4% of job postings mention AI. But what they’re really referring to isn’t machine learning—it’s agents.

Of the 528,697 job postings analyzed, only 21,640—about 4%—contain at least one mention of AI or machine learning. Amid the endless talk of “AI is changing everything,” this comes as a bit of a reality check. But the figure is misleading: it doesn’t mean that AI isn’t important. It means that demand is concentrated in specific areas—and the language of that demand has changed significantly over the past year.

Most telling of all: the term “agentic AI” is mentioned in 7,908 job postings, surpassing “machine learning” (7,712) and “generative AI” (6,032). Just a year or two ago, ML would have been the undisputed leader. Now, employers aren’t writing about abstract “machine learning,” but about agents, pipelines, and specific tools.

Where There’s a Concentration, and Where There’s Silence

In the Software & Data category, AI is mentioned in 34% of job postings. By comparison: Retail & Hospitality—0.51%, Healthcare—0.37%, Construction—0.25%. In other words, AI in job postings is almost entirely about technical roles and related functions. For most professions, AI was never mentioned in job descriptions, and it still isn’t.

Within the “Software & Data” category, the variation is also telling. Pure AI/ML roles account for 76% of mentions, as expected. But “Full-stack” accounts for 37.5%, “Cloud & DevOps” for 35.7%, and “Backend” for 29.4%. Employers aren’t looking for a separate “AI engineer” for every task. They’re incorporating AI skills into standard engineering roles—just as they once did with Docker or Kubernetes.

What They Actually Ask You to Do

The most valuable part of the study isn’t the percentages, but the quotes from real job postings. The tasks are surprisingly specific:

• Build AI agents, prompts, and retrieval pipelines; connect models to corporate data and tools.
• Architect MCP servers that expose enterprise APIs as context for agents.
• Assess LLM risks: prompt injection, data leaks through tools, and excessive access.
• Automate SOC processes through LLM-assisted incident investigation.
• Design handoff points between agents and humans—where AI makes a decision and where a human confirms it.

This isn’t just “using AI to boost productivity.” These are engineering challenges with specific architecture, responsibilities, and reliability criteria. One job posting states explicitly: “AI accelerates implementation, but engineers remain responsible for system design, debugging, and code review.”

What This Means in Practice

First: the hype surrounding “AI skills mandatory for everyone” isn’t backed by data. Outside of technical roles, AI is a rarity in job postings. If you’re not in software, product, IT, or security, your employer isn’t in a hurry to require AI skills just yet.

Second: where AI is mentioned at all, the language has shifted from the general to the specific. Employers already know what they want: not just an “ML model,” but an agent with RAG, MCP, and guardrails. If you’re positioning yourself as an AI engineer, the abstract phrase “I work with LLMs” no longer cuts it—you need to provide specifics: which agents, which pipelines, and which reliability metrics.

Third: the emergence of security challenges related to LLMs (prompt injection, tool data exfiltration, over-broad tool access) in job postings signals that the market is ripe for a distinct AI security niche. For now, this falls under general cybersecurity roles, but the terminology has already taken hold.

Data restrictions: Mentioning the term does not equate to requiring the skill. The authors are candid about this. But even as an indicator of the direction—the picture is clear: AI hasn’t flooded the market, but has been integrated selectively into engineering roles, and the language of job postings has already caught up with practice.

Source: Hacker News - Newest: “AI” “LLM”