The State of Hiring for AI Fluency 2026

This is not a new failure mode in hiring. Hiring has leaned on the same proxies for decades: years of experience, pedigree across companies and universities, confident interview answers, and the right keywords on a resume. And for decades, the research has quietly been telling us that those proxies do not work (see TestGorilla's State of Skills-Based Hiring 2025). A widely-cited meta-analysis by Van Iddekinge and colleagues found that the correlation between years of experience and actual job performance was roughly 0.06, effectively zero. Yet these signals remain the default inputs into most hiring decisions, not because they predict performance, but because they are easy to collect, easy to compare, and easy to feel confident about.

AI fluency is now being filtered through that same broken funnel. And the requirement for it has made this gap more acute, more visible, and more critical to address.

The skill organizations say they need is an applied one: integrating AI into real workflows, exercising judgment over outputs, adapting as tools and constraints shift. But the signals hiring processes are built to capture are different in kind. They are declarative — what the candidate knows, what they have used before, how fluently they can describe it in a thirty-minute conversation. The result is predictable. Candidates who articulate AI well are rewarded; candidates who apply it well may or may not be, because the process has no reliable way to distinguish between them. And once the wrong candidate is hired, the consequences compound: slower execution, inconsistent output, and misplaced confidence in AI-generated work that no one on the team is equipped to audit.

The gap, then, is not about effort or intention. It is about evidence. Specifically, about the kind of evidence current hiring processes are built to collect, and why that evidence cannot tell you what you actually need to know. Until that changes, the paradox holds: rising investment in AI-fluent talent, alongside falling confidence that the process is finding it.

Three people, three different vantage points, one shared diagnosis: the gap between what organizations say they are hiring for and what they can actually identify and measure is widening, not closing.

To understand why, and to build something more useful than another list of observations, we surveyed nearly 2,000 senior hiring leaders across the United States and the United Kingdom. 56% identified as senior leaders, or senior decision-makers within hiring - the people shaping hiring processes in real time. The sample spanned organizations hiring across one to 250-plus roles per year, across industries including technology, healthcare, finance, consulting, and education, and across hybrid, remote, and on-site teams.

We asked a focused set of questions: how organizations are defining AI fluency today, how they are measuring it, where those processes break down, and what actually happens once candidates join. We also drew on the expert conversations from the February event, and on a selection of external research, including frameworks from Zapier and the Microsoft Work Trend Index.

The findings that follow reflect a market in the middle of working something out. There are no clean answers yet in a technology whose fundamentals and use cases are evolving faster than the speed of light, but there are clearer questions, and a framework for thinking about them more rigorously than most hiring processes currently allow.

By the end of this report, you will have:

• A practical, nuanced understanding of how to define AI Fluency for your specific use case.

• A thorough framework created and used by TestGorilla to implement in your AI hiring practices that allows you to identify truly AI-fluent talent across all roles.

By most measures, this looks like a system that is working. It isn't. And the reason is instructive.

Building a definition and building a measurement are two different things. A definition tells you what you are looking for. A measurement tells you whether you found it. Most organizations have done the former while believing they have done the latter.

The infrastructure is real, but it is measuring the wrong inputs. And infrastructure that measures the wrong thing does not produce better outcomes. It produces more confident wrong ones!

We explore 3 tensions worsened by the gap between hiring for AI fluency requirements, infrastructure, and signals.

Hiring has often leaned on superficial signals. When you cannot directly observe whether a candidate will perform well in a role, it’s natural to look for indicators that may correlate with performance:

But these are not measures of capability. They’re just signals we’ve assigned a mark of quality that might yield a higher-quality candidate. 

The problem with these signals is that they erode. We’re building Apollo 11 using the same rules from Apollo 1 - it won’t likely make it to outer space, forget the moon. 

These are just proxies. And as the research on hiring proxies consistently shows, proxies feel more reliable than they are. The degree requirement felt reliable. The years of experience felt reliable. The confident interview answer felt reliable. None of them predicted performance with any meaningful consistency. AI fluency, filtered through the same process, is producing the same result.

A candidate who has never built an AI workflow in a professional context can, with modest preparation, describe one convincingly. They can name the right tools, frame the right trade-offs, and deploy the right language. An interview rewards articulation, but it was never designed to test execution.

Jason Miller, Head of People Intelligence and AI at Natera, put it with characteristic directness at our February event: "Putting ChatGPT on your resume is the equivalent of saying proficient in Microsoft Office. We are getting to a point where there are pre-established table stakes, and no one has really figured out yet how much AI we allow in an interview, or how to cut through the performance."

The distinction matters because what organizations say they need from AI-fluent hires bears no relationship to what a tool-awareness threshold actually screens for. Efficiency gains, productivity improvements, and revenue impact (the outcomes organizations consistently cite as the measures of AI performance) require judgment, verification, and the ability to adapt as conditions shift. None of those capabilities is visible from a list of tools a candidate can name.

When you hire a skilled prompt writer, you get excellent prompts. That’s not nothing. But it is not a measure of the outcomes organizations are trying to drive. A candidate who has built genuine fluency has not just written good prompts - they have used AI to redesign how work gets done, caught errors before they caused damage, and made judgment calls about when not to use it at all. The threshold needs to reflect that, not the lowest version of it that still counts as engagement with AI.

Only 26% of organizations currently require candidates to demonstrate independent AI use and verify results as part of the hiring process. That number should be the floor. Currently, it functions as the ceiling.

The second is role simulation: Asking a candidate to walk through how they would approach a specific, realistic task using AI, with a constraint introduced mid-way, surfaces adaptation in real time. The strongest candidates reframe and adjust. Candidates who are just performing stall, repeat vocabulary, or describe what they would theoretically do rather than demonstrating how they actually think.

The third is a collective debrief structured around shared evidence: As Lou Adler observes, "You get everybody in the interviewing room, and you debrief collectively, and you start sharing that evidence. And actually, the truth comes out. By everybody sharing the evidence, you see things you didn't think you actually heard." Individual interviewers are susceptible to confident performance. A structured debrief using shared, specific evidence is considerably harder to deceive.

The interview, as currently structured, is not designed to observe metacognition. Romina Da Costa, TestGorilla's Director of Talent and Assessment Science, frames the underlying principle with precision: "What we get from a psychometric lens is that we are not really looking at the output (what did they create with AI) but that metacognitive process of how they got to that output, and why they defend it as the best output given the constraints of time and resources and what they had in front of them. And how did they adapt and evolve, because the technology isn’t standing still. That is the key differentiator."

Redesigning the interview is the difference between a process that reliably produces confident wrong hires and one that finds people who can actually do the work.

62% of US hiring managers cite the challenge of defining AI fluency differently for technical versus non-technical roles, compared to 46% in the UK. Paradoxically, the US is more aware of the complexity of the problem and less equipped to resolve it. 57% of US organizations report a lack of internal alignment on how much AI augmentation is actually required in a given role, versus 41% in the UK. Awareness of a problem, it turns out, is not the same as having solved it.

Where the gap becomes most concrete is in the threshold data. In the US, 45% of organizations set the minimum bar for AI fluency at tool awareness, or simply knowing what tools exist and where they might broadly apply. In the UK, that figure is 29%. British organizations, by contrast, have a meaningfully higher concentration, setting the bar at independent use and verification: the standard that actually requires a candidate to demonstrate applied judgment rather than simply name the right tools. One market is screening for vocabulary. The other is screening for capability.

Among organizations that have not defined AI fluency at all, the failure modes diverge further. In the US, 29% say they do not need a definition, while a further 34% say they struggle to define it. In the UK, only 22% say they do not need a definition, while 47% say they simply have not gotten around to it yet. 

Put plainly, the UK failure is a capacity problem. The US failure, in a significant number of cases, is a conviction problem. Organizations that have decided the question is not worth answering are not going to stumble into better hiring outcomes.

None of this is abstract. Set a lower bar, with less internal alignment, in a market that is more likely to treat the definition question as already resolved, and you get 33% of organizations experiencing AI-driven errors often. Raise the bar, acknowledge what you have not yet done, and that figure drops to 13%.

Garbage in, garbage out is a principle most technology teams understand intuitively. It applies equally well to hiring. The quality of the standard set upstream determines the quality of the outcome downstream. That relationship is not theoretical. It is measurable, and it is measured here.

The difficulty is real and structural. AI fluency for a software engineer looks like chaining model calls with fallback logic and building evaluation tests to catch hallucinations. For a revenue operations or marketing leader, it looks like architecting multi-step AI agents that orchestrate outreach, scoring, and follow-up across a pipeline, as well as knowing where human judgment needs to stay in the loop before a prospect is touched. For an HR professional, it looks like redesigning screening workflows and using AI to reduce bias in job description language.

Different tools. Different risk profiles. Different vocabulary. But the underlying behaviors -  judgment over outputs, ethical awareness, the ability to communicate AI use to colleagues who need to understand or audit it - remain the same across all three.

Most organizations have defined AI fluency at the tool level. A tool-level definition cannot cross functional lines. Which means the 54% who cannot define it across functions are failing because they started from the wrong end of the problem. And unsurprisingly, in the UK, 47% of those who haven’t defined AI fluency yet say they have “not gotten round to it yet,” not due to indifference or confusion, but capacity. The intention exists. The bandwidth does not.

These are not abstract benefits. They are operational advantages that compound over time. A clear definition makes every downstream process more consistent and more predictive: job description writing, screening, interviewing, onboarding, and performance management.

Zapier has published not one but two AI fluency frameworks in the first quarter of 2026 alone. That pace is worth pausing on. In a technology environment where the tools and their capabilities are shifting faster than most job descriptions can track, even organizations that have invested seriously in defining fluency are finding that their definitions require updating. 

Zapier's frameworks structure assessment around AI Mindset, Strategic Acumen, and Technical Skills, with a maturity progression from Unacceptable through Capable and Adoptive to Transformative, where Transformative means AI has fundamentally changed not just how someone works, but how the people around them work. It is a deployable framework, particularly for organizations at an early stage of AI hiring maturity. Its scope, by design, does not extend to systems thinking, ethical judgment, or the collaborative behaviors that determine whether individual fluency scales into organizational impact.

Both frameworks reflect the same underlying recognition: tool lists are insufficient, behavioral dimensions are necessary, and the definition needs to survive the next model release. Neither claims to have closed the problem. Both have moved the conversation forward.

“Tool agnostic” is an emerging conversation. A framework built on specific tools has a shelf life measured in months. A framework built on the behaviors that make someone effective with any tool, be it judgment, adaptability, ethical reasoning, or collaborative transparency, survives the next model release and the next wave of capability only when combined with actual AI skills.

Romina Da Costa, TestGorilla's Director of Talent and Assessment Science, framed the design challenge precisely at our February event: "We advocate for taking a really broad and holistic view of AI fluency. It is not a singular thing. The most lasting skills go beyond the mastery of any single tool into something that is completely tool agnostic."

Where the TestGorilla framework extends further (which we’ll introduce in a moment) is in the dimensions that determine whether individual AI fluency scales into organizational impact. The systems thinking, ethical judgment, and collaborative behaviors that allow AI use to be understood, audited, and built upon by others. 

These are the dimensions our survey data identifies as hardest to assess, and the ones most directly linked to the failure modes described earlier. 

Each pillar describes a distinct behavioral domain. Together, they define what it means to be AI-fluent in a way that is specific enough to measure, broad enough to travel across functions, and durable enough to remain valid as the tool landscape evolves.

That dependency is a feature rather than a bug. Tom Booth, Talent Lead at payments fintech Primer, spoke to this directly at our February event: "Company size changes the emphasis quite meaningfully. At earlier stages, the bias is typically toward experimentation and speed. As companies grow, the systems thinking side and the downstream impact of AI use become more important."

Jason Miller, Head of People Intelligence and AI at Natera, working in a regulated healthcare environment, placed responsible and ethical AI use at the top of his list: "If we use AI and it mishandles patient data, or it hallucinates test results, that is not a little bit of egg on our face. That is a lawsuit."

This framework is meant as a diagnostic tool, not a fixed hierarchy. The right weighting depends on your industry, your organizational stage, and the specific role.  What the TestGorilla 5 Pillars do, however, is ensure that no critical dimension gets quietly skipped. 

What AI-fluent looks like in practice: 

• An engineer who can chain LLM calls with fallback logic and build evaluation tests to flag hallucinations before they reach production. 

• A marketer who can run AI-assisted A/B tests and audit copy for demographic bias. 

• A customer success manager who can use AI to synthesize ticket patterns and surface systemic product issues, not just close individual tickets faster.

What this pillar guards against: 

The “Confidence vs Competence Problem,” i.e., the candidate who can describe AI tools accurately but cannot operate them under real conditions. 

This is the most common false positive in current hiring. Detecting it requires moving beyond theoretical questions into scenario-based tasks that mirror actual job workflows.

What our survey reveals: 

For 31% of hiring managers, a primary challenge is distinguishing candidates who truly understand AI from those who just use the right terminology.

What AI-fluent looks like in practice: 

When a tool a candidate relies on is deprecated or updated, they do not wait for retraining.  They identify an alternative, test it against their existing workflow, document what changed, and share the updated approach with their team. 

What it guards against:  

The candidate who has mastered a specific platform but cannot transfer that mastery when the platform changes. 

Given that the Microsoft and LinkedIn 2025 Work Trend Index found AI adoption nearly doubled in six months, the half-life of tool-specific knowledge is short. Learning agility is what converts today's investment in AI-fluent talent into long-term returns.

How practitioners test for it: 

Tom Booth described his team's approach at our February event: "We might ask someone to answer a question or build something in a live interview. Then, once that is complete, we throw in a requirement change, take away a tool, or introduce a constraint. The stronger candidates do not just tweak the output. They step back, reframe, and adjust their whole approach."

What AI-fluent looks like in practice: 

A product manager who, before deploying an AI-generated feature recommendation, asks: 

• Who consumes this output? 

• What assumptions are embedded in the model? 

• What happens downstream if the recommendation is wrong? 

• What is the blast radius if it goes wrong at scale? 

And then designs a human-in-the-loop step into the workflow accordingly.

What this pillar guards against: 

The individually productive hire who does not scale. 

This is precisely the failure Tom Booth described: “A candidate who 10x'd their own output but produced documentation so light that no one else could follow it, and who bypassed security and compliance processes in the process. Individual AI productivity that cannot operate within a system is not a productivity gain. It’s a liability.”

What our survey reveals: 

23% of hiring managers specifically flag difficulty assessing cognitive skills like systems thinking and verification of AI output. This pillar is the hardest to detect in a standard interview and the most consequential to miss.

What AI-fluent looks like in practice: 

A recruiter who, before deploying an AI screening tool, asks: 

• Has this model been audited for demographic bias?

• What data was it trained on?

• What is our process for candidates who are screened out? 

• And: Who documents the answers so the organization can demonstrate compliance if the process is challenged?

What this pillar guards against: 

The fast mover who creates slow-burning liability. This is also a validity signal. If a candidate can't audit an output for bias or hallucinations, they lack the "systems thinking" required for senior roles i.e. an awareness of the overall context and implications of AI use, downstream impacts on data security, things like compliance with EU AI acts, SOC2, reputational damage, etc.

Organizations that prioritize execution speed without ethical guardrails do not avoid consequences. They only defer them and find out they come back to bite later.

What our survey reveals: 

Responsible and ethical AI use ranks as the most frequently cited #1 skill that is looked for in candidates. Ahead of even applied AI use and workflows. 20% of respondents ranked it as the single most important pillar overall. 

Interestingly, practitioners are ahead of job descriptions on this. The market has not yet caught up to what hiring managers already know they need.

Isabel Berwick, writing for the Financial Times and speaking at our February event, made the broader case: "In an agent-led future, our human work will be much more about collaboration, innovation, and ethical judgment. That human capacity is going to be needed more than ever."

What AI-fluent looks like in practice: 

A data analyst who does not just produce an AI-assisted report but documents which inputs were used, which outputs were verified manually, where AI judgment was applied and where human judgment overrode it. All so that the next person who picks up the work can understand, audit, and extend it without starting from scratch. 

What this pillar guards against: 

The AI fluency silo. 

A highly capable individual who cannot communicate their AI use creates knowledge concentration rather than knowledge distribution. 

The organizational multiplier effect of AI fluency only compounds when it is transferable. When it is not, it stops at the individual and leaves when they do.

How practitioners assess it: 

A talent leader for a large Dutch tech company described at our February event: "In tech assessments, candidates can use AI as long as they justify and show the use. What are the prompts they are using? How are they using it? Feel free to use AI, as long as you can explain it and the logic behind it." The output matters. The transparency about how the output was reached matters more.”

These are not interview questions. They are evaluation dimensions. And they require a different kind of evidence than a thirty-minute conversation can produce.

When we asked nearly 750 hiring managers to rank the 5 pillars in order of importance, the answers were almost uniformly distributed. No pillar dominated. No pillar was ignored. 

That finding explains a pattern visible elsewhere in the data. Hiring managers are not unaware of the full picture. 49% report a lack of internal stakeholder alignment on how much AI augmentation a given role actually requires. 54% cannot define fluency consistently across technical and non-technical functions. When internal agreement is out of reach, organizations default to the one dimension that requires no negotiation: tool awareness. It is not that they believe tool awareness is sufficient. It is that tool awareness is the only thing everyone in the room can agree on.

The framework provides the vocabulary to assign that weight rigorously rather than by default.