AI competency levels: a four-level scale with observable behaviors

Four AI competency levels suit most employers. Each level should describe what a person does, in terms a manager can observe. "Understands prompting" is not a level. "Checks every figure in an AI draft against the source before sending" is. This guide gives a four-level scale with observable behaviors, and shows how to place people on it.

How many levels to use

Public scales range from three levels to seven. Each choice fits a different purpose.

• UNESCO's AI competency framework for teachers uses three levels: Acquire, Deepen and Create (UNESCO).
• The US Office of Personnel Management uses five: Awareness, Basic, Intermediate, Advanced and Expert. Each step changes how difficult a situation the person can handle and how much guidance they need (OPM).
• SFIA 9 uses seven levels of responsibility, from Follow to Set strategy, inspire, mobilize (SFIA).

Three levels leave too little room between a beginner and an expert. Seven levels suit a career framework that covers a whole working life. For a skill like AI use, four levels give enough steps to plan training. Managers can still tell them apart.

The OPM scale shows the pattern to copy. It does not define levels by what a person knows. It defines them by how hard a situation they can handle, and how much help they need to handle it (OPM). Both of those can be observed.

What should change from one level to the next

Each step up should change at least one of four things. Name the change in the level description, so two managers read it the same way.

• Independence: from following a template to working without one.
• Range: from familiar, well-defined tasks to new and unclear ones.
• Checking: from catching obvious errors to catching the subtle ones that matter.
• Reach: from improving one’s own work to improving the methods of a team.

SFIA uses a similar set of factors. Its level 3, Apply, describes people who handle varied work, some of it complex, using established methods under broad direction (SFIA). That description works because it names the type of work and the amount of direction.

A four-level scale

Adapt the wording to your own roles. Keep the behaviors concrete. Each list below describes what a reviewer would see in a work sample or an observed session.

Level 1: Foundational

• Uses an approved AI tool for familiar tasks, such as summarizing a document or drafting an email.
• Writes short requests and accepts the first or second result.
• Catches obvious errors, such as a wrong name or a broken format, but misses errors of fact.
• Knows which data must not go into an AI tool, and follows the policy when reminded.
• Needs a template or a colleague for anything new.

Level 2: Proficient

• Uses AI across most everyday tasks in the role without help.
• Gives the tool the background, examples and limits it needs, and gets a usable result in a few tries.
• Checks facts, figures and quotes against sources before the work leaves their hands.
• Produces work that meets the standard of the job without a reviewer fixing it.
• Applies the data rules without prompting, including on new tools.

Level 3: Advanced

• Handles unclear or new tasks, and decides when AI is the wrong tool.
• Combines several tools or steps into a workflow that others can repeat.
• Finds subtle errors, such as a reasonable-sounding claim with no source, or a calculation that uses the wrong period.
• Writes down how they checked the work, so a reviewer can follow it.
• Helps colleagues fix their requests and checks.

Level 4: Expert

• Sets the team’s methods: templates, checklists and review steps for AI-assisted work.
• Judges where AI adds value and where it adds risk, and can explain the reasons.
• Produces consistent results under time pressure and on high-stakes work.
• Spots when a tool change or a new model affects quality, and adjusts the team’s process.
• Teaches others, and their teaching shows up in the quality of the team’s work.

What a level is not

Several common measures look like levels but do not measure competence. Leave them out of the level descriptions.

• Hours of training completed. Training is an input. A level describes the output.
• Number of tools used. A person can use five tools badly or one tool well.
• How often a person uses AI. Frequent use without checking the work is a risk, not a higher level.
• Job title. A senior manager can be at level 1 in AI use, and a new hire at level 3.

How to place someone on the scale

1. Choose one or two real tasks from the person’s role.
2. Ask them to complete the tasks with the AI tools they would use at work.
3. Collect the final output and the drafts, and ask them to note what they checked.
4. Have a reviewer who does not manage the person score the work against the level descriptions.
5. Record the level with the date and the tasks used, so a later review can compare like with like.

Self-ratings help you plan the assessment, but they do not replace it. Use them to choose which tasks to test, not to set the level. The guide on AI skills survey questions explains how to run the survey side. The AI skills matrix shows behaviors for six competencies side by side.

Setting a target level per role

Not every role needs level 4. Set the target by the risk and volume of AI-assisted work in the role. A person who sends AI-drafted text to customers needs strong checking. A person who builds team workflows needs range and reach. Most office roles can aim for level 2 as a baseline, with level 3 for people who design processes or review others’ work.

Write the target into the role profile. Then plan training to close the gap between the current level and the target. A companion guide covers how to build an AI competency framework for employees, including the evidence for each level. Job seekers reading this may prefer AI proficiency levels explained.

How Emeri reports levels

Emeri uses a four-band scale: E1 Foundational, E2 Proficient, E3 Distinguished and E4 Exceptional. The certification assesses people on real work tasks and reports one band. The bands page describes each band, and the methodology explains how the grading works. The free check gives an indication of where a person stands in about 15 minutes.