A South African scholar looking for an academic role recently encountered a different kind of offer: train an AI system to create assessments, teach undergraduates and grade essays. The proposed work paid 600 rand, or about $37, an hour—far above South Africa’s reported national minimum wage of 30.23 rand.
The author of a first-person account published by *Rest of World* ultimately stopped pursuing the role. But the episode captures a growing operational reality for AI companies and knowledge workers alike: the valuable input is no longer just labeled data. It is expert judgment.
From annotation to professional discretion
For years, outsourced AI labor has often meant tasks such as data labeling, transcription and content review. The newer demand is more consequential. Platforms and AI vendors are recruiting specialists who can demonstrate how professionals make decisions: which concepts matter in a course, how to distinguish memorization from understanding, or why one essay deserves a higher mark than another.
That is attractive to model builders because these decisions are difficult to reduce to simple rules. They combine subject expertise, context and accumulated professional practice. The source points to similar demand for legal, clinical and engineering expertise, where systems are being trained or evaluated against practitioners’ judgments.
For employers, that changes the nature of AI implementation. A model may generate drafts or recommend an answer, but making it dependable in a specific workflow requires a clear account of what “good” looks like. Organizations will increasingly need to identify which decisions can be standardized, which need expert review, and which should remain human-owned.
The economic tension is real
The dilemma is particularly sharp in markets with high unemployment and lower professional wages. The author describes being recruited in South Africa, where youth unemployment was reported at 47.4% in the second quarter of 2026. For a skilled worker, high-paying AI-training work can be a rational short-term choice even if it may eventually make parts of that worker’s role more automatable.
This creates a potential asymmetry for global AI builders. Regions with a young, educated workforce and constrained local job markets can become sources of relatively inexpensive professional knowledge. The immediate benefit is income and exposure to frontier technology work. The longer-term risk is that local workers supply the expertise for products whose economic gains accrue elsewhere—or which reduce demand for the same work.
The issue is not unique to Africa, nor is replacement inevitable. AI systems often change jobs before eliminating them, and expert participation can improve quality and safety. Still, executives should not treat expert contributors as interchangeable annotation capacity. The data being captured may encode a profession’s methods, standards and judgment calls.
A governance issue, not just a hiring issue
One detail in the account is notable: the candidate says the 45-minute interview was conducted by an AI, which also delivered feedback. That points to a loop many companies are beginning to create—using AI to recruit people to improve AI.
That loop needs governance. Companies commissioning expert-training work should be able to answer basic questions: What capabilities is the work intended to enable? How will contributors be compensated as their expertise becomes reusable product value? What safeguards prevent a model from making high-stakes decisions without appropriate oversight? And how are contributors informed about data retention, reuse and downstream deployment?
For founders and operators, the practical takeaway is to design expert-data programs as partnerships rather than one-off labor procurement. Use clear scopes, meaningful consent, domain-appropriate evaluation and human escalation paths. Where a system will affect hiring, education, healthcare or legal outcomes, document the limits of automation and the accountability of the people deploying it.
What to watch next
Watch for more competition for credentialed experts, more AI-mediated hiring, and sharper worker demands for transparency around model-training contracts. The central business question will be whether companies can access professional judgment responsibly—not merely cheaply.
AI’s next bottleneck may not be compute or raw data. It may be the willingness of skilled people to teach machines how their work is actually done.
