Annex 19 now wants "safety intelligence." That is not the same as another dashboard.

On 26 November 2026 the second amendment to ICAO Annex 19 becomes applicable. It was adopted on 23 June 2025 and it does a few concrete things: it pushes the State Safety Programme harder, it extends formal SMS obligations to certified RPAS operators flying internationally and to certified heliports, and it introduces a new chapter on what it calls "safety intelligence," supported by Doc 10159. The direction is plain enough — away from compliance you can tick, toward decisions you can defend with data.

I think the direction is right. I have spent twenty years watching safety systems pass an audit on Tuesday and fail an operation on Thursday, and a regime that asks "what is your data actually telling you" is a better regime than one that asks "is the binder up to date." So far, so good.

But watch what is already happening to the word. "Safety intelligence" is being read, by a remarkable number of vendors, as a product you can buy — a dashboard, a model, a feed of green and amber tiles that promises to find the next accident for you. And that reading is wrong in a way that matters, because it quietly moves the accountability from a person to a piece of software that cannot hold it.

What the amendment is actually asking for

Read Chapter 5 and Doc 10159 as what they are: a requirement to turn safety data into safety decisions, with the governance and the Just Culture protections that make the data trustworthy in the first place. The strengthened safety-performance requirements — indicators, targets, continuous monitoring — are the machinery. The intelligence is what a competent human does with the machinery's output.

An AI is genuinely good at the machinery. It will cluster your occurrence reports, surface a drift in a leading indicator weeks before you would have noticed it by eye, and draft the monthly summary faster than any of us. I use these tools and I would not go back. The floor on routine analysis has risen, and that is a gain for everyone.

Where it stops is the part Annex 19 actually cares about. A model can tell you that reports of a particular taxiway incursion are up forty percent this quarter. It cannot tell you whether that is because the risk got worse or because you finally built a reporting culture where people stop hiding it — and those two readings call for opposite responses. One is a hazard. The other is the system working. Knowing which one you are looking at is judgment, and judgment is exactly what does not arrive in the box.

The accountability the model cannot carry

Here is the test I apply to any tool that claims to do safety intelligence: when it is wrong, who answers for it?

A safety manager signs the safety case. An accountable executive carries it to the regulator and, in the worst case, to a commission of inquiry. The model that produced the confident wrong paragraph carries nothing — it cannot be cross-examined, it cannot be held to account, and it will generate the same plausible output for the operator who can evaluate it and the operator who cannot. In a regulated, safety-critical field, that asymmetry is the whole problem. The buyer who most needs to know whether the output is correct is, by definition, the buyer least able to judge it.

So the honest way to read Amendment 2 is this. The regulation is raising the value of the human in the loop, not lowering it. The data work gets cheaper and faster, which means the scarce thing — the judgment that the data is being read correctly, tailored to your operation and your authority, and signed by someone who will stand behind it — gets more valuable, not less. The states that treat "safety intelligence" as a procurement line will produce very tidy dashboards and learn nothing. The ones that treat it as a discipline, with a named, accountable person reading the instruments, will get the thing the amendment was written to produce.

Buy the machinery, of course. It is good and it is here. But do not mistake it for the intelligence. An AI cannot be accountable for a safety decision. A person can — and from November, the regulation will expect one to be.

---

Robert Strauss holds a PgD in Aviation Safety Management from City, University of London, and has spent more than 20 years across airports, air traffic control, flight academies and operators. Sources: ICAO Annex 19, Safety Management (SKYbrary); ICAO Safety Management — Standards and Recommended Practices; EASA: ICAO Annex 19 SMS and reporting SARPs transposed into EU law.