Everyone agrees there’s an “AI skill gap.” Far fewer agree on what it actually is. Is it a shortage of machine-learning PhDs? A lack of prompt engineers? Or something quieter and more structural inside the organizations trying to adopt AI?
To move past the headlines, I did what a data person does: I went to the source. For my MSc capstone I pulled together more than 1.3 million AI-related job postings across the US, UK, Canada and Australia, encoded every skill requirement into embeddings, and clustered the results to see which capabilities cluster together, and which ones employers keep asking for but rarely find.
The findings surprised me, and they changed how I think about “readiness” entirely.
The gap isn’t where you think
The loudest narrative says companies are desperate for deep-learning specialists. The data says otherwise. The most frequently unmet requirements weren’t exotic research skills, they were the connective ones: people who can translate a business problem into a data question, evaluate model output critically, and communicate uncertainty to a decision-maker.
The loudest narrative says companies are desperate for deep-learning specialists. The data says otherwise. The most frequently unmet requirements weren’t exotic research skills, they were the connective ones.
When I clustered roles by their semantic skill profiles, a clear structure emerged. Technical depth and business fluency rarely lived in the same posting, and the roles bridging them commanded the largest premiums and stayed open the longest.
What the postings actually asked for
Aggregating the top co-occurring skills across all four markets produced a consistent shortlist. These appeared far more often than the discourse would predict:
- Data literacy — reading, questioning and reasoning about data, independent of any specific tool.
- Applied AI judgment — knowing when a model helps, when it doesn’t, and how to check its work.
- Communication — turning a technical result into a decision a stakeholder can own.
- Domain grounding — enough context to ask the right question in the first place.
