The skill mix AI can't replace, according to a review of 187 studies

A new systematic review published in Discover Artificial Intelligence looked at 187 peer-reviewed journal articles on AI and workplace competency, pulled from an initial pool of 1,352 records and screened through PRISMA 2020 guidelines. The conclusion cuts against a decade of workforce planning: the old model of "AI arrives, workers learn AI skills" doesn't hold up. What actually predicts value now is a bundle, some technical fluency paired with judgment, communication, adaptability, and the ability to know when to trust a system and when to override it.

That reframing matters for anyone building a training budget, a hiring rubric, or a coaching program around the assumption that "AI literacy" is a checkbox. It isn't one skill. It's a combination, and the combination is what's scarce.

Five buckets replaced one checkbox

The researchers grouped the emerging capabilities into five categories:

• AI-enabled functional skills. Using the tools, interpreting analytics output, understanding what a system is actually doing when it produces a recommendation.
• Judgment and orchestration. Knowing when to trust an AI output, how to interpret it, and how to coordinate humans and machines on the same task.
• Learning and adaptability. A standalone category because the tools change faster than any training cycle can track.
• Skill portfolios. Technical, managerial, and interpersonal capability traveling together rather than sitting as separate lines on a resume.
• Psychological readiness. Confidence using new technology without spiraling into anxiety or technostress, which turns out to predict whether any of the other four ever get applied on the job.

None of these function well alone. A manager with strong functional skills but no judgment will defer to a bad recommendation. One with judgment but no adaptability will fall behind within a year as the tools shift beneath them. The value sits in the overlap, not in any single ring of the diagram.

‍

Management is disappearing in pieces, not as a whole

One of the more counterintuitive findings involves what happens to management itself. The reviewed studies generally find that AI is better at replacing pieces of management than replacing management as a category. Routine activities such as information gathering, monitoring, and relatively simple decisions become easier to automate, and firm-level research on robotics adoption shows this playing out already: fewer managers, wider spans of control, and decision rights moving away from traditional hierarchies.

What rises in that vacuum is everything routine automation can't touch:

• Judgment
• Sensemaking
• Coordination
• Interpretation
• Leadership

These become more valuable precisely because the easy 80% of the job gets absorbed by the system, leaving the hard 20% as the entire job description. What a manager brings now is the ability to interpret what the system surfaces, weigh it against context the system doesn't have, and make a call.

Competency models are aging out faster than companies can update them

The most practical finding for HR leaders concerns the machinery organizations use to manage all of this. Traditional competency systems follow a familiar sequence: define the job, define the competencies, assess the employee, train the gaps, revisit on a fixed schedule (usually annually). That cycle assumed something the researchers argue no longer holds, that a job's task composition stays reasonably stable between review periods.

AI breaks that assumption by continuously changing what a role actually requires. A competency model built in January can be measuring the wrong things by summer, not because the employee changed, but because the tools they use and the tasks those tools touch changed instead. The paper's recommended shift is direct: move from a static competency library to a dynamic capability system, one where competencies get sensed, updated, and redeployed continuously rather than assessed and filed away.

What this means for anyone building people programs

The through line across all three findings is the same. Static, single-skill thinking, whether it's a training checklist, a management job description, or a competency model, is losing ground. What's gaining ground:

• Continuous assessment instead of annual review
• Blended technical and human judgment instead of isolated skills
• Systems built to track that blend in real time

For CHROs and heads of L&D, the near-term task isn't picking the next AI training module. It's building the infrastructure to notice, continuously, which combination of skills each person and each role actually needs next, and closing that loop faster than the tools themselves change.