Applied Frontier

Research agenda · Last updated November 2026

The six questions a country must answer to build an AI-era workforce

Direct answer

Applied Frontier’s research agenda is organized around six open questions: which job families the AI economy will need, how they will be priced, what replaces the degree as a signal, how to read individual contribution through jargon, which skills to train and stop training, and whether the Capital × Labor productivity model still holds. For each, we state what is known, what is not, and what we are doing about it.

Which job families will be needed in the next three and six years?

What we know. Demand is moving away from producing analysis, code and prose and toward specifying, verifying, deploying and owning outcomes from AI systems.

What we do not know. The names, the volumes, or the stability of those roles beyond an 18-month horizon.

Our approach. Build a job-family map on two horizons, 2029 and 2032, sourced from employer hiring plans rather than job titles, and update it quarterly as the models change.

How will those jobs be priced in local and global labor markets?

What we know. The premium for commoditized cognitive skills has collapsed, and the premium for judgment, accountability and domain context has not yet been set.

What we do not know. Whether AI-era jobs will price globally, as software did, or locally, as trades do, and what the spread between the two will be.

Our approach. Track compensation for the emerging job families across at least five labor markets, and publish the spread.

If universities can no longer produce an interpretable signal, what will employers and workers use?

What we know. The type of work and the type of employer already act as a proxy for transferable skills; a candidate from a known firm carries a readable signal that a degree no longer does.

What we do not know. How to filter an individual’s real contribution out of the noise and jargon people use to describe it, which is a non-trivial measurement problem.

Our approach. Test whether verified work artifacts, structured references and short demonstrated tasks produce a signal employers trust more than a transcript.

Which skills should we train, and which should we stop training?

What we know. Every hour spent training a skill a model already performs at expert level is an hour and a dollar spent on a depreciating asset.

What we do not know. Where the line sits today, and how fast it moves.

Our approach. Maintain two public lists, revised as capabilities change. Version 1 is published as The Two Lists.

Is Capital × Labor = Productivity still valid?

What we know. The classic production function treats capital and labor as complements; AI is capital that substitutes for cognitive labor directly.

What we do not know. The right functional form when a unit of capital can produce the output of a knowledge worker, and what that implies for wages, tuition and national education budgets.

Our approach. An empirical version of the equation using firm-level data, with an explicit depreciation rate on human skills tied to model capability.

What does a US college degree actually cost and return in 2026–27?

What we know. The US spends on the order of several hundred billion dollars a year on higher education and produces roughly four million degrees a year (approximate).

What we do not know. What share of the 2030 graduating class the AI economy will absorb, and how long a $200,000 degree takes to reach $200,000 in earnings. No university will say.

Our approach. Put a number on each and compare cost per graduate with China, so that countries designing new systems know the benchmark they have to beat.

How the questions get answered

Not with reports alone. Each question becomes a hypothesis, a targeted pilot in a real market, and a measured result that is published with its sources and revised when the models change.

Questions about the agenda

Who is answering these questions today?
Nobody with a credible, published, regularly updated answer. That is the gap Applied Frontier is built to fill.
How often is the research updated?
Quarterly at minimum, because the underlying models change at that pace and any job-family map older than a year is already wrong.
Can I contribute data or a case?
Yes. Employers, universities, ministries and multilateral agencies with hiring, enrollment or outcome data are invited to work with us.