Meet eAvy: the agentic
safety copilot.

See the highlighted block above the contents list. The rest of this article explains what makes a copilot agentic, what eAvy reads and proposes, and the hard line it is built never to cross.

Almost every safety platform now has “AI.” In most of them it is a single button that reads the words on the screen and hands you a paragraph. It is useful for ten seconds and forgotten by the next record. It does not look at anything you have not opened, it does not act, and it leaves no record of who agreed with it.

An agentic copilot is a different kind of tool. Three things separate it from a summarize button:

• It reaches across the operation. It does not stop at the text on screen. It reads the records linked to the occurrence in front of you and searches the rest of the operation for events that look like it.
• It takes several steps.It is multi-turn: it can read, then follow a link, then compare, then draft — the way a colleague would work a problem — rather than returning one canned answer.
• It proposes the next move, under approval. It does not just describe; it drafts the actual next step. And because it can propose real changes, every one of those changes is held behind a human decision that is recorded.

That last point is the one that matters most for a regulated operation. A summarize button is safe because it cannot do anything. A copilot that can draft a classification or a corrective action has to earn that reach with oversight you can defend in an audit. eAvy is built around exactly that trade: more reach, matched by a hard approval gate on every change.

Open eAvy on an occurrence and it starts where a good safety analyst would. It reads the report itself, then it pulls the thread.

• The occurrence and its linked records.Not just the report text — the hazards, findings, actions and other records already connected to it. eAviora keeps records as one connected operation, so the links are there to follow.
• Look-alike occurrences.It searches the operation for events that resemble the one in front of you. The three similar approaches from last quarter, the recurring ground-handling event — the pattern a busy team does not have time to go hunting for.
• The barriers and corrective actions on the trail. From the occurrence it follows the links out to the barriers that were meant to catch the event and the corrective actions already raised against them. That is how it understands not just what happened but what was supposed to stop it and what is already being done.

Because it reads across the linked records rather than a single page, eAvy can answer the operational questions a summary never could: is this part of a pattern, which barrier let it through, and is there already an action in flight?

Reading is half the job. The reason a copilot beats a search box is that it drafts the next step for a person to approve. With the picture assembled, eAvy proposes — it never finalises.

It drafts the next step. A suggested classification for the occurrence. A link to the look-alike events it found. A corrective action aimed at the barrier that failed. Each arrives as a draft on a one-click approval card, with the confidence eAvy attached to it shown plainly so the reviewer can weigh it before deciding.

Behind the copilot, eAviora also runs a small fleet of specialist safety agents, each focused on one job and bound to the stage of the workflow where that job belongs:

• A root-cause analyst — works the chain of contributing factors behind an event.
• A risk assessor — proposes severity and likelihood so the event lands in the right risk band.
• A corrective-action proposer — drafts the action aimed at the weakness the event exposed.
• A compliance-gap detector — flags where a requirement looks unmet.

These specialists do focused work at the right point in the process and feed their proposals into the same approval cards. eAvy is the copilot you talk to; the specialist agents are the expertise working quietly behind the stages. In both cases the rule is identical: the AI suggests, a person decides.

An agentic copilot is only safe for a regulated operation if it is built to stop in the right places. In eAviora those limits are not settings you can toggle off — they are how the system works.

• Every write is a one-click approval. eAvy never changes a record silently. Each proposed change pauses as an approval card. The operator accepts it, modifies it, or rejects it.
• Every decision is logged.Accept, modify or reject — the outcome, the person and the confidence the AI reported are recorded as a defensible oversight trail. When an auditor asks who agreed to a classification and on what basis, the answer is on file.
• Low confidence goes to a human. When eAvy is unsure, or its suggestion falls outside your agreed catalogue of classifications and values, the output is queued for a person to handle. It is never quietly written.
• The AI cannot move or close a record.The steps that advance a record through its workflow, and the governance sign-offs that close it, can never be set by the AI — and never through the API. A human moves the record forward.

This is the difference between a demo and a system you can put in front of a regulator. The reach is real, but the safety-critical decisions stay with named people, and the trail proves it.

Reach across the operation is powerful, which is exactly why it has to respect who is allowed to see what. ICAO Annex 19 protects the confidentiality of safety reports so people keep reporting. eAviora enforces that protection on every read — and a copilot does not get an exception.

A confidential report stays invisible to an analyst who does not have the clearance to see it. eAvy reads only what the person using it is already allowed to read, so it cannot surface a protected report to someone who should not see it. There is no back door through the assistant: the AI inherits the same confidentiality the person inherits, on every single read.

That is the contract that makes an operation comfortable letting a copilot reach across its records. The reporting culture Annex 19 is meant to protect is not weakened by adding intelligence on top of it — it is the boundary the intelligence respects.

More on how eAvy fits the wider AI picture in AI in aviation safety management, the modules it works across in the SMS module and safety analytics, and the oversight and data commitments on the trust page. To see it on your own operation, talk to us.