AI Assistance vs AI Decision-Making: EASA's AI Levels Explained for Aviation Work

An engineer at ExampleMRO asks an assistant to assemble a defect review pack. Finding the source paragraphs is one task. Presenting options for a human decision is another. Taking the decision itself changes the assistant's role again.

EASA's AI levels describe these differences in work and authority. They are not a ranking of how clever a model is, or a certificate that a product is safe to use.

Which EASA documents are we reading?

EASA published Concept Paper Issue 2, Guidance for Level 1 & 2 machine-learning applications in March 2024. The RMT.0742 terms of reference use that paper as the basis for rulemaking on AI trustworthiness.

The table below follows Proposed Issue 3, June 2026, Table 4, page 31. This is proposed guidance, not law. The C1/C2 discussion comes from September 2026 workshop slides and is also draft material. Sources were checked on 24 September 2026.

EASA AI levels, from information to automation

These are paraphrases of Proposed Concept Paper Issue 3, Table 4. Responsibility remains with the end user at Levels 1A, 1B and 2A; the proposal calls for a responsibility-scheme assessment at 2B and 3A. Do not describe every level as a human checking every decision.

What is in RMT.0742's scope?

The September 2026 roadmap presentation describes the consolidation scope as supervised machine learning for Levels 0, 1 and 2A. Levels 2B and 3 belong to the broader exploration shown alongside it. That is a statement of the work programme, not permission to deploy a tool. See AI Days 2026, Day 1 morning, slide 12.

Keep the version distinction visible: the earlier NPA 2025-07(B), proposed DS.AI.010 describes a Level 1/2 scope, including cooperation and collaboration, subject to risk and technology exclusions. The later workshop direction does not amend that published proposal into law. Neither should be presented as an enacted general rule for office AI.

Where lookup and drafting fit

Applied to office work, a source lookup that helps a person acquire information is a Level 1A-style use. A pack comparing options so a person can choose a course of action is Level 1B-style support. Drafting may support either task; its label depends on the intended use and interaction, not the presence of a text box.

Those examples are our interpretation of the framework, not an EASA classification of Avioverse. A system that can also act needs assessment of its full intended capability; calling one screen “assistance” does not settle its level.

C1 and C2: two proposed paths for organisational use

EASA's organisation workshop first asks whether incorrect performance of the activity could increase the relevant risk. For activities presenting that risk, its draft branches are:

• C1 — verify outputs: detect and address errors before use, through systematic personnel checks or automated checks by non-AI software. Define acceptance criteria and error handling.
• C2 — qualify the system: where that output-checking process is absent, establish performance objectives and qualification evidence, with in-service monitoring.

C1 also includes organisational procedures, training, human-factors and ethics assessments; it is not simply an “accept” button. C2 carries organisational adaptations as well as qualification. These are workshop proposals, not current legal duties. See AI Days 2026, Day 2 morning, slides 23–27.

What stays human in our aviation-workbench examples

The levels do not give an office assistant operational authority. In the ExampleMRO workflow:

• MEL deferral: AI may organise the defect facts and supplied MEL conditions; the authorised process decides the deferral.
• Certificate of release to service (CRS): AI may flag missing work-pack information; authorised personnel issue the release.
• Finding classification: AI may prepare evidence and a possible classification; the responsible reviewer confirms it.
• Risk acceptance: AI may assemble hazards, controls and open questions; the authorised person or group accepts the risk.

These are the boundaries of the preparation workflow described here, not a claim that EASA prohibits automated decisions at every AI level. The working rule is the one the other role articles point here for: the tool prepares, a person decides. For the practical handoff from draft to checked work, use the human review gate. Avioverse's role is to help prepare that work; this article claims no EASA approval or qualification for the product.