where the Premium Doubles
The 2 percent is an average, and the heterogeneity analysis is operationally useful.
In firms with more than 250 employees, high prompting skills raised hiring probability by 9 percent — double the pooled effect, significant at the five percent level. On a median split of headcount, above-median firms were 7 percent more likely to prefer the high-prompting applicant. In SMEs, the effect stayed close to the pooled average of about 4 percent.
The same pattern appears by AI adoption. Firms that have adopted or plan to adopt AI in physical or non-physical work processes preferred high-prompting applicants by roughly 9 percent, against 2 to 4 percent among non-adopters. Notably, among firms that had adopted generative AI specifically, the interaction was only marginally significant — the premium tracks firms industrialising AI into processes more than firms whose staff have ChatGPT open.
Read that as a competitive signal rather than a pricing signal. If you are a 300-person firm competing for the same technician against a 3,000-person firm mid-rollout, the larger firm values prompting fluency at roughly twice what you do. It will outbid you on that attribute. It will not outbid you on social skills or domain match, because everyone prices those high.
The Honest Counter
Three limits, stated plainly.
A different experiment found the opposite kind of compensation. Stephany, Teutloff and Leone ran a paired conjoint with 1,725 recruiters across the UK, US and Germany and found AI skills raised interview-invitation probability by 8 to 15 percentage points, and that they partially or fully offset disadvantages of older age and lower formal education (Stephany et al., 2026). That is a real tension — but note the margins. Their outcome is an interview invitation, where a signal is cheap to act on; IZA's is a hire with a salary attached. And offsetting a demographic disadvantage is not the same as offsetting a capability gap. Both can be true: prompt fluency gets a 55-year-old office assistant into the room, and still does not cover a missing qualification once someone has to pay for it.
Stated preference is not a payroll record. Nobody audited what these 992 firms actually paid. The estimates are choices in a survey instrument, not transactions — which is the standard trade for getting clean causal variation.
It is German, vocational and early. The authors flag their findings as time-specific to a phase of early but rapidly increasing GenAI diffusion. This is a labour market anchored in formal vocational qualification, where "occupation-specific skills" is an unusually crisp category. Do not port the point estimates to US hiring without saying so. The ordering of the attributes is what travels; the decimals are not.
What to Change in the Screen Before Q4 Reqs Close
Four moves, none of which requires a new budget line.
- Stop letting AI fluency buy a discount on domain depth. If your shortlist logic lets a strong prompting signal pull a candidate past a requirement you marked as critical, you have built a compensatory rule that 992 hiring decision-makers declined to build. Mark the blocker requirements and make them block.
- Reprice the AI-skills line in the req. If you were anchoring on a double-digit posting premium, the within-role evidence says 2 to 2.5 percent. Spend the difference on the attribute worth 14.6 percent.
- Move prompting from "required" to "assessed". It is cheap to teach relative to domain expertise and social capability, and the distinction between intermediate and high mattered little in the pooled sample. Screen for the floor; do not pay for the ceiling.
- Make the weighting explicit and inspectable. The useful mechanic is criticality weighting — a blocker requirement counting several times an important one and an order of magnitude more than a nice-to-have — so that a strong score on a cheap attribute cannot arithmetically outvote a missing essential one. Scovai's scoring model is documented publicly if you want a worked reference for how that arithmetic is structured.
The deeper point is about where requirements are heading. Prompting skill did not loosen the job spec. It added a line to it, while the lines that were already there got more expensive.
Before your next req goes out, find the clause where AI fluency is listed as an alternative to something. Delete the word "or".
