Short answer
AI is useful in a safety management system where people read, sort and search large volumes of reports and data: classifying incoming reports, linking duplicates, watching safety performance indicators and finding similar past events. Risk acceptance, investigation conclusions and accountability stay with the people your SMS names. Every AI output should be reviewable, traceable and easy to override.
Most aviation organizations that run a safety management system have the same problem. Reporting works. People submit hazard reports, occurrence reports and audit findings. The difficulty starts after that: someone has to read every report, code it, link it to a hazard, decide whether it is new, and notice when a category starts to rise. Safety teams are small, and the backlog grows.
This is where AI can help, provided it is introduced with the same care as any other change to a safety process. This guide maps realistic AI uses to the ICAO SMS framework and sets out what should not be handed to a machine.
The SMS framework in brief
ICAO Annex 19, Safety Management, requires service providers such as operators, approved maintenance organizations, air navigation service providers and certified aerodromes to implement an SMS. ICAO's Safety Management Manual, Doc 9859, describes the framework as four components:
- Safety policy and objectives: management commitment, accountabilities, key safety personnel, emergency response planning and SMS documentation.
- Safety risk management: hazard identification, and safety risk assessment and mitigation.
- Safety assurance: safety performance monitoring and measurement, management of change, and continuous improvement of the SMS.
- Safety promotion: training and education, and safety communication.
Annex 19 also sets principles for the protection of safety data, safety information and related sources. Any tool that touches reports has to respect those protections, and the reporting culture they support.
Where AI helps
The strongest uses sit in safety risk management and safety assurance, where the work is reading, sorting and comparing.
Report triage and classification
Language models can read a free-text report and suggest an occurrence category, the related hazards in your register, the affected department and an initial severity indication. The safety officer sees the suggestion beside the report and confirms or corrects it. Corrections are stored and used to improve later suggestions.
The benefit is speed and consistency. Reports are coded on arrival instead of in batches, and coding is less dependent on which person did it. The taxonomy should be the one your organization already uses, whether that is an internal list or a common taxonomy such as the ICAO and CAST occurrence categories.
Duplicate and related-event detection
The same event is often reported more than once, by crew, ground staff and maintenance, in different words. Text similarity models can flag likely duplicates and related reports so the team reviews them together. This also helps when the same hazard appears at different stations or on different fleets.
Trend monitoring
Safety performance indicators are usually reviewed monthly or quarterly. A monitoring model can watch the underlying data continuously and flag when a category moves outside its normal range, with the reports behind the change listed. The safety review board still decides what the trend means and what to do about it.
Search across past investigations
Investigators and safety managers often ask whether something similar has happened before. Search that understands meaning, not only keywords, can surface earlier reports, investigation findings and mitigations, with links to the original documents. This supports better risk assessment and avoids repeating mitigations that did not work.
Drafting routine outputs
Acknowledgements to reporters, monthly summaries and safety bulletins can be drafted from structured data and reviewed before they are sent. Drafting saves time only if review is quick, so the drafts should follow fixed templates.
| SMS component | Possible AI support | Decision that stays human |
|---|---|---|
| Safety risk management | Report classification, duplicate detection, search for similar past events | Hazard identification sign-off, risk assessment, risk acceptance, choice of mitigation |
| Safety assurance | Continuous monitoring of indicators, alerts on unusual trends, audit finding analysis | Interpreting trends, deciding corrective action, judging SMS effectiveness |
| Safety promotion | Drafting bulletins and lessons learned from reviewed cases | What is communicated, to whom, and how |
| Safety policy and objectives | Little direct use | Policy, objectives, accountabilities |
What must stay human
An SMS is built around accountability. The accountable executive and the people the SMS names are responsible for safety decisions, and that cannot move to software. In practice:
- Risk acceptance and mitigation. A model can rank or compare. It cannot accept a risk on behalf of the organization.
- Investigation conclusions. Causal analysis needs interviews, context and judgment. AI can organize evidence, not decide causes.
- Confidentiality decisions. Who sees a report and what is shared stays with the safety office, under your just culture policy.
- Final classification. Suggestions are confirmed by a trained person, at least until the organization has evidence that a category can be handled with lighter review.
A useful test: if the regulator asked who made this decision, the answer should be a named role, not a system.
Protecting reporters and data
Reporting depends on trust. If staff believe their reports will be analyzed in ways that could identify or blame them, reporting drops. Practical safeguards include:
- Removing or masking names and identifying details before analysis where the task does not need them.
- Running models on infrastructure approved for safety data, with no reuse of your data to train third-party models.
- Applying the same access rules to AI outputs as to the reports themselves.
- Telling staff, in plain words, how reports are processed and who reviews the results.
Introducing AI as a managed change
Adding AI to the SMS is itself a change to a safety process, so it belongs in your management of change procedure. A sensible sequence:
- Pick one process. Report classification is a good first candidate because results are easy to check.
- Run in parallel. For a defined period, the model suggests and staff classify as usual. Compare results.
- Set acceptance criteria. Agree in advance what level of agreement is needed before suggestions are shown to staff as defaults.
- Keep the override visible. Every suggestion can be corrected in one step, and corrections are logged.
- Review regularly. Include model performance in your safety assurance reviews, the same way you review any other tool.
What regulators are saying
Aviation authorities are actively working on AI. EASA, for example, has published an AI roadmap and guidance material for machine learning applications, focused mainly on AI inside aircraft systems and operations. Tools that support an organization's own SMS processes, with humans making the decisions, are a different and lower-risk use, but the same principles apply: know what the tool does, test it, monitor it and keep people accountable. Check your national authority's position before you rely on AI outputs in any formal submission.
A realistic first step
Take the last two or three years of your safety reports, with their final classifications. Use them to test how well a model would have coded them, which categories it handles well, and where it struggles. That exercise uses data you already hold, does not change any live process, and gives the safety review board evidence to decide whether to go further.