Who decides it’s an emergency?

The right question

In 112, the question is what authority AI receives.

A similar family of technologies can perform very different public functions. One model can transcribe “my chest hurts”. Another 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 merely carry information, or does it change a person’s chance of receiving attention, sooner, and with particular resources?

Where does operator assistance end, and where does AI influence over emergency classification or response priority begin?

The AI Act expressly lists systems intended to evaluate and classify emergency calls, dispatch first response services, or establish dispatch priority as high-risk. It provides exceptions for some narrow or preparatory tasks, but only when the system creates no significant risk and does not materially influence the decision outcome. A provider relying on the exception must document that assessment.

What the procurement proves

A real contract. A real LLM requirement. An operational role that remains only partly visible.

Award notice CAN1172992 became public on 18 August 2026. The contract was signed on 28 July, runs for 20 months, and covers RTT, video, SMS and Chatbot/Smart IVR platforms to be integrated with Romania’s existing 112 system.

24.55

million lei

The value awarded to RADCOM, against an estimated procedure value of RON 38.46 million.

1 → 3

LLM libraries

One was mandatory. One additional library earned 6 points; at least two additional libraries earned 12.

225.52

million lei

The value of the wider 112 programme, run by STS and the Interior Ministry until 6 March 2028.

The LLM criterion was justified by “reliability and general performance”, including modules’ ability to account for context, spoken dialect and terminology. The public award record does not show how many libraries RADCOM proposed, which models they are, which functions they receive, or RADCOM’s score on that criterion.

Government Decision 485/2026 describes an ambition wider than a basic chatbot: AI and ML for conversation, analysis, correlation and prediction to identify the nature of the emergency, its complexity and immediate risks, and to support decisions and resource allocation. This does not prove that every function sits inside the RADCOM contract. It does show why the boundary must be disclosed component by component.

Interactive experiment

Move the boundary of authority.

Follow a fictional emergency call through eight stages. Choose what AI does at each point. The result is not an automated legal opinion; it maps where a function moves from carrying information into evaluation, priority or dispatch.

auxiliary function boundary case / document it likely high-risk use

The categories draw on the AI Act and the Commission AI Act Service Desk’s current examples. Those examples are guidance, not a binding determination of the future STS system.

An important correction

A human in the loop is not a legal eraser.

Current Commission guidance draws the boundary through function. An operator’s ability to confirm or reject an output matters for oversight, but it does not automatically make the system low-risk.

Outside 5(d), in the official example

It transcribes, but does not evaluate.

A system that cleans up or transcribes a difficult call and marks relevant information, without determining urgency or priority, is not treated as an emergency-call evaluation system.

Source: AI Act Service Desk, “Essential services”.

High-risk, in the official example

It listens for danger and helps classify.

A real-time assistant that looks for signs of a life-threatening emergency and helps the call-taker classify severity is high-risk, even when a human operator makes the final decision.

Source: AI Act Service Desk, “Essential services”.

The consequence for 112: “the operator decides” is necessary information, but it is not enough. We need to know what the model produced before the decision, how that output was displayed, and how strongly it structured the human’s options.

What the tender rewarded

Architecture received points. Operational safety needs different measures.

The number of LLM libraries was a material technical criterion. But three models are not automatically safer than one, and the public notice does not show the acceptance matrix for errors that matter in an emergency service.

The 100 points

award criteria

40141212125

Price40
24/7 support14
TC conference access12
At least 3 LLM libraries12
Text-to-speech12
Tutorials and lessons5
PM experience2
Methodology3

Weights and descriptions come from the consolidated procurement record.

1 + 2

One mandatory; two more for maximum points.

STS linked plurality to reliability and performance, including context, dialect and terminology. The public record does not show whether RADCOM offered one, two or at least three libraries, nor how their outputs are combined.

↓Missed signalsHow many critical signs are not detected or never reach the operator?

↕MisclassificationHow many cases are wrongly escalated or down-triaged?

msLatencyHow much time does each component add, including at peak demand?

≈Differential degradationDialect, accent, stress, noise, disability and different types of speech.

24/7AvailabilityWhat level is guaranteed and how does degraded mode work?

↩FallbackWho takes over, what disappears, and what is logged when outputs conflict?

The implementation meets the law

The high-risk regime becomes applicable during the project.

The 2026 amendment to the AI Act moved the application of Annex III high-risk requirements to 2 December 2027. Romania’s 112 programme runs into March 2028.

6 Mar 2025
The wider 112 programme startsSTS and the Interior Ministry; total funding of RON 225.52 million.
28 Jul 2026
The RADCOM contract is signedRTT, video, SMS and Chatbot/Smart IVR platform.
2 Aug 2026
The general chatbot disclosure rule appliesWhen a person interacts directly with AI, they must be informed, subject to the exceptions in Article 50.
18 Aug 2026
Award notice CAN1172992 is publishedValue: RON 24,548,231; winner: RADCOM.
2 Dec 2027
Annex III high-risk duties applyRisk management, data, logging, documentation, human oversight, robustness, security and accuracy.
6 Mar 2028
Announced end date of the 112 programmeThe regulatory boundary arrives before implementation is due to finish.

Confidence register

What we know. What we do not. What should not be guessed.

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/library was mandatory.
At least three libraries in total could earn 12 points.
The official programme includes AI/ML assistance, analysis and decision support.
The AI Act lists emergency-call evaluation and response prioritisation as high-risk uses.

Not yet publicly documented

How many LLM libraries did RADCOM offer and will it deliver?
Does the chatbot speak directly to callers or only assist operators?
Does it produce scores, categories or severity alerts?
Does it recommend resources or alter call priority?
What can an operator change, ignore or stop?
What Article 6 assessment was made for each component?
What are the acceptance thresholds, logs and fallback procedure?

Confidence: very high on the award, the LLM criterion, the programme’s official ambition and the legal text. Deliberately undetermined on the final system’s classification: that depends on the intended purpose, effective workflow and influence of each component.

The missing governance file

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 proposalThe LLM components, each role, logical architecture and accepted technical score.

Functional specification 4.2.2.3Inputs and outputs of Chatbot/Smart IVR, and who receives each output.

The AI Act classification assessmentIntended purpose and the reasoning for high-risk status or an Article 6(3) exception.

Testing and acceptance matrixErrors, latency, availability, test populations and rejection thresholds.

Human oversight designWhat operators can change, which warnings they see, how automation bias is managed, and who can stop a component.

Logging and traceabilityWhich output, version, input and human intervention can be reconstructed after an incident.

Fallback and degraded modeBehaviour during unavailability, conflicts between models, or excessive latency.

Provider and deployerWho bears each legal duty: producer, integrator, STS or another actor.

Public instrument

A freedom-of-information request aimed at the real boundary.

The request seeks functional descriptions and governance documents, not vulnerabilities, passwords, network topology or operational details that could endanger the service. It is in Romanian, ready to adapt and send to STS.

Către: Serviciul de Telecomunicații Speciale
În atenția compartimentului de informare și relații publice

Subiect: Solicitare de informații de interes public: CAN1172992 / componentele AI și LLM ale platformei Chatbot/Smart IVR pentru SNUAU 112

Stimată doamnă / Stimate domn,

În baza Legii nr. 544/2001 privind liberul acces la informațiile de interes public, vă rog să-mi comunicați, în format electronic, următoarele informații și documente referitoare la contractul atribuit prin anunțul CAN1172992 pentru „Platformă software pentru comunicații de urgență prin servicii multimedia de tip RTT, video și SMS”, inclusiv Platforma Chatbot/Smart IVR:

1. Contractul semnat la 28.07.2026 și anexele sale, inclusiv propunerea tehnică a ofertantului câștigător, în măsura în care acestea pot fi comunicate; pentru pasajele exceptate, vă rog să transmiteți documentele în formă parțială, cu anonimizarea informațiilor protejate și indicarea punctuală a temeiului juridic al fiecărei omisiuni.

2. Secțiunea 4.2.2.3, „Cerințe funcționale specifice asociate Platformei Chatbot/Smart IVR”, și orice document ulterior care detaliază intrările, ieșirile și rolul funcțional al componentelor/librăriilor LLM.

3. Numărul de componente/librării LLM ofertate și acceptate în propunerea RADCOM, punctajul acordat pentru factorul 2.3 și partea relevantă a raportului procedurii / fișei de evaluare care justifică punctajul.

4. Pentru fiecare componentă AI/ML sau LLM prevăzută: scopul intenționat, utilizatorul outputului, dacă interacționează direct cu apelantul și dacă produce ori influențează scoruri, clase de urgență, alerte de severitate, prioritatea apelurilor, recomandări de resurse sau dispecerizarea acestora.

5. Orice evaluare, notă, registru de risc sau opinie juridică prin care STS ori contractantul a stabilit aplicabilitatea art. 6 și a Anexei III pct. 5(d) din Regulamentul (UE) 2024/1689, inclusiv justificarea unei eventuale încadrări în excepțiile art. 6 alin. (3) și documentarea prevăzută de art. 6 alin. (4).

6. Matricea de testare și acceptanță pentru funcțiile AI/ML și LLM: tipurile de eroare măsurate, pragurile minime, latența, disponibilitatea, comportamentul în mod degradat și testarea pe limba română regională, dialecte, terminologie, vorbire afectată de stres, zgomot și nevoile persoanelor cu dizabilități. Nu solicit seturi de date care conțin date personale ori detalii care ar permite compromiterea sistemului.

7. Descrierea mecanismelor de supraveghere umană: ce output poate modifica sau respinge operatorul, ce acțiuni nu pot fi executate de AI, cum este prevenită acceptarea automată a recomandărilor și cine poate suspenda o componentă.

8. Categoriile de loguri și informații de trasabilitate păstrate pentru outputurile AI/LLM, versiunea modelului, intervenția operatorului și decizia finală, fără a include date cu caracter personal ori detalii tehnice exploatabile.

9. Procedura de fallback pentru indisponibilitate, latență excesivă, output contradictoriu între librării sau rezultate sub pragul de încredere.

10. Alocarea rolurilor de „provider” și „deployer”, în sensul Regulamentului (UE) 2024/1689, între STS, RADCOM și eventualii furnizori ai componentelor AI/LLM.

11. În cazul unei interacțiuni directe între Chatbot/Smart IVR și apelant, modalitatea prevăzută pentru informarea acestuia că interacționează cu un sistem AI, conform art. 50, inclusiv soluția pentru comunicări de urgență în care timpul și accesibilitatea sunt critice.

Precizez că nu solicit cod sursă, credențiale, adrese, topologii de rețea, configurații de securitate, vulnerabilități, date operaționale în timp real sau alte informații a căror divulgare ar putea afecta securitatea SNUAU. Solicit descrierile funcționale, criteriile de guvernanță și documentele de evaluare care permit înțelegerea frontierei dintre asistarea personalului și influențarea clasificării ori dispecerizării unei urgențe.

Pentru orice informație exceptată de la comunicare, vă rog să aplicați principiul accesului parțial și să indicați pentru fiecare pasaj omis categoria de informație protejată, temeiul legal și motivul pentru care anonimizarea nu ar permite comunicarea restului documentului.

Vă rog să confirmați înregistrarea cererii și să transmiteți răspunsul la adresa de e-mail: [ADRESA DE E-MAIL].

Cu stimă,
[NUME]
[DATA]

Sources and method

The chain of evidence.

Contract facts, programme ambition, legal rule and interpretive examples are kept separate. No fictional interface element is presented as a feature of the contracted system.

01Consolidated procurement record: CAN1172992Value, winner, duration, platforms and award criteria, including the LLM factor. 02Romanian Government Decision 485/2026The official case for the 112 investment, including AI/ML and decision support. 03Interior Ministry: “Extending and improving 112”Period, objectives, components and funding value. 04Regulation (EU) 2024/1689: the AI ActArticle 6, Annex III point 5(d), high-risk duties, transparency, provider and deployer roles. 05Regulation (EU) 2026/1744The amended calendar: 2 December 2027 for Annex III high-risk systems. 06European Commission: AI Act Service DeskCurrent examples for transcription, operator support, triage, priority and dispatch.

Methodological limit: Service Desk examples aid interpretation but do not decide this contract. Legal classification requires the intended purpose and the real workflow. This page should be updated if the technical proposal, full contract, testing matrix or Article 6 assessment becomes public.