Korean Air: AI helps find maintenance records

Artificial intelligence (AI) can help a mechanic find an earlier maintenance case by describing the problem in ordinary words. That is the use Korean Air announced on 14 August 2026. To use a result, someone still needs to open the source and check that it applies to the aircraft and situation being examined.

What Korean Air announced

In its release of 14 August 2026, the airline describes a tool developed with LG CNS and AWS in around six months. A mechanic enters a symptom in their own words. The system searches for past work and similar cases.

It draws on around 90 database tables and millions of maintenance records. A table is an organised set of data within a database, such as a list of maintenance actions. This does not mean 90 independent databases.

The tool also brings abbreviations and free text into a common form. It extracts principal actions and fault areas. The airline describes analysis by aircraft model or registration, to find recurring faults and examine repair histories. Operational messages exchanged between aircraft and ground are also integrated.

Korean Air expects faster analysis. The release gives neither measured time savings nor an error rate. It does not present the tool as an authority that can approve maintenance or release an aircraft to service.

Why finding a case is not enough

Two reports may describe the same symptom using different words. A search that connects those descriptions can find a document that is hard to locate by its title. But similar symptoms do not establish that the same answer applies.

The earlier case may involve another aircraft, a different part or a manual version that has changed. A summary may omit an important condition. The result helps someone choose what to read; checking the document then establishes what can be taken from it.

For a European team subject to Part-M, the part of the rules covering continuing airworthiness, this distinction connects with an existing requirement. M.A.305(g) requires clear and accurate entries, with corrections that leave the original entry visible. This reminder concerns that European framework, not Korean Air’s own obligations. It does not prescribe an AI search engine.

Fictional case: a useful lead, a different aircraft

Kepler invented the following example in full. It retains a teaching scenario dated 8 September 2026. It is neither a Korean Air incident nor a test of its tool.

Léa is looking for the reason a warning light previously came on intermittently on fictional aircraft EX-204. The assistant suggests a report and summarises it as “similar case resolved”. When she opens the document, it concerns EX-203. It also refers to an old manual version and an appendix whose link no longer works.

Léa checks the current version. It covers configuration C2, meaning a particular equipment state of the aircraft. EX-204’s records show configuration C3. Simply changing the version number in the answer would not resolve the issue: the team must establish which instructions actually cover this aircraft.

Marc is assigned to this records review in the example. He keeps the earlier case as a lead but rejects the conclusion that it is already usable for EX-204. He requests the missing appendix and the reference applicable to configuration C3.

The search has advanced: the team knows what it found and what is missing. It has not yet established the technical answer. An inaccessible link alone proves neither that a document does not exist nor that the aircraft is unfit to fly.

Keep a note the next colleague can understand

For handing over such a search, Kepler suggests a simple note adapted to the organisation’s procedures:

• Question: locate the earlier warning report for EX-204.
• Source opened: EX-203 report, manual version cited and current version consulted.
• Discrepancy: different aircraft, different configuration and inaccessible appendix.
• Decision: historical lead retained; conclusion for EX-204 not accepted at this stage.
• Next step: request the appendix and the reference covering C3; Marc resumes the review after receipt and verification.

Also retain the displayed answer, its date and the tool version if available. A later search may rank results differently. A colleague should be able to understand the decision using the evidence visible at the time.

This note is a suggested working method, not a new regulatory form. Technical decisions remain with authorised personnel using applicable data and procedures.

Before trying a similar tool

Ask the supplier to follow a result through to its original document. Then check how the tool handles an old version, two different aircraft and a missing document. Can the user recognise uncertainty and continue checking without accepting the summary?

That is the useful criterion: find a lead more easily and be able to check what it establishes. Our guide to preparing an aviation AI trial helps structure that supplier discussion.