AI in 112 Emergency Calls 2027 – High-Risk Assessment, Human Oversight, and Testing Protocols | Word/PDF/Excel

Are you considering AI to analyze emergency calls, support dispatchers, or prioritize ambulance and rescue services? This niche document package helps emergency call centers, regional authorities, healthcare and rescue organizations, and suppliers document whether the AI system is subject to the EU's high-risk rules, how human oversight works, and what critical tests must be carried out before any real-world deployment is considered.

The special value: Annex III point 5 d of the AI Act covers AI systems intended to evaluate and classify emergency calls or direct/prioritize emergency services and emergency medical care, as well as certain acute patient triage systems. The special Annex III high-risk rules will apply from December 2, 2027. This package is a manual preparedness resource, not a medical decision-making system.

Who can benefit from the package?

Regional authorities, ambulance and emergency call center operations, suppliers of AI-based prioritization support, integrators, procurers, medical management, IT security, data protection officers, and those responsible for the EU AI Act. The documents provide a framework for source, risk, and decision review and should be adapted to the organization's actual legal basis, medical procedures, and expertise. Purely administrative AI may require a different classification than AI that actually affects prioritization.

File by file – concrete deliverables

Excel – nine tabs and eight control signals

Overview, System, Classification, Data Protection, Supplier, Safety Testing, Human Control, Pilot Decision, and Read Me. Excel records the number of systems and specific open questions such as unclear classifications, FRIA/DPIA investigations, missing supplier documentation, failed safety tests, failed human control, and blocked decisions. Status selections, source references, and formula calculations have been tested with synthetic data.

Important: A zero in the overview only means that no relevant deviations have been recorded. It does not mean that a system is safe, legal, clinically approved, or ready for deployment. Excel can never prioritize a patient or emergency situation.

Legal boundaries 2026/2027

AI Act – Annex III 5 d and Article 6. AI that evaluates or classifies emergency calls, determines or prioritizes dispatching police, emergency services, and medical assistance, or is part of certain acute patient triage systems, is considered high-risk when the criteria are met. The classification depends on actual intended use and the provisions of Article 6 – not solely on the supplier's marketing or the fact that the system uses AI.

Date. According to the consolidated AI Act after the 2026 Digital Omnibus, the Annex III high-risk rules will apply from December 2, 2027. Other AI Act rules, GDPR, and Swedish operational rules may apply earlier.

Article 26 – operation and human control. Deployers of high-risk AI shall, when the rules apply, among other things follow relevant instructions for use, assign competent human oversight, monitor use, and manage logging and serious risks. Read Article 26. It is not enough to write a person's name on a checklist if the person lacks the actual mandate or practical opportunity to intervene.

Article 27 – FRIA. A fundamental rights assessment before first use may be required for relevant high-risk AI systems deployed by public bodies or private actors providing public services. The scope must be assessed separately from GDPR. Read Article 27.

GDPR and patient data. Processing of sensitive health data requires a separate legal basis under GDPR Article 6 and an applicable exemption under Article 9, and IMY's guidance may trigger a DPIA under Article 35. The Patient Data Act (2008:355) applies within its scope to a healthcare provider's processing of patient data – not automatically to every 112 call.

Ambulance calls – SOSFS 2009:10. The Swedish National Board of Health and Welfare's consolidated regulations contain provisions on agreements with emergency call centers and priority levels for ambulance calls: Priority 1 life-threatening, Priority 2 acute but not life-threatening, Priority 3 other medical/monitoring needs where a reasonable wait is not assessed to worsen the patient's condition. The product templates are not medical guidelines, prioritization support, or instructions for a real emergency call.

How you use the package

1. Define the exact AI system and whether it can affect call classification, priority, or dispatching.
2. Investigate the EU AI Act classification, the organization's role, and applicable Swedish operational regulations.
3. Assess data protection, potential FRIA, DPIA, and access to confidential/health data.
4. Request the supplier's evidence and test dangerous, synthetic error scenarios – especially incorrectly low priorities.
5. Verify the human's actual ability to stop the AI and perform safe manual prioritization.
6. Make a documented STOP/INVESTIGATE/isolated test decision. Real-world deployment requires special authorized and professional controls that the package does not perform.
7. Record only de-identified statuses and evidence IDs in Excel and follow up on new model versions and deviations.

Frequently Asked Questions

Is AI that sorts 112 calls always high-risk?

AI systems intended for emergency call classification or dispatch prioritization are typically covered by Annex III 5 d. Exact classification requires assessment under Article 6 and intended use. Administrative support functions may need to be assessed differently.

Can the template be used to determine ambulance priority?

No. The templates are only for preparatory risk, testing, and compliance work. They do not provide medical advice and are not suitable as decision support for real patients or emergency calls.

Is FRIA the same as GDPR's DPIA?

No. FRIA under AI Act Article 27 and DPIA under GDPR Article 35 have different legal bases and requirements; both may be relevant.

Does it include real technical or medical validation?

No. The package provides forms and control questions. Real test specifications, operational requirements, and professional approval must be carried out in a separate process.

What happens if an AI suggestion could lead to dangerous undertriage?

The case must be reviewed by a competent safety/medical function. The package contains stop signals and documentation fields but no automatic solution for incorrect AI results.

Can Excel store real 112 calls and patient data?

No. Only de-identified case, test, and evidence IDs should be used. Personal and health data must be handled in authorized operational systems.

Are the PDF files digitally fillable?

No. The PDFs are searchable and printable; Word and Excel are the editable originals.

Are the work forms available in English?

The five work forms and Excel are in Swedish. A separate English User Guide describes the workflow for English-speaking users.

Version and legal transparency

Version 1.0 · November 12, 2026 · Sweden/EU. AI-assisted review against official legal sources has been performed. No external legal review, clinical validation, medical risk acceptance, actual FRIA/DPIA, patient safety certification, or testing of real emergency calls has been performed. The documents should be reviewed by competent legal, medical, and technical functions before use. The product remains a DRAFT, and the ZIP file is not linked to Digital Products until a separate publication decision is made.