Measure AI Literacy in Employees [2026]

If you need to measure AI literacy in employees, you need more than training records and login counts. The EU AI Act Article 4 now requires providers and deployers of AI systems to ensure a sufficient level of AI literacy among their staff — and the only defensible way to demonstrate that is to measure what people can actually do with AI, not just what courses they completed.

This guide walks HR, L&D and compliance leads through the obligation, the measurement methods available, the tools on the market, and a concrete plan to build an evidence trail that holds up to scrutiny.

What It Means to Measure AI Literacy in Employees

Measuring AI literacy means assessing whether each person can use AI tools effectively, critically and safely in the context of their role — not whether they attended a workshop or opened a chatbot. It produces evidence of capability, not evidence of exposure.

That distinction matters because Article 4 of the EU AI Act does not ask you to train people. It asks you to ensure a sufficient level of AI literacy. Training is one input. Measurement is how you know whether the input worked.

Three things a literacy measurement should tell you:

1. Where each person stands right now — a baseline, not a guess.
2. Where the gaps are — by dimension (can they prompt well? can they spot errors? do they understand data privacy implications?) and by team.
3. Whether the level is sufficient — relative to the risk profile of the AI systems they work with.

Without measurement, "sufficient" is an assertion. With measurement, it is a finding.

Why Measure Now: The EU AI Act and the Literacy Gap

Article 4 of the EU AI Act has applied since 2 February 2025. The text is short and direct:

Providers and deployers of AI systems shall take measures to ensure, to their best extent, a sufficient level of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf.

This is not a future obligation. It is current law. And it applies to every organisation that deploys AI systems within the EU, regardless of where the organisation is headquartered.

The gap between confidence and capability

The problem is not that organisations are ignoring AI. It is that most have no idea whether their people are actually literate.

The ETS 2025 Human Progress Report: HR Edition, based on a survey of over 1,000 HR leaders, found that 82% of HR decision-makers are prioritising AI literacy for employees in the next 18 months — yet only 39% of U.S. employees felt proficient (43% globally). Meanwhile, 67% of those HR leaders say AI literacy is an important skill they look for in new hires (ETS, 2025).

A separate HiBob UK survey of 200 business leaders found that 77% expect AI proficiency to become a baseline requirement within two years (HiBob, 2025).

So the intent is there. The measurement is not.

Self-assessment is unreliable

Across 1,213 people who predicted their own AI literacy score before taking an AI literacy assessment, the average predicted score was 62.8 out of 100 while the average actual score was 43.5 — an overestimation gap of 19.3 points. Among students, the gap was 31.9 points. Even engineers overestimated by 15.4 points. Self-reported confidence is not a proxy for literacy.

If your current "measurement" is a self-assessment survey, you are almost certainly overestimating your workforce's readiness — and your Article 4 compliance position.

What Article 4 actually requires you to show

The regulation does not prescribe a specific assessment method. But it does require you to demonstrate that you took measures to ensure literacy, not just that you offered training. A regulator or auditor will look for:

• Evidence that you identified which staff interact with AI systems
• Evidence that you assessed their literacy level
• Evidence that you addressed gaps
• Evidence that the level reached is proportionate to the risk

Training completion certificates show the first step. They do not show the last three. For a deeper breakdown of the Article 4 requirements, see EU AI Act Training: Article 4 Checklist.

Four Ways to Measure AI Literacy in Employees

Not all measurement methods produce the same kind of evidence. The table below compares four approaches across what they capture, what they miss, and the strength of evidence they provide for Article 4 compliance.

Training completion records

Every LMS tracks completions. The data is easy to pull. But a completion record tells you someone sat through a course, not that they absorbed it. Two people can complete the same module and walk away with very different levels of understanding. For Article 4, completion records are necessary documentation but insufficient evidence of literacy.

Knowledge tests

Multiple-choice and true/false tests measure recognition and recall. They can verify that someone knows what a large language model is, or can identify a definition of prompt injection. They cannot verify that someone can write an effective prompt, detect a hallucinated citation, or make a sound judgement about when to trust an AI output. For a detailed comparison of test formats, see our analysis of why applied assessment outperforms multiple-choice.

AI usage analytics

Usage dashboards answer the question "are people using AI tools?" — which is a valid question, but a different one from "can people use AI tools well and safely?" High usage can coexist with poor practice. Low usage might reflect a thoughtful decision not to use AI for a particular task. Usage data is a useful complement to literacy measurement, not a substitute for it.

Applied assessments

Scenario-based and conversational assessments put the person in a situation and observe what they do. The best ones adapt to the person's level, cover multiple dimensions of literacy, and produce evidence that links each score to something the person actually said or did. This is the method that most directly answers the Article 4 question: does this person have a sufficient level of AI literacy for their role?

AI Literacy Assessment Tools: What's Available

Four tools are worth evaluating if you are building a measurement programme. Each takes a different approach.

ETS Futurenav Adapt AI

ETS Solutions launched Futurenav Adapt AI on 10 June 2025. It includes three assessments: Reflect (gauges employees' perceptions and frequency of using AI tools), Reason (an adaptive, scenario-based measure of AI knowledge), and Apply (scenario-based tasks that mirror workplace AI use and prompt engineering). The three-part structure separates self-perception from knowledge from application, which gives a more layered picture than a single test.

CodeSignal AI Literacy Assessment

CodeSignal's AI Collection includes an AI Literacy Assessment that covers foundational AI knowledge through interactive simulations, a Prompt Engineering Assessment, and an AI Researcher Assessment. It covers both technical roles and non-technical functions such as sales, marketing and HR.

Worklytics

Worklytics is a people analytics platform that measures AI adoption from usage data across tools such as GitHub Copilot, ChatGPT Enterprise, Claude Enterprise, Microsoft Copilot and Google Gemini, and connects usage to performance metrics. It measures how much people use AI, not an individual's literacy. It is most useful as a complement to a capability assessment — showing you adoption patterns alongside skill levels.

AISA

AISA is a 20–40 minute conversation with an AI interviewer that adapts each question to the previous answer, so it works for every role, technical or not. It produces a 0–100 score across five dimensions: Prompting & Communication, Critical Thinking, Technical Understanding, Workflow & Application, and Safety & Responsibility. Every score is linked to evidence from the person's own answers. Results hold up on retake — 98% of retakers land in the same or an adjacent tier, with a correlation of r = 0.79 between attempts. Every person gets a free full report and a free certificate. Results are integrity-verified.

For organisations, AISA provides a team dashboard with an organisation average, every person's report, and AI Fluency Analytics showing where the team has strength in depth, lone experts and blind spots. The scoring rubric is published and public. For a broader look at how AISA fits into employee assessment programmes, see AI Fluency Assessment for Employees.

A Step-by-Step AI Literacy Measurement Plan

A measurement programme does not need to be complex. It needs to be systematic. Here are five steps.

Step 1: Define the level each role needs

Not every role needs the same depth of AI literacy. A customer service agent using an AI-powered chatbot needs to understand its limitations and know when to escalate. A data scientist building models needs deep technical understanding. A compliance officer needs to understand AI governance frameworks and risk classification.

Map your roles to the AI systems they interact with, and define what "sufficient" looks like for each. Article 4 explicitly ties the required level to the person's role and the context of use. Section 6 of this post offers a qualitative framework for this.

Step 2: Baseline everyone

Run an initial assessment across the workforce — or at least across every team that interacts with AI systems. The goal is a clear picture of where people stand today, before any targeted intervention.

The AISA data across 2,123 completed assessments shows an overall average composite score of 46 out of 100, with the median also at 46. The weakest dimension across the population is Technical Understanding at 38.8, followed by Safety & Responsibility at 40.8. These are exactly the dimensions that matter most for responsible AI deployment and Article 4 compliance.

Step 3: Train to the gaps

Once you have baseline data, you can target training where it will have the most impact. If your marketing team scores well on Prompting but poorly on Critical Thinking, they need training on evaluating AI outputs — not another prompt engineering workshop.

This is where the measurement layer and the training layer work together. AISA is the measurement layer; your existing training providers, internal programmes, or platforms like Coursera, LinkedIn Learning or specialist AI literacy courses are the training layer. The baseline tells you what to train. The re-measurement tells you whether the training worked.

Step 4: Re-measure

After a training cycle (typically 3–6 months), re-assess. Compare scores to the baseline. Look for movement in the specific dimensions you targeted. If scores have not moved, the training did not work — and you need a different approach.

Re-measurement also refreshes your compliance evidence. AI tools and capabilities change quickly. A person who was literate six months ago may have gaps if their role now involves new AI systems or higher-risk applications.

Step 5: Keep the records

Article 4 compliance is an ongoing obligation, not a one-time exercise. Maintain a record of:

• Which roles interact with which AI systems
• Baseline assessment results (date, score, dimensions)
• Training delivered (content, date, participants)
• Re-assessment results
• Any remediation actions taken

This is your evidence trail. If a regulator asks how you ensured sufficient AI literacy, you can show a systematic process with measurable outcomes — not just a stack of training certificates. For a quality framework to evaluate your assessment approach, see the AISA Assessment Quality Framework.

What "Sufficient" Looks Like by Role

Article 4 does not define a universal threshold. It ties the required level to the person's role, the AI systems they use, and the context. Here is a qualitative framework for three broad categories.

Frontline and operational roles

These are people who use AI tools as part of a defined workflow — customer service agents using AI chatbots, warehouse staff interacting with AI-driven logistics, administrative staff using AI writing assistants.

Sufficient looks like: They understand what the AI tool does and does not do. They can recognise when an output looks wrong and know the escalation path. They understand basic data privacy rules — what they can and cannot input. They do not need to understand how the model works technically, but they need to know its boundaries.

In AISA terms, this maps roughly to competence in Prompting & Communication and Safety & Responsibility, with at least developing-level awareness in Critical Thinking.

Specialist and technical roles

These are people who configure, customise or build with AI — developers using code generation tools, data analysts using AI for pattern recognition, product managers specifying AI-powered features.

Sufficient looks like: Everything above, plus the ability to evaluate AI outputs critically, understand model limitations, construct effective prompts for complex tasks, and integrate AI into multi-step workflows. They should be able to identify when an AI output is plausible but wrong, and know how to verify it.

AISA data shows that even engineering roles average 54.4 out of 100, with Technical Understanding at 50.7 and Safety at 47.1. There is room to grow even among technical staff. For more on how AI certification compares to assessment for these roles, see our comparison.

Leadership and governance roles

These are people who make decisions about AI adoption, set policy, or oversee compliance — CIOs, Chief AI Officers, compliance leads, senior managers approving AI use cases.

Sufficient looks like: They understand AI capabilities and limitations well enough to make informed strategic decisions. They can evaluate vendor claims critically. They understand the regulatory landscape, including the EU AI Act's risk classification system. They can assess whether their teams have sufficient literacy and allocate resources accordingly.

AISA data shows that people motivated by leadership score an average of 51.4 — higher than the overall average of 46, but still in the Developing tier. Leaders who want to understand what the data says about leadership AI skills can explore the patterns in more detail.

---

Related reading: EU AI Act Training: Article 4 Checklist — the compliance requirements broken down into actionable steps.

Related reading: AI Fluency Benchmarks: 9 Personas [2026] — where different AI user types actually score, from Bystander to Architect.

Related reading: What Is an AI Competency Assessment? [2026] — how competency assessment differs from knowledge testing and usage tracking.

Frequently Asked Questions

How do you measure AI literacy in employees?

You measure AI literacy by assessing what employees can do with AI, not just what they know about it. Applied assessments — scenario-based or conversational — produce the strongest evidence because they test prompting, critical evaluation, workflow integration and safety behaviour in context. Training completions and usage analytics are useful supplements but do not measure capability on their own.

Does the EU AI Act require an AI literacy assessment?

Article 4 of the EU AI Act requires providers and deployers of AI systems to take measures to ensure a sufficient level of AI literacy among their staff. The regulation does not prescribe a specific assessment method, but you need evidence that the level is sufficient — and assessment is the most direct way to produce that evidence. The obligation has applied since 2 February 2025.

Is AI usage data a measure of AI literacy?

Usage data measures adoption, not literacy. It tells you how often someone uses an AI tool and which features they access, but it cannot tell you whether they use it well, evaluate outputs critically, or handle sensitive data appropriately. A person with high usage and poor judgement is a higher risk than a person with moderate usage and strong critical thinking. Usage analytics are most valuable when paired with a capability assessment.

How often should employee AI literacy be measured?

Most organisations should re-measure every 6–12 months, or whenever there is a significant change — a new AI system is deployed, a role's AI responsibilities expand, or a major model update changes tool behaviour. The EU AI Act's requirement is ongoing, not one-time, so your measurement programme should be too. Keeping dated assessment records creates the compliance trail regulators expect.