What is GDPval, and why do real occupational tasks matter for AI?

GDPval in plain language

GDPval is an evaluation that asks AI models to do the kind of work professionals are paid for, then has experienced practitioners judge the results. A task hands the model a request plus the reference files a person in that job would receive, such as last year's figures in a spreadsheet or a set of meeting notes, and asks for a finished deliverable: a memo, a slide deck, a schedule, a forecast.

The name points to the design. Occupations were drawn from the sectors that contribute most to US gross domestic product, with a focus on roles whose work happens mainly on a computer. GDPval was released publicly in September 2025 with a research paper and an openly published sample of tasks. This page describes the design in general terms and gives no task counts or scores; read those, and confirm the design details, in the primary release, because they change as new models are tested.

For anyone who works with established companies, one point stands out: the yardstick is real work product, not exam questions.

How does GDPval work?

Each task follows a request, deliverable and review pattern that anyone who has run a professional practice will recognize.

How is GDPval different from earlier AI benchmarks?

Earlier benchmarks mostly asked models to choose or produce a correct answer. GDPval asks whether the output would be accepted by someone who does the job for a living.

The coding row is covered in how AI coding agents are trained, and the multi-day gap in why agents need records of multi-week projects.

Why do real occupational tasks matter for AI?

Building a task like this takes knowledge an AI developer cannot easily produce alone. Someone has to know what a good coverage summary, structural calculation note or month-end variance commentary looks like, and what separates it from a weak one. That knowledge lives inside the firms that produce those deliverables every week.

Evaluations built on genuine work products therefore pull attention toward company records. An established professional-services firm, engineering practice or operations team holds thousands of examples of the pattern GDPval uses: a request, the files used to answer it, the deliverable, and a reviewer's verdict in the form of redlines, approvals or a client's reply. Those records also show what happened next, which a single task cannot.

Simulated tasks can approximate this, with trade-offs set out in synthetic environments versus real business logs. Recency counts as well, because methods, rules and tools change; data freshness is part of the same picture.

What GDPval means for a referral partner

GDPval is a public research benchmark. SourceX has no role in it, and nothing here suggests that its authors license data from any company. For a partner, its value is as a clear public illustration of the evidence AI developers use to judge their models.

Use it as a lens on the companies you know. The request, deliverable, review test:

• Request: does the company keep the original asks, such as client briefs, statements of work, tickets or internal requests, and not just final files?
• Deliverable: are finished work products stored with dates and versions in shared drives, document systems or project tools?
• Review: can you see what reviewers said, through redlines, approval steps, QA scorecards or client sign-off?
• Outcome: is there a record of what followed, such as a renewal, a change order, a claim paid or a project closed?
• Baseline: is it a US company with 50+ full-time employees at peak (contractors excluded), several years of documented operations, the right to license what it created, and an owner, CEO or CFO who can approve?

A company that ticks most boxes is worth running through the company fit checker, a preliminary, non-binding screen. If it looks promising, you make the introduction and the company works with SourceX on qualification, its data inventory, pricing and buyer review, as described in how it works. You never export, upload or describe the company's confidential files.

If a deal closes, partners earn 25% of the eligible platform fees SourceX actually collects from the referred company's licensing deals, capped at $100,000 per referred company, and the reward becomes payable only after the buyer pays and SourceX receives its fee. No reward is guaranteed.

What GDPval does not tell you

Treat GDPval as one well-designed measurement, not as a verdict on jobs or on what any company's records are worth.

• Each task is a single request and a single deliverable. Real work involves clarifying questions, drafts, feedback and handoffs across days, which a one-shot task cannot capture.
• Expert preference is a judgment. Two reviewers in the same occupation can disagree, especially on style.
• It covers computer-based work. Physical, on-site and relationship-heavy parts of a role sit outside its scope.
• Doing well on tasks drawn from a role is not the same as automating that role.
• Benchmarks age, and once tasks are public they can leak into training data. Annual reviews such as Stanford HAI's AI Index, which covers technical performance alongside investment, adoption and policy, help track how benchmark results are read over time.

Next step

If you know a firm whose archive is full of requests, deliverables and reviews, register as a partner and introduce it, or send the owner to apply directly at sourcex.si/apply using your referral link.