AI Fluency Levels: Defining What Good Looks Like by Role
AI literate is not a yes or no answer. Four defined levels, what someone at each level can actually do, and how to decide which roles need which.
Ask most organisations whether their staff are AI literate and you get a yes or a no. Neither is useful. Literacy is not binary, and treating it as binary produces the two most common training errors: buying one module for everybody, or buying deep training for the people who shout loudest rather than the people most exposed.
What works better is a defined ladder. Four levels, each described by what a person can actually do, mapped to roles by exposure rather than seniority.
Level 1 — Aware
Can: recognise when they are interacting with an AI system, including where it is embedded in software they already use. Knows the organisation acceptable-use position. Knows what must never be entered into an external tool. Knows who to ask.
Cannot yet: judge whether an output is reliable.
Who needs it: everyone. This is the floor, it is short, and it is deliverable internally.
The most common gap at this level is not attitude, it is detection — people do not know which of their existing tools now contain an AI component, because it arrived in a vendor update nobody announced.
Level 2 — Capable
Can: use AI tools for real work. Frames a task well, recognises when an answer is thin, verifies facts and references before relying on them, and knows which tasks are a poor fit.
Cannot yet: design how a team should use AI, or assess a tool before it is bought.
Who needs it: anyone producing work with AI assistance — drafting, summarising, analysis, research, coding. In most organisations this is the largest group and the one receiving the least appropriate training, because generic productivity courses stop at prompting and never reach verification. Our guide to checking AI output is the core skill at this level.
Level 3 — Accountable
Can: act on AI output that affects someone else — a credit decision, a recruitment screen, a clinical pathway, a case priority. Understands what the system is doing well enough to exercise meaningful oversight, knows the limits of that oversight, and knows when to override or escalate.
Who needs it: a small group, usually much smaller than people assume, and usually not the group the training budget went to. These are the roles where the regulatory obligation is heaviest and where a generic module is plainly inadequate.
The test for this level is not whether someone can use the tool. It is whether they can explain, to a person affected by the decision, what part the system played and what part they played.
Level 4 — Governing
Can: decide whether a system should be deployed at all. Classifies systems, assesses vendors, owns the inventory, maintains the training record, signs off on controls, and understands how the obligations interact with the rules already in place.
Who needs it: one or two people, plus whoever on the senior team signs. Boards are frequently missed here — directors who approve AI-informed decisions often do not consider themselves users of AI at all.
How to map roles to levels
Map by exposure, not grade. Three questions per role:
- Does this role act on AI output that affects a person outside the organisation? If yes, level 3 minimum, regardless of seniority.
- Does this role produce work with AI assistance? If yes, level 2.
- Does this role decide what gets deployed or bought? If yes, level 4.
Everyone else is level 1. A graduate analyst scoring applications sits above a director who only reads summaries — which is counter-intuitive and usually correct.
Why a ladder helps with the legal position too
The EU AI Act asks for AI literacy appropriate to the role and the context, not a uniform standard. A defined ladder with roles mapped to levels is a direct answer to that, and it is far stronger evidence than a list of people who completed a course. The July 2026 amendment made clear that no specific level must be guaranteed in any individual — but the duty to take measures is unchanged, and a ladder is the clearest way to show the measures were proportionate. See our guide to the Article 4 training duty.
Common mistakes
- Training to level 2 across the board. Expensive, and still leaves level 3 roles under-prepared.
- Assuming seniority equals exposure. It routinely does not.
- Treating the ladder as fixed. People move up it, and tooling changes what a role is exposed to. Re-check annually.
- Measuring completion instead of capability. A completion rate tells you who sat through something. It does not tell you anyone reached a level.
Frequently asked questions
Is there an official AI fluency framework we must follow?
No. There is no prescribed framework and no approved standard under the AI Act. A defensible internal ladder is sufficient, and arguably better, because it reflects your actual systems.
How long does each level take?
Level 1 is short. Levels 3 and 4 are substantial and ongoing. Beware anything selling all four in a single day.
Should we test people?
For levels 3 and 4, some form of assessment is worth having, because the whole point is demonstrable capability rather than attendance.
Does a certificate prove a level?
Only if it was assessed. See our guide to self-awarded versus accredited certificates.
Our CPD courses are built around roles rather than a single general module, and our guide to tailoring AI training by role covers how to scope the map.
This page was last updated:
Learnsignal Education Team
Expert Tutor at Learnsignal
Qualified professional with years of experience helping students advance their professional careers.
View all posts by Learnsignal Education Team
