AI in aviation is best understood one task at a time. Forecasting where contrails may form, identifying maintenance patterns and helping a person review information have different inputs, consequences and approval needs. Start with the decision a system supports, then examine the evidence for the claimed improvement. This guide provides two documented examples and a repeatable way to assess new claims.
On a narrow screen, scroll tables sideways. Keyboard users can focus a table and use the arrow keys.
Name the task and the decision
Ask the supplier or researcher to identify the input, output and user. Does the system predict a future condition, classify a record, recommend an action or execute it? Find out which part uses machine learning. A dashboard, a conventional optimization method and a generative assistant can all appear in the same workflow while performing different jobs.
| Question | Evidence to seek | What to record |
|---|---|---|
| What changes? | A clearly defined decision and baseline process | Task, user and claimed benefit. |
| How was it tested? | Relevant flights, aircraft, conditions and comparison group | Population, dates, sample size and exclusions. |
| What improved? | A defined outcome with uncertainty and tradeoffs | Metric, units, denominator and side effects. |
| Who reviews it? | Human responsibilities and escalation path | Reviewer, authority and fallback. |
| What is approved? | Evidence covering the stated installation and operation | Jurisdiction, approval scope and limitations. |
Use these questions before comparing percentages. A reduction in maintenance alerts, cancellations or contrail formation measures different things. A supplier’s estimate can be informative when its method and scope remain visible.
Example: contrail-avoidance research
A March 2026 research preprint reports a randomized trial of dispatcher-led contrail avoidance integrated into standard flight planning. Its abstract reports an 11.6% reduction in contrail formation rate for 1,232 flights marked eligible for avoidance, relative to controls. For the 112 flights that flew the avoidance plan as intended, the reported reduction was 62.0%. The authors found no statistically significant fuel-use difference between the groups.
Those figures answer different questions. The eligible-flight result captures the intervention across its assigned population; the completed-plan subset describes a narrower group. Keep both populations visible when explaining the result. Contrail formation rate also differs from total aviation climate impact. The study is a dated preprint, and its observed fuel result leaves uncertainty rather than establishing zero fuel cost in every operation.
Google’s accompanying explanation describes the flight-planning integration and the broader trial. The research paper is the better place to check the analysis and definitions behind a headline percentage.
Example: predictive maintenance
In its October 2024 account of Skywise Fleet Performance+, Airbus describes predictive models using machine learning and natural language processing. It reports that easyJet avoided 44 cancellations in July 2024 and 35 in August after using the service and taking proactive maintenance action.
That is a supplier-reported operational example. To evaluate transfer to another fleet, ask how an avoided cancellation was estimated, which components were monitored, what maintenance action followed and what effort false alerts created. Keep the time period and operator attached to the numbers. Deployment evidence and an independently controlled estimate of benefit have different strengths.
The useful workflow connects an indication to qualified engineering review, approved maintenance activity and a recorded outcome. A predictive score becomes valuable through that complete process. Ask whether the organization measures missed problems as well as successful warnings.
Treat approval and monitoring as specific questions
EASA’s AI publication library includes its 2024 Issue 2 concept paper and a proposed Issue 3 published June 3, 2026. Check the document’s status and applicability. A concept paper or roadmap describes guidance development; an individual product needs evidence appropriate to its intended use.
For a purchase or evaluation, request a written account of responsibility when the model is uncertain, unavailable or wrong. Ask how data changes are detected, who authorizes updates and how performance is monitored after deployment. Keep sensitive operational information within the organization’s approved systems.
A useful one-paragraph assessment can state: “This system supports [task] for [user]. Evidence from [population and dates] found [defined outcome] against [baseline]. The remaining questions are [limits], and [role] reviews the output.” If a claim cannot yet fill those fields, record the gap and seek the missing evidence before relying on it.