There's a straightforward way to measure how well employees use and understand AI. It doesn't involve how often they use it, the number of tools they've adopted, whether they've completed training, or whether their department has an enterprise license. It's whether they can write a good prompt.
That may sound simplistic, but it isn't. The quality of a prompt, including how precisely users define a task, how much context they provide, and how well they iterate, may be the closest thing we have to a real measure of AI fluency. But most enterprises have little or no visibility into this data.
If they did, they'd know whether workers are using AI as a glorified search tool and document generator (spoiler: most are) or as a true thinking partner that helps them do things they otherwise couldn't.
It would also confirm what most organizations already suspect: AI fluency is not evenly distributed. A small number of employees are prompting with sophistication and getting outsized results, but the majority are still finding their footing, according to recent research by WRITER.
That discrepancy shows output quality, how work gets done, and business results. Identifying it is the first step in closing it.
Fluent users also weave AI into their workflows across multiple steps rather than posing a single big query. And they iterate, asking follow-up questions, editing responses, and validating results rather than accepting the first answer.
Perhaps most importantly, good prompting means evaluating output with a critical eye: demanding citations and explanations and pushing back on confident-sounding assertions even when the topic is outside your domain.
“[AI] is not the same as human cognition, but it allows us to apply computational thought to solve problems and scale up the amount of thought we can solve problems with," says Dr. Jules White, senior advisor to the chancellor for generative AI at Vanderbilt University and a professor of computer science.
"What we're really trying to measure is how effectively someone can figure out how to direct and use all this computational thought to solve problems,” he adds. “We're trying to measure a combination of creativity and problem-solving skills."
Researchers at the Ringling College of Art and Design and Cork University Business School are among them. Their Framework for AI Fluency breaks prompting competency into three components:
- Describing a task clearly enough for AI to execute it
- Delegating the right work to the right tools
- Discerning whether the output is good
The research on what makes prompts effective points in the same direction. A 2025 University of China study on AI literacy found that inadequate prompting produces low-quality or irrelevant outputs. And weak critical evaluation prevents users from assessing the logical coherence of what AI returns.
A landmark analysis led by Sander Schulhoff of Learn Prompting reviewed more than 1,500 academic papers and over 200 prompting techniques. It found that technique matters—a lot. Breaking a problem into sub-problems, asking the model to critique its own output, and providing concrete examples aren't stylistic choices. They're the difference between a tool you use and a capability that works for you.
Most organizations claiming to be AI-first are basing that claim on the only measure they're tracking: AI usage. But usage isn't proficiency. In fact, high adoption rates have become so unremarkable that tracking them has largely lost meaning.
"Most leaders don't care about [AI] usage anymore because it's so high, and honestly, it's an expectation," says Brian Conaway, vice president of AI and people science AI at ServiceNow.
The intelligence about who's prompting well and who isn't is what most enterprises lack.
If visibility into prompt quality is so valuable for understanding AI fluency, why is it so opaque? There are three main reasons.
Even when employees use sanctioned tools, with some exceptions, most AI interfaces aren’t designed to expose prompt data to their employers. This is partly product design and partly contractual.
The major model providers—including Anthropic, OpenAI, and Google—have constructed enterprise agreements to protect user input, not expose it. Usage analytics show volume, not what the prompt or output said. But even if they did, most enterprises wouldn't want to kick that hornet's nest.
Prompts are unfiltered and reveal a lot about what someone might be struggling with or avoiding. Unlike email, prompts are one-sided and confessional by nature. Monitoring them would likely trigger a backlash over privacy that's not worth the insight gained. (The privacy of prompts is also dictated by data privacy policies within organizations.
When audit logs are exposed, they’re typically used only for litigation or compliance purposes.
"If you're trying to inspire people to be creative, they have to feel safe," White says. "I wouldn't want to work for a company that took away my privacy."
A 2026 study by BlackFog found that nearly half of workers use AI tools (many are free versions) without IT's approval, and that 63% believe it's OK to use unsanctioned tools when no approved option exists.
Illustrating this gap, a study by Harmonic Security found that only 40% of organizations have purchased official AI subscriptions. That means the majority of employees using AI are doing so through personal or free-tier accounts that leave no enterprise footprint.
When employees use unsanctioned accounts, IT has little to no visibility into how these tools are used or what potentially sensitive data is being entered into them.
Even where approved tools exist, the primary way employees interact with AI is by copying and pasting information from one place to another. A developer might copy a block of code. A marketer might paste a creative brief or competitive analysis. In both cases, the interaction happens inside a browser, not a connected system that logs anything.
LayerX's Browser Security Report 2025 found that 77% of employees paste data into AI prompts, with half of that being potentially sensitive information. None of it leaves a trace in corporate systems. The prompt and its outputs disappear into the ether.
That means the most telling moments in any AI interaction—how clearly someone defined a problem, how much context they provided—leave no record.
Enterprises can't monitor their way to AI fluency, but they can build the environment for it. The strategies that work don’t involve surveilling workers' AI use but rather elevating their knowledge. Two approaches have shown early results.
One of the most effective ways to build AI fluency across the enterprise is by leaning on your most advanced users. An AI champions program embeds curious, capable, trusted advocates in each business unit to lift others who may be struggling and help them apply the broader AI strategy to their everyday work.
"The champions idea does work [because] it's effective to get people together in small groups to see interesting things from their peers," White says. "It's important to be empathetic and inspirational with each other."
The case for organizing champions by line of business comes down to context. A champion inside the sales organization understands customer relationship management (CRM) workflows and sales cycles. A champion on the legal team knows which AI outputs need scrutiny and why.
Generic AI training teaches people what the tools can do. A line-of-business champion teaches colleagues what the tools can do specifically for them, within workflows they already know.
Champions are also peers, not instructors. Employees tend to experiment more willingly when they're guided by someone doing a similar job rather than by someone from the corporate AI enablement team.
The newsletter Lead with AI notes that “champions succeed because they act as workflow translators, helping teams understand where AI fits, where it adds value, and how it reshapes decisions, handoffs, and accountability.”
More plainly, it says, "people copy people."
Enterprises investing in AI tools should also invest in teaching people how to use them. A prompt library is a shared repository of high-performing, high-quality prompts that gives employees a solid starting point, reducing the trial and error that could slow adoption.
When employees can see how a well-constructed prompt is structured—including the context it provides, the constraints it sets, and the output it specifies—they absorb those practices. It closes the gap between the most advanced users and everyone else.
A well-maintained prompt library also becomes a record of how the organization uses AI. As users refine prompts and share improvements, institutional knowledge compounds. But libraries can quickly go stale without active curation. The organizations that get the most value from them prioritize maintenance as a standard practice.
Many waves of enterprise technology came with a fluency benchmark: typing speed, spreadsheet proficiency, Boolean search skills. AI is different in at least one important way: The gap between basic use and expert use does not scale evenly, and it's invisible.
A worker who asks AI to summarize something and a worker who constructs a multistep reasoning task with specific output constraints and several iterations may be using the same tool, but they're not doing the same thing. In a usage dashboard, they look identical. That's the problem.
Section AI's AI Proficiency Report says, “In 2025, ‘AI proficiency’ meant something pretty basic. Do your people know how to use AI safely and write a decent prompt?”
Organizations have spent the last year focused on that bar, with predictable results: Employees now know what AI is and how to use it. But as AI advances, the bar is rising fast. Enterprises have optimized for the baseline. Now they need to set their sights on hitting the higher bar.
Find out how ServiceNow can help you put AI to work for people.
