Four levels of proficiency
Levels describe how a person works, not how long they have worked. A person may be Advanced in one domain and Aware in another. Levels are not seniority, not job titles, and not years of experience.
Each level links to a filtered view of every statement written at it.
| Dimension | L1 Aware | L2 Applied | L3 Proficient | L4 Advanced |
|---|---|---|---|---|
| Autonomy | Works under direction. Recognises when to escalate. | Works to established procedure with occasional guidance. | Works independently. Sets own approach within policy. | Sets the policy others work within. |
| Complexity | Familiar, well-defined situations. | Routine situations with some variation. | Non-routine situations with competing constraints. | Ambiguous situations with no established practice. |
| Judgement | Follows given rules. | Applies rules and recognises edge cases. | Weighs trade-offs and justifies choices. | Decides where no adequate rule exists, and creates one. |
| Scope of impact | Own task. | Own work and immediate team. | Function or product line. | Organisation, sector or profession. |
| Accountability | For own output. | For own output and its documented basis. | For work reviewed, and for method. | For governance, standards and outcomes at scale. |
Aware
The person can participate in work involving AI without being a hazard to it. They recognise what a system is doing, know the vocabulary well enough to follow a conversation, and — most importantly — know the limits of their own competence and escalate rather than proceed. L1 is not a low-value level. Most organisational harm from AI comes from people at L0 acting as though they were at L2.
Applied
The person does the work. They use AI systems to produce results in routine situations, following established procedure, and they document what they did well enough that someone else can check it. They recognise when a situation has left the routine and refer it upward. L2 is the level at which most professional roles operate and the level most credentials target.
Proficient
The person decides how the work should be done. They design approaches rather than follow them, weigh competing constraints, justify choices to others, and review the work of people at L2. They are the level at which a novel problem first gets a considered answer rather than an escalation. L3 carries responsibility for method, not merely for output.
Advanced
The person sets the standard others work to. They operate where no adequate practice exists and establish one, and they hold accountability for outcomes across an organisation, a sector or a profession. L4 is rare and should be treated as such. A framework whose Advanced level is widely claimed has miscalibrated it.
Two tests, stated explicitly
Every disagreement about levelling will occur here, so the tests are published rather than left to interpretation.
The test is not difficulty. It is who decided the approach.
A person performing a complex task by following a procedure someone else designed is at L2. A person performing a simpler task, having decided the approach themselves and being able to justify it against alternatives, is at L3.
Could this person review someone else’s work in this domain and say why it is wrong?
If yes, L3. Producing correct work is L2; recognising and diagnosing incorrect work is L3.
The reserved fifth level
An L5 is reserved and deliberately undefined. The intention is a level describing contribution to the state of the art — establishing practice that did not previously exist in the field, rather than within an organisation.
There is not yet sufficient evidence about what that looks like in AI competence to define it responsibly, and defining it prematurely would produce a level that credentials claim and nobody can assess.
It is reserved rather than omitted so that later definition does not require renumbering.
Every statement opens with a verb from its level
This is the mechanism by which L2 in one domain means the same thing as L2 in another. Without it, levels drift domain by domain and the framework’s central claim fails.
These describe an internal state rather than an observable action and cannot be assessed.