EASA Safety Management AI Prepares the Work

A pilot report mentions workload. A ramp supervisor notes confusion during a turnaround. Maintenance raises a handover concern. Training reports a repeated misunderstanding. Compliance monitoring finds a procedure that people are not following.

That is an ordinary week for a safety manager: partial information, from different sources, at different levels of confidence. None of it becomes a safety conclusion just because it can be summarised neatly.

That is where AI fits: preparation before judgement. It can organise notes, extract themes, prepare hazard discussions, build SRB packs and draft action tables. It should not accept risk, approve mitigations, assign final severity, close actions, decide cause or become the official safety reporting system. Humans classify, judge, approve and accept risk through the organisation's safety process.

Start by separating what is known

Safety work begins with messy input. Some information is formal. Some is a line in a meeting note. Some is confidential. Some is a weak signal that may never become a hazard, and some is only rumour that needs careful handling. The first job is not to sound certain. It is to sort the material.

AI can prepare a review note that separates:

• observed facts;
• reported statements;
• assumptions;
• missing information;
• possible hazard areas;
• affected operation or role;
• existing controls;
• records needed;
• questions for the safety manager.

That structure slows the process down in a useful way. A polished paragraph can make a concern look more mature than it is. A structured note keeps the uncertainty visible.

"Crew workload caused the event" is a conclusion. "The report mentions high workload during the sector; roster data and crew statements have not yet been reviewed" is a preparation note. The second one tells the safety manager exactly what to do next.

Occurrence triage needs careful language

Early triage is where wording causes the most trouble. A first note may say "unstable approach," "handover missed," "tooling issue" or "fatigue concern." Those phrases may matter, but they are not the whole event. The safety manager needs context: aircraft, phase of operation, source of report, immediate action, people involved, records available, repeated history and any urgent containment need.

AI can prepare a triage note with fields such as:

• reported event or concern;
• source and date;
• known facts;
• unconfirmed statements;
• immediate action already taken;
• possible hazard area for review;
• missing information;
• decision needed from the safety manager.

It should suggest questions, not conclusions. "Does this require immediate escalation?" is safer than "This is a high-risk event" when the evidence has not been reviewed. The occurrence reporting guide covers the reporting side that triage feeds.

Hazard notes need operational texture

Hazard records usually start as rough observations: poor lighting on a ramp, unclear maintenance handover, repeated ground equipment defects, "dispatch keeps missing things." Those statements may point to a real issue, but the safety team needs a condition it can review.

AI can draft that first version from safe inputs or approved summaries. It can remove emotional wording, flag vague claims and separate the hazard from the consequence and the control. A stronger hazard note can include:

• short title;
• source or context;
• condition being reported;
• operation affected;
• possible consequence;
• existing controls for review;
• available evidence;
• uncertainty or limits;
• proposed owner for review;
• next information needed.

That structure helps the safety manager avoid two common errors: ignoring a useful early signal, or treating an untested concern as a proven hazard.

The tool prepares the wording. The safety manager decides whether it enters the hazard log, how it is described, who is involved and what risk assessment is required. AI should not assign final severity or likelihood. It can prepare prompts such as "review whether current controls cover night operations" or "confirm whether this affects contracted ground handling," which keeps the output useful without turning it into an automated risk assessment. How the likelihood and severity scales themselves are calibrated is covered in the aviation risk matrix.

Occurrence summaries protect uncertainty

Summaries help when they make the event easier to review. They are dangerous when they flatten it.

A management note may turn "reported concern" into "confirmed issue." It may drop the sample size or leave out that the investigation is still open. If the reporter says the event may have involved fatigue, the summary should not say fatigue caused it. If there is one occurrence, it should not be called a trend. If an action is proposed, it should not be described as implemented.

AI can prepare an occurrence review note with:

• event context;
• simple timeline;
• known facts;
• reported statements;
• immediate response;
• affected process or role;
• possible contributing areas for review;
• records still needed;
• action status;
• decision required.

That gives the safety manager and operational leaders a clearer starting point, and it saves the Safety Review Board from spending the first half of a meeting working out what is known and what is assumed. The follow-up that comes after, from analysis to action, is covered in From Occurrence to Action.

Official occurrence reports, confidential statements, investigation records and safety data remain in the approved safety management system. A personal workspace should not become a shadow occurrence database.

Classification is preparation, not the decision

AI can sort safety information into possible categories for review. It can group items by topic, department, aircraft area, human factors theme, training issue, procedure weakness, fatigue concern, supplier interface or repeated defect. It can suggest whether an item may need investigation, monitoring, escalation or closure review.

Those suggestions stay drafts. The safety manager confirms the category, severity, likelihood, risk level, escalation path and action requirement according to company procedures. A tool may recognise similar words in several reports, but it does not know the full operational context unless humans and approved systems provide it.

AI prepares the sorting. People with authority decide the significance.

Risk review preparation is not risk acceptance

Risk review belongs to people with operational knowledge and defined authority. They understand current pressures, crew or staff capability, equipment status, supplier performance, procedures, training and the consequences of changing a control.

AI can still prepare the pack. A useful risk review pack can show:

• topic and reason for review;
• background and recent events;
• current controls;
• evidence available;
• evidence missing;
• affected roles or locations;
• previous actions;
• possible options for discussion;
• decision requested;
• proposed effectiveness check.

This matters most when a safety topic crosses departments. A turnaround issue may involve ground operations, flight operations, load control, contracted handling, training and compliance monitoring. A prepared pack shows those interfaces before the meeting, so the review does not shrink into a narrow discussion about one paragraph in a procedure.

The final risk assessment, acceptance, mitigation approval and review interval remain with the approved safety process. AI prepares the room; it does not make the decision in the room.

Safety actions need proof of work

"Brief all staff" is the classic weak action. Who writes the briefing? Who receives it? Is it a one-off notice or a training change? How is completion recorded? Which procedure changes? Does it apply to contractors? Who checks whether the control worked?

Action trackers can look tidy while still being weak. An action may have an owner and a due date but no completion evidence. It may be marked complete before the mitigation is tested. It may sit overdue for three meetings with no escalation. It may close the immediate issue and leave the wider hazard untreated. "Monitor" and "review process" have the same problem as "brief staff": nobody can say what done looks like.

AI can turn weak action wording into a draft action table for review:

This helps safety, compliance and training work from the same picture, and it avoids closure based only on an email being sent or a sentence being added to a tracker.

Closure requires evidence and review. The safety manager, responsible manager or board confirms whether the action is complete and whether the residual risk remains acceptable.

SRB packs should help leaders decide

A weak SRB pack wastes the meeting. Open actions are unclear, mitigations are listed without effectiveness notes, trends are mixed with single events and the decision required from the board is buried in text. The opposite failure is just as bad: uncertainty hidden behind traffic-light colours and short labels.

AI can group open actions, overdue investigations, hazard reviews, occurrence themes, training impacts, compliance monitoring links, supplier issues and resource questions, then draft a one-page summary for each topic that needs board attention. Each issue summary can show:

• why the topic is on the SRB agenda;
• current status;
• evidence available;
• uncertainty;
• existing controls;
• options or questions for discussion;
• decision needed;
• owner after the meeting.

This is where a preparation tool saves real time. The board spends less time decoding the pack and more time deciding what the organisation will do. It still owns those decisions: AI does not approve mitigations, allocate resources, accept risk or declare safety performance acceptable. What the rules say about safety performance indicators, and who reviews them, is in the SPI guide.

SMS and compliance monitoring should meet

Safety management and compliance monitoring feed each other. An audit finding may show that a control is not implemented. A safety report may show that a procedure is not practical. An overdue corrective action may increase exposure. AI can prepare short link notes between the two, for example a safety report linked to a procedure gap, an occurrence theme linked to repeated audit observations, or a hazard action that needs a manual amendment. That keeps teams from running parallel trackers that never meet. It does not decide significance: safety and compliance leaders decide whether an item needs escalation, board review or management action. AI for EASA management systems covers how those links fit into the wider system.

Confidentiality protects reporting culture

Safety information is sensitive because people must trust the system. Reports may include personal data, operational errors, just culture considerations, customer information, regulator communication, investigation material or proprietary controls. That information belongs in approved company systems with controlled access, retention and audit trails. It should not be copied into an uncontrolled personal AI layer.

A personal layer can still be useful. It can hold public learning notes, generic hazard templates, non-confidential SRB pack structures, personal certificates and safe question sets. It helps a safety professional prepare how to think, without storing the company's controlled safety data.

The boundary protects trust. If people believe safety reports are being fed into uncontrolled tools, reporting culture suffers. Just culture and reporter protection explains why that trust is worth protecting.

What changes for a safety team

Clearer hazard notes. More careful occurrence summaries. Action tables that name their evidence. Risk review and SRB packs that lead to a decision instead of a discussion about what the pack means. Assumptions stay visible, and the prompts to check training impact, manual amendments, compliance monitoring links and effectiveness evidence are already there.

That gives safety managers, post-holders and review boards more time for judgement.

Official records stay in company systems. For safety management, that is not a limitation. It is the control. Where a draft stops and a person decides is in assistance versus decision-making.