# Who decides it’s an emergency?

## AI, LLMs and Romania’s 112 system

**Checked 19 August 2026.** Romania is buying a 112 platform that must include at least one LLM component or library. The public documents do not yet show where auxiliary automation ends and influence over classification, priority or dispatch begins.

This page does not claim that the final system will legally be high-risk. The effective functions and RADCOM’s technical proposal are not public at enough granularity to support that conclusion.

## The right question

Similar technologies can transcribe “I have chest pain”, translate a call or extract explicit facts. Another configuration can infer that the call is critical, move it ahead of others and recommend which crew should go.

The technical label does not settle the issue. The boundary is the concrete role of each output: does it only carry information, or does it change the chance that a person receives attention sooner and with particular resources?

The AI Act lists as high-risk systems intended to evaluate and classify emergency calls or to be used for dispatching or prioritising first-response services. Narrow preparatory tasks can be excepted when the system creates no significant risk and does not materially influence the decision outcome. A provider relying on the exception must document its assessment.

## What is publicly documented

- The contract was awarded to RADCOM for RON 24,548,231.
- The platform includes RTT, video, SMS and Chatbot/Smart IVR.
- At least one LLM component or library was mandatory in Chatbot/Smart IVR.
- At least three libraries in total could earn 12 technical points out of 100.
- The official programme includes AI/ML assistance, analysis and decision support.
- The wider 112 programme has announced total funding of RON 225.52 million and an end date of 6 March 2028.

## The working model

A call can, conceptually, pass through these stages:

1. Transcription: voice to text.
2. Translation: the same words in another language.
3. Fact extraction: explicit facts for the case file.
4. Summary: neutral condensation or a severity assessment.
5. Urgency estimate: how quickly should the response arrive?
6. Resources: what type of help is recommended?
7. Priority: where does the call enter the queue?
8. Dispatch: what intervention is suggested or triggered?

The early stages may be auxiliary, depending on purpose and influence. An output that evaluates severity, changes priority or influences resources moves the analysis toward the uses listed by the AI Act.

## What should be measured

Any claim about an AI component should be testable through omissions, false classifications, latency, degradation across accents, dialects, stress or disability, availability and fallback. Final confirmation by an operator does not automatically erase the influence a system had on evaluating the call.

## Timeline

| Date | Event |
| --- | --- |
| 6 March 2025 | The wider 112 programme begins with STS and the Interior Ministry. |
| 28 July 2026 | The RADCOM contract is signed. |
| 2 August 2026 | The general chatbot transparency rule applies, subject to Article 50 exceptions. |
| 18 August 2026 | Award notice CAN1172992 is published. |
| 2 December 2027 | Duties for Annex III high-risk systems apply under the updated timetable. |
| 6 March 2028 | The announced end date for the 112 programme. |

## What is not public yet

- How many LLM libraries RADCOM offered and will deliver.
- Whether the chatbot speaks directly to callers or only assists operators.
- Whether it produces scores, categories or severity alerts.
- Whether it recommends resources or changes call priority.
- What an operator can change, ignore or stop.
- What classification assessment was made for each component.
- What acceptance thresholds, logs and fallback procedure apply.

Confidence is very high on the award, the LLM criterion, the programme’s official ambition and the legal text. The final system’s classification remains deliberately undetermined: it depends on intended purpose, the effective workflow and each component’s influence.

## The documents with the highest public value

The public does not need source code or details that expose critical infrastructure. It needs the boundary of authority, acceptance criteria and the allocation of responsibility:

- RADCOM’s technical proposal and the role of each LLM component;
- functional specification 4.2.2.3 for Chatbot/Smart IVR;
- the AI Act classification assessment and reasoning for any exception;
- the testing and acceptance matrix;
- the human oversight design;
- logging, traceability and fallback;
- the allocation of provider and deployer roles.

## Freedom-of-information request

The request in the HTML edition can be adapted and sent to the Romanian Special Telecommunications Service at `office@sts.ro`. It seeks functional descriptions and governance documents, not passwords, vulnerabilities, network topology or live operational data. Fill in your name, date and contact details before sending.

## Sources

1. [Consolidated procurement record for CAN1172992](https://sicap.ai/licitatii/contract/100649221)
2. [Government Decision 485/2026 and its case for the investment](https://legislatie.just.ro/Public/DetaliiDocument/311577)
3. [Interior Ministry: Extending and improving the 112 emergency service](https://www.mai.gov.ro/extinderea-si-eficientizarea-serviciului-de-urgenta-112-oferit-cetatenilor/)
4. [Regulation (EU) 2024/1689, the AI Act](https://eur-lex.europa.eu/eli/reg/2024/1689/oj)
5. [Regulation (EU) 2026/1744](https://eur-lex.europa.eu/eli/reg/2026/1744/oj)
6. [European Commission: AI Act Service Desk, essential services](https://ai-act-service-desk.ec.europa.eu/en/essential-services)

The Service Desk examples aid interpretation but do not decide this contract. The page should be updated if the technical proposal, full contract, testing matrix or classification assessment becomes public.

_Research and interface: [Marius Comper](https://mariuscomper.uk/en/)._