You've probably added "AI proficiency" to a job description in the last six months. Maybe you listed ChatGPT or Copilot under "preferred qualifications." Maybe you asked a candidate in a phone screen, "How do you use AI in your work?"
They said something about writing emails faster. You nodded. Neither of you learned anything useful.
This is where most companies are right now. They know AI fluency matters. They have no idea how to screen for it. And the gap between "says they use AI" and "has fundamentally changed how they work with AI" is enormous.
Zapier just published the most detailed AI fluency hiring rubric anyone has made public. Their V2 rubric, released March 2026, breaks down what AI fluency looks like across every department. Engineering. Legal. Sales. All of it. It's worth reading in full. But the underlying framework applies to any company trying to hire for AI fluency, not just Zapier.
Here's what they got right, what's harder than it looks, and how to actually screen for AI skills during your hiring process.
"AI fluency" is too vague to hire for
The first problem is definitional. When a hiring manager says they want someone who's "good with AI," they usually mean one of three different things:
- Tool familiarity. Can this person use ChatGPT, Claude, Copilot, or whatever the team runs on? This is table stakes. It tells you almost nothing about how someone will actually perform.
- Workflow integration. Has this person changed how they do their core job because of AI? Do they have repeatable systems, or are they still copy-pasting prompts they found on Twitter?
- System thinking. Can this person design AI workflows that other people use? Can they identify where AI should and shouldn't be involved? Can they evaluate outputs critically and own the results?
These three levels are not a continuum you naturally progress through. Plenty of people plateau at tool familiarity for years. And the signals for each level are completely different.
Zapier's rubric names four tiers: Unacceptable, Capable, Adoptive, and Transformative. The useful insight is in how they define the floor. Their new "Capable" minimum requires AI to be embedded in core work with repeatable systems and clear impact. Using AI for one-off tasks, no matter how frequently, doesn't meet the bar.
That distinction matters. Frequency of AI use is a terrible proxy for fluency. Someone who uses ChatGPT twenty times a day for email drafts is less fluent than someone who built one automated workflow that changed how their team operates.
What Zapier's rubric actually measures
Zapier assesses four components across the hiring process: AI mindset, strategy, building, and accountability. The fourth one, accountability, is new and arguably the most important.
Mindset is the easiest to fake
AI mindset is about orientation. Does this person default to checking whether AI can help with a task? Are they curious about new tools and approaches?
The problem: mindset is easy to perform in an interview. Everyone knows the "right" answers. "I'm always looking for ways AI can improve my workflow" costs nothing to say.
Mindset questions are useful for detecting people who are actively resistant to AI (that's a real signal), but they won't separate genuine practitioners from people who've rehearsed the talking points.
Strategy reveals whether someone has actually built anything
Strategy means this person can explain their approach to using AI: which tools for which tasks, where AI adds value and where it breaks, how their usage has changed over time.
This is where interviews get interesting. Zapier's rubric asks candidates to explain tool and model choices, describe limitations they've hit, and show how their workflow evolved. A strong answer has version history. "I started using AI for X, realized it was unreliable for Y, switched to Z for that part, and now my system looks like this."
The absence of iteration is a red flag. If someone describes their AI usage as a static list of tools, they're probably still in the experimentation phase.
Building is the hardest to assess without watching someone work
Building means this person creates things with AI. Workflows, automations, prompts that other people use, internal tools. This is where the gap between "uses AI" and "fluent in AI" becomes obvious.
Zapier redesigned their talent assessments to watch candidates use AI in real time. They want to see how people prompt, push back on bad output, and adapt. A rough result with visible reasoning and iteration beats a polished result with no process behind it.
This matters. The candidate who shows you their messy, iterative process of getting AI to do something useful is demonstrating the skill you actually need. The candidate who shows you a perfect output probably spent three hours pre-generating it.
Accountability separates practitioners from enthusiasts
This is the component Zapier added in V2, and it addresses the biggest risk in hiring AI-fluent people: someone who ships fast with AI but never checks whether the output is actually good.
Accountability means defining what "good" looks like before starting, evaluating outputs critically, catching problems before they ship, and owning the outcomes. Zapier's framing: "With AI, you can delegate the work, but not the accountability."
This will become more important as AI tools get better. The real risk is someone who uses AI for everything and has no quality bar. You've probably already seen this in the wild. AI-generated code that technically works but nobody can maintain. Marketing copy that reads like a prompt response. Analysis that looks thorough but falls apart under scrutiny.
When you hire for AI fluency, you're also hiring for judgment about when AI output is and isn't good enough.
How to screen for AI fluency at each stage
Zapier built AI fluency touchpoints into four stages: application, screen, skills test, and executive interview. You don't need to copy their exact process, but the principle of compounding signals across stages is sound.
Applications: look for evidence, not buzzwords
Adding "AI proficiency" to a job posting tells you nothing. Candidates will list every AI tool they've heard of.
Instead, ask candidates to describe a specific workflow they've built or changed using AI. Frame it as an open-ended question in the application: "Describe one way you've used AI to change how you do your core work. What was the before and after?" Answers that name specific tools, describe iteration, and quantify impact are strong signals. Answers that read like a LinkedIn post about the future of AI are not.
Screens: probe for the learning curve
Phone or video screens are good for testing whether the application answer was real. Ask follow-up questions that force specifics.
"Walk me through the last time an AI tool gave you a bad result. What happened, and what did you change?" This question separates people who use AI tools from people who understand them. The inability to describe a failure is a red flag. Everyone who uses AI seriously has hit its limits.
"How is your AI workflow different today than it was six months ago?" This is Zapier's "slope" concept. You're looking for an evolution story. Someone who's been using the same three prompts since last year is different from someone who's actively experimenting and building on what they've learned.
Skills tests: watch the process, not just the output
This is where most companies miss the biggest opportunity. If your skills test allows AI (and it should, because that's how people actually work), the process matters more than the final deliverable.
Zapier's approach: observe candidates working with AI in real time. See how they prompt, how they handle bad output, whether they iterate or accept the first response. Anthropic's research on AI fluency backs this up. The strongest fluency signals come from people who iterate and use AI as a thought partner. Reaching for the most tools is a weaker signal than knowing how to get more out of fewer tools.
If you can't observe in real time, ask candidates to submit their process alongside the deliverable. Screen recordings, prompt histories, or a written explanation of their approach all work. A candidate who can articulate "I started here, hit this wall, adjusted my approach, and arrived at this result" is showing you the skill that matters.
Final interviews: test judgment, not just capability
At the final stage, you should know whether someone can use AI effectively. The remaining question is whether they know when not to, and whether they'll hold themselves accountable for what AI produces.
Ask about a time they decided not to use AI for something, and why. Ask about their quality control process. How do they verify AI output before it goes out? What's their threshold for "good enough"?
For management candidates, Zapier requires evidence of leading teams through AI adoption. A manager who's personally fluent but whose team still works the old way hasn't demonstrated the skill you need. Ask what they changed about their team's process, not just their own.
Why department-specific rubrics matter
One of the most useful things about Zapier's rubric is that it's department-specific. What "Capable" looks like for an engineer (AI embedded across implementation, debugging, testing, and documentation) is completely different from what it looks like for someone in a legal role (reusable prompt libraries for contract review, faster and more consistent redlining).
Generic AI fluency rubrics fail because the application of AI varies so much by function. An engineer who built an AI testing pipeline and a marketer who built an automated content system are both demonstrating strong AI fluency, but you'd never assess them with the same questions.
If you're building your own rubric, start with one department. Pick the team where AI fluency has the most immediate impact on output. Define what "minimum acceptable" looks like for that specific role, using concrete examples of the work product you expect. Then expand from there.
The temptation is to build a universal rubric. Resist it. A rubric that's specific enough to be useful for engineers won't be specific enough for marketing, and vice versa.
Building this into your screening process
Most of the advice above boils down to one structural shift: stop treating AI fluency as a line item on a requirements list and start treating it as a skill you assess across multiple touchpoints.
Add an AI-specific question to your application. Probe for specifics in your screen. Design your skills test to reveal process, not just output. And in final interviews, test for judgment and accountability alongside capability.
You don't need Zapier's resources to do this. You need a clear definition of what "good enough" looks like for the specific role, a set of questions that reveal trajectory and iteration, and a skills test that lets you see the work behind the work.
If your candidate screening software already includes async exercises or structured interviews, you're halfway there. The adjustment is small. The signal is worth it.
Zapier's full V2 AI Fluency Rubric is published on their blog. If you're building your own rubric, it's the best starting reference available right now.
