How to build an AI competency framework for employees

An AI competency framework for employees has three parts. It names the skills you expect. It defines levels for each skill. It states the evidence that shows a person is at a given level. The third part decides whether the framework works. Without evidence, the framework is a list of words that managers read in different ways.

You do not need to start from nothing. UNESCO, the OECD with the European Commission, and the SFIA Foundation have each published a framework. None of them was written for a general office workforce. Each one still offers a structure you can borrow. This guide shows what to take from each, and how to fill the gaps.

Start with the work, not the technology

Write the skills as things people do at work. "Understands large language models" cannot be observed. "Checks every figure in an AI-drafted report against the source data" can. A manager can see the second one happen, or see it missing.

A useful test: could two managers watch the same employee and agree whether the skill was shown? If not, rewrite the skill until they could. Keep the list short. Five to eight skills is enough for most organizations. A longer list spreads attention thin and makes assessment slow.

For office roles, a workable set looks like this:

• Asking clearly: writing a request that gets a usable result on the first or second try.
• Giving context: supplying the background, data and limits the tool needs.
• Checking the work: finding errors, testing claims against sources, and knowing when not to trust an output.
• Building a workflow: turning repeated tasks into a set of steps that others can follow.
• Finishing real work: producing an output that meets the standard of the job, not a first draft.
• Keeping it safe: protecting confidential data and following the organization’s AI policy.

What the public frameworks offer

Each public framework solves part of the problem. Read them for structure, not for wording to copy into an employee handbook.

• UNESCO AI competency framework for teachers (2024): 15 competencies in five dimensions, including "Ethics of AI" and "AI for professional learning." It uses three progression levels: Acquire, Deepen and Create (UNESCO).
• UNESCO AI competency framework for students (2024): 12 competencies in four dimensions, with three levels named Understand, Apply and Create (UNESCO).
• AI Literacy Framework (European Commission and OECD): written for primary and secondary schools. It groups AI literacy into four domains: Engage with AI, Create with AI, Manage AI and Shape AI. A draft appeared in May 2025 (AILit).
• SFIA 9 (SFIA Foundation, October 2024): a skills framework for digital professionals with seven levels of responsibility. The levels run from Follow at level 1 to Set strategy, inspire, mobilize at level 7 (SFIA).

UNESCO's frameworks show a clean way to group skills into a few dimensions with a small number of levels. The AI Literacy Framework shows how to separate using AI from managing and shaping it. SFIA is the closest fit for technical staff. Its AI guidance names skills such as Machine learning (MLNG) and Artificial intelligence (AI) and data ethics (AIDE) (SFIA).

The gap is the same in all four. They describe what people should know and do. They say little about the evidence an employer should collect from an office worker. You have to write that part yourself.

Define the levels by what changes

Use three to five levels. Fewer than three gives no room for growth. More than five produces levels that nobody can tell apart. UNESCO uses three. SFIA uses seven, but most roles span only three or four of them.

Each level should change one or more of these things:

• Independence: does the person need a template or a colleague, or do they work alone?
• Range: does the skill hold on familiar tasks only, or on new and unclear ones too?
• Checking: does the person catch obvious errors, or the subtle ones that matter?
• Effect on others: does the person improve only their own work, or the team’s methods as well?

SFIA describes each level through autonomy, influence, complexity, knowledge, and business skills and behavioural factors (SFIA). Those factors map well onto the four changes above. A companion guide, AI competency levels with observable behaviors, writes out a full four-level scale.

Write the evidence for each cell

The public frameworks above leave this step to you. For every skill at every level, write down what you would accept as proof. Good evidence is something a reviewer can look at after the fact.

• A work sample: the person completes a real task with an AI tool, and a reviewer grades the output against a rubric.
• A before-and-after record: the draft the tool produced, the final version, and a note on what the person changed and why.
• A workflow document: the steps, prompts and checks a person built, which a colleague then follows.
• An observed session: a manager watches the person work through a task and records what they checked.
• A third-party assessment: a test scored by someone outside the team, against a published standard.

Self-ratings are not evidence on their own. In an Aalto University study, about 500 people answered logic questions from the LSAT, half of them with ChatGPT. The ChatGPT users scored better, and they overestimated how well they had done. Those who rated their own AI literacy higher overestimated more (Aalto University). Use a survey to learn who uses which tools. Use work samples to learn how well. The guide to AI skills survey questions covers the survey side.

Match levels to roles

Not every role needs the top level in every skill. Set a target level per role, and keep the targets few. A support agent may need a high level in checking the work and keeping it safe. An analyst may need a high level in building a workflow. A short role profile is easier to assess and easier to train toward.

Emeri publishes role guides that list the tasks behind each skill, for example AI skills for operations and AI skills for customer support. Use them as a starting point for your own role profiles.

Test the framework before rollout

1. Pick ten employees across three roles.
2. Ask two managers to place each person at a level for each skill, working separately.
3. Compare the ratings. Where the managers disagree by more than one level, the skill or level description is unclear.
4. Rewrite the unclear cells, and repeat with a new group.
5. Only then attach the framework to training plans, hiring or reviews.

Review the framework each year. AI tools change quickly, and a skill written around one product will age badly. Write skills around the task and the judgment, not the tool.

Where Emeri fits

Emeri assesses six competencies on real work tasks and reports the result as a band from E1 to E4. The AI skills assessment framework sets out how the six competencies are defined, and the bands page explains each level. Teams that want a baseline before writing their own framework can start with the free check, which takes about 15 minutes.