Open almost any tech job description written in 2026 and you will find AI near the top of the requirements. The trouble is what those requirements say. "Experience with LLMs." "Familiarity with Claude, Cursor or similar tools." Lines like these tell candidates that AI matters, but they do not tell anyone what good looks like.
This article uses the job description data in Lenny Rachitsky's State of the tech job market: late 2026 to show how fast AI requirements have spread, and then sets out a practical way to define AI fluency for product, engineering and design roles. It is written for hiring managers and talent acquisition teams at tech companies. Our view: AI fluency belongs in your hiring standard, but as observable outputs at clear levels, not as a list of tools that will change within a year. It expands one part of our guide to the tech job market in 2027.
What the job description data shows
The report counts how often skills are mentioned in job descriptions across global openings at tech companies.
| Function | Two years ago | Today |
|---|---|---|
| Product management | Top skills were SQL, Jira, AWS and Python | "LLM" is the most-mentioned skill; "Claude" is third; Cursor, OpenAI and MCP are in the top eight; "evals" is in the top ten. SQL is still holding strong |
| Engineering | Python was the most-mentioned skill | Python is still first and still growing; "LLM" has just passed Java; Java is declining; Kubernetes is holding its ground |
| Design | Figma led every other skill | Figma is still first, and demand for it has grown; "Claude" became the second most-mentioned skill in spring 2026, passing CSS; Cursor has also passed CSS |
Two things stand out. First, AI terms have moved in beside the core skills of each function rather than replacing them. Second, much of the new vocabulary is tool and product names. These are counts of words in job descriptions. They show what employers ask for, not what candidates can do.
Why tool names make a weak standard
Tool names are easy to write and easy to match on a CV, which is why they spread. They make a poor hiring standard for three reasons:
- They date quickly. Two years ago the top PM skills were SQL, Jira, AWS and Python; today AI terms fill much of the top ten. We expect the tool names to keep changing through 2027, and requirements written around them to age with them.
- They test exposure, not judgment. Having used a tool says nothing about whether someone can tell when its output is wrong.
- They screen out good people. Access to AI tools differs by employer, market and budget. A strong candidate whose current employer uses a different tool can fail a keyword screen for the wrong reason.
Define AI fluency as outputs, at three levels
A better standard describes what a person produces with AI, at increasing levels of responsibility. The examples below are our own, not the report's. Adapt them to your roles.
Product managers
| Level | Observable output |
|---|---|
| Working | Uses AI to draft and check specs, research summaries and analyses, and can explain what they verified by hand |
| Strong | Has defined how the quality of an AI feature is measured, written or run the evals, and made a ship decision based on the results |
| Leading | Has set the team's approach to AI features: what to build, how to measure quality and when to stop |
Engineers
| Level | Observable output |
|---|---|
| Working | Uses AI coding tools daily and reviews generated code with the same care as a colleague's |
| Strong | Has built and shipped a feature that uses an LLM, including evaluation, monitoring and handling of failures |
| Leading | Has designed systems where AI components are reliable, observable and safe, and set standards for others |
Designers
| Level | Observable output |
|---|---|
| Working | Uses AI tools to explore ideas and produce prototypes faster, and knows where the tools fall short |
| Strong | Has designed the user experience of an AI feature, including how users understand, correct and trust its output |
| Leading | Has defined patterns for AI features across a product, so behaviour is consistent and understandable |
Each role then states the level it needs. Most mid-level roles need "Working" across the board and "Strong" in the area that matters most for the team.
Write it into the job description
Replace tool lists with one or two output statements. Two examples:
-
Before: "Experience with LLMs, Claude, Cursor or similar tools."
-
After (engineering, Strong): "Has shipped at least one feature that uses an LLM, and can show how its output quality was measured and monitored in production."
-
Before: "AI-savvy product manager."
-
After (PM, Strong): "Has defined quality criteria for an AI feature and used evaluation results to decide whether to ship."
Tool names can stay in a "nice to have" line. They should not decide who passes screening.
Keep the core skills in the standard
AI fluency sits on top of each role's foundation. The data makes this clear: Python is still the top engineering skill and still growing, Figma demand has grown, and SQL still holds strong for PMs. A scorecard that rewards AI fluency and ignores the core skills will hire people who can prompt but cannot build.
A practical rule: AI fluency is one dimension of four or five on the scorecard, with its own written description of "strong" and "not enough". It should rarely be the deciding factor on its own.
Assess it with a work sample
Interview questions about AI produce rehearsed answers. A short work sample shows the real thing. Give candidates a task they complete with AI tools during the interview, then score:
- how they break down the problem;
- how they check and correct the AI's output;
- what they decide to do themselves.
Score judgment, not speed or familiarity with one tool. Our guide to assessing tech candidates when every CV says LLM covers work samples in more detail.
Roll the standard out across teams
A definition only helps if every hiring manager uses it the same way. Four steps keep it consistent:
- Agree the levels once. Product, engineering and design leaders review the three tables above, adjust the wording and sign off. Store the final version where every hiring manager can find it.
- Tag every open role. For each requisition, record the level needed in each function's AI dimension. This takes minutes and stops each manager inventing their own bar.
- Train interviewers on examples. Show interviewers two real answers, one strong and one not enough, for each level. Examples make the difference between levels concrete in a way definitions alone do not.
- Review after a quarter. Look at the people you hired. Did the AI dimension predict how they performed? Adjust the wording where it did not.
The same levels also help internal mobility. An engineer moving into a PM role, or a designer taking on research, can see which outputs they need to show at the next level.
A note for APAC teams
Regional teams often reuse job descriptions from headquarters, and the skill rankings in the report come from global job descriptions. We expect those AI requirements to travel into APAC hiring through shared templates. Before using a headquarters template, check two things: whether the tools it names are available and commonly used in each market, and whether candidates can show the output in the language they work in. To check, count how many of your job descriptions in each market name specific AI tools, and how many describe outputs.
Next step
Pick one open role in each function and rewrite its AI requirement as an output at the level the role needs. Then add the matching dimension to the scorecard.

Author
Moka AI Editorial Team
The Moka AI Editorial Team creates insightful, practical, and engaging content on HR, recruitment, and AI innovation. We explore emerging trends, share industry insights, and turn complex ideas into actionable knowledge to help HR professionals navigate the evolving world of work.



