Emergency Dispatch AI Is High-Risk Under the EU AI Act. The Deadline Just Moved.

Emergency dispatch AI is not a general-purpose productivity tool in the eyes of EU law. AI that evaluates emergency calls or decides who gets sent where is a high-risk system. Those obligations were due this month, and have just been pushed to December 2027 — but not everything moved with them.

Annex III of the EU AI Act lists the uses that attract the high-risk classification. Point 5(d) covers ‘AI systems intended to evaluate and classify emergency calls by natural persons or to be used to dispatch, or to establish priority in the dispatching of, emergency first response services, including by police, firefighters and medical aid, as well as of emergency healthcare patient triage systems.’

That is a precise description of a category of tool being actively sold into emergency management today. Any agency evaluating call-triage, dispatch-prioritisation, or resource-allocation software is evaluating a high-risk system, whether or not the vendor presents it that way.

The deadline moved. The obligations did not.

The postponement

Under the Digital Omnibus agreement, obligations for standalone Annex III high-risk systems move from 2 August 2026 to 2 December 2027. Systems embedded in products already covered by sectoral legislation — medical devices, machinery, vehicles — move from 2 August 2027 to 2 August 2028. The agreement was reached provisionally on 6 May 2026 and confirmed by member state representatives on 13 May 2026, with formal adoption and publication in the Official Journal following. Organisations planning against these dates should confirm the final published text rather than the political agreement.

What did not move

The postponement is narrower than the headlines suggest. Set the dates against each other and only the high-risk obligations moved:

So an agency that read ‘delayed to 2027’ and closed the file has misread it. The high-risk engineering and documentation burden moved. The prohibitions and transparency duties did not.

Sixteen months is a build window, not a reprieve.

That additional time is not an invitation to defer the question. It is the window in which governance can be built before the tooling is embedded, which is the sequence the sector’s own research says agencies routinely get backwards.

RAND’s 2026 market review, produced for the Markle Foundation’s AI for Disasters and Emergencies Initiative with Aspen Digital, catalogued 1,179 AI-enabled products with potential relevance to emergency management. Around half are purpose-built for the sector or have documented use in it. Roughly 40 percent are general-purpose tools with no emergency-management-specific design — which matters considerably when a system drifts into an Annex III use case and the obligations attach to the deployer as well as the provider.

Jeremy Greenberg, Senior Adviser at Aspen Digital and previously director of FEMA’s Response Operations Division, frames the sequencing: ‘With better governance comes better adoption.’ Not the reverse. His characterisation of the technology’s role — ‘AI is designed as an augmentation tool, not a replacement tool’ — happens also to describe what the high-risk regime requires. Human oversight is not an ethical nicety under the AI Act. It is a design obligation.

What this means in practice

• Establish whether any tool in use or under evaluation touches Annex III 5(d). Call triage, dispatch prioritisation, and resource allocation are the obvious candidates. The classification follows the use, not the vendor’s marketing category.
• Document where human judgement remains final, particularly for life-safety decisions, and make that boundary a procurement requirement rather than an afterthought.
• Check the transparency obligations now, since those are live and were not deferred.
• Treat workforce training as a condition of deployment. A system staff cannot operate accountably is not adding capability, and will not satisfy a human-oversight requirement.

This is the boundary RAYVN is built around, with or without AI in the picture. The platform assembles a shared, role-appropriate operational picture faster than any individual could put one together manually, and the decision stays with an accountable person who owns it — recorded, timestamped, and answerable afterwards. Speed in the information layer, human ownership in the decision layer.

This article summarises the EU AI Act and the Digital Omnibus agreement for general information and reflects our reading of the regulations as they stand. It is not legal advice and should not be relied on as such. For a definitive view on how these obligations apply to your organisation, consult a qualified legal adviser.