AI Is Now the Second Most-Requested Skill in Job Postings — What Employers Actually Want in 2026

The headline number: AI ranks second across all roles

The “one in five” figure comes from an analysis of more than 360,000 postings spanning late 2025 to mid-2026. AI’s second-place ranking across all roles is the clearest signal yet that the skill has moved from differentiator to line item.

Several supporting data points cover U.S. job postings specifically. Dice analyzed seven million U.S. tech job postings and found AI skill requirements rose from 15% of postings in January 2024 to 75% by mid-2026. Lightcast data cited by the Bipartisan Policy Center found U.S. postings explicitly requiring AI-related skills increased 144% year over year. In U.S. data and analytics roles specifically, AI appears in close to 45% of postings.

The composition matters as much as the growth. When a skill appears in three-quarters of technical postings, it stops functioning as a filter and starts functioning as a floor. The interesting question is no longer whether a role mentions AI. It is which kind of AI work the role actually involves, and how deep the expectation runs. The phrase itself remains slippery, as examined in What Does ‘AI Experience’ Actually Mean? A Hiring Phrase Without a Shared Definition.

Category 1: AI literacy, the broadest reach and the weakest moat

AI literacy is the widest category in the data. It covers a general understanding of what AI tools can and cannot do, using them effectively, and critically evaluating their outputs. It is the layer that reaches beyond engineering into marketing, operations, finance and support roles.

Prompt engineering sits inside this category, and it is among the fastest-growing sub-skills. Dice found 253% year-over-year demand growth for prompt engineering. LinkedIn’s 2026 labor market report showed U.S. jobs requiring AI literacy skills, including prompt engineering, grew 70% year over year.

Then comes the reality check. Across the broader dataset, prompt engineering appears in 7% of all postings. ChatGPT proficiency appears in 6%. LLM familiarity appears in 4%. Those are respectable absolute numbers, but they do not describe a market clamoring for dedicated prompt specialists. Employers are not hiring standalone prompt engineers. The skill appears as one component of broader roles, much the way Excel proficiency became a baseline expectation for analysts two decades ago.

Analysts describe the most valuable position as “prompting plus something”: AI amplifying domain expertise rather than replacing it. A marketer who can direct a model through a campaign brief, or a paralegal who can pressure-test a contract summary, is more useful than someone whose only credential is familiarity with a chat interface. The moat is not the prompting. The moat is the something.

Category 2: Agentic AI, fastest growth by rate of change

If AI literacy has the broadest reach, agentic AI has the steepest trajectory. Stanford’s 2026 AI Index found the share of AI job postings mentioning agentic AI rose from 0.69% to 18.9% in a single year. Dice’s analysis shows demand for agentic AI grew 587% year over year, with AI agents themselves growing 503%.

The concrete skills behind the label are specific and checkable. Postings ask for retrieval-augmented generation, vector databases and agent orchestration frameworks. A 2026 analysis of 390 AI engineer postings found that 56% mentioned AI evaluation, 50% mentioned AI agents, and 26% required RAG experience specifically. Engineers mapping a learning path can find the same toolkit covered in 5 Free Courses to Learn AI Engineering in 2026: From LLM Basics to Production Systems: model APIs, embeddings, vector stores, RAG, agent workflows, evaluation, serving and monitoring.

Agentic expectations are not confined to engineering roles. For non-technical professionals, familiarity with what agentic workflows are and can accomplish is now expected in roles that work alongside engineering teams. Product managers, analysts and operations leads are increasingly asked to scope what an agent should do, where it should stop, and how its output gets checked. That is literacy in the service of oversight, not implementation.

Category 3: Machine learning, the foundational skill that grew more specific

Machine learning remains the most consistently requested foundational technical skill across AI hiring history, concentrated in engineering, data science and research roles, though the terminology appears more broadly. What changed in 2026 is the level of specificity. Postings increasingly request lifecycle experience such as model evaluation, fine-tuning and machine learning operations rather than general ML knowledge.

That shift explains why evaluation skills now show up across categories, not just ML roles. The 390-posting analysis found 56% of AI engineer postings mention evaluation, a higher share than mention agents. Evaluation is the skill that connects literacy, agentic work and traditional ML: someone has to decide whether the system’s output is good enough, and increasingly that someone is expected to have done it before. The same responsibilities keep appearing under different titles, a pattern unpacked in ML Engineer vs. AI Engineer vs. LLM Engineer: What Each Role Actually Builds.

Category 4: Data engineering, the foundational layer beneath AI roles

Data engineering work is unglamorous and foundational: building the pipelines and storage systems that feed machine learning models. Every agent and every fine-tuned model sits on top of data infrastructure someone has to design, maintain and monitor. Recruiter attention and posting volume do not always point at the same roles, and data engineering roles illustrate that gap.

What the data shows about AI skill demand

The through-line across all four categories is that AI skills are additive to domain expertise, not a substitute for it. The postings reward “prompting plus something,” evaluation as a cross-cutting hiring signal, and concrete, checkable experience such as MLOps and RAG. Those are the items a candidate can point to: a deployed retrieval system, an evaluation harness with documented results, a model that went through fine-tuning and monitoring in production.

The second lesson is about reading the numbers. Growth-rate figures and absolute prevalence answer different questions. Agentic AI’s 587% year-over-year growth and prompt engineering’s 7% absolute prevalence can both be true, and together they describe a small base expanding quickly inside a broad but shallow literacy layer. A 587% increase on a small base is a signal about direction, not scale. A 7% share is a signal about scale, not momentum. Hiring decisions and learning plans should weigh both.

Caveats and methodology: how to read these numbers responsibly

Posting analyses measure what employers write, not what they hire for. A requirement listed in a job description may be aspirational, copied from a template, or negotiable in practice. The data describes stated demand, and stated demand is a proxy.

Scope matters too. The Dice analysis of seven million postings, the Lightcast figures cited by the Bipartisan Policy Center, and LinkedIn’s labor market report all cover the United States. Those numbers do not generalize globally, and hiring patterns in other regions may differ in both level and composition. The 360,000-posting dataset and the 390-posting AI engineer analysis are the figures here not tied to a single national market.

Finally, year-over-year percentages can inflate small bases. Read the growth figures as directional, the prevalence figures as structural, and the combination as the most honest picture available of what employers say they want in 2026.