AI Emergency Assistant

This Azure-based AI assistant was developed through an intermediary for an anonymized international organization. At 5A Technologies, we build on the resulting solution architecture, RAG and Azure delivery expertise. The assistant guided employees to relevant company information during operational incidents.

Given a concrete question such as “There is flooding at facility 7C. What should I do?”, the solution could retrieve the applicable protocols, location-specific contacts and available follow-up actions.

• Context: operational incidents in an international organization
• Project: professional delivery through an intermediary
• Status: completed and technically handed over
• Role of 5A Technologies: carrying forward expertise in AI solution architecture, RAG implementation, Azure deployment and technical handover
• Core principles: enterprise grounding, location context and predefined actions

Key design principle
The assistant surfaced documented information and available actions. The model did not independently decide which undocumented action should be taken.

The organization had sites across several countries. During an incident, a local responsible person needed to know quickly:

• which protocol applied;
• which steps needed to be taken first;
• which contacts were responsible;
• which escalation route applied to the facility involved.

This information was distributed across an extensive internal knowledge source and was not always easy to retrieve quickly when time mattered.

Bound the incident and location

The user described the incident and the location involved in natural language. The solution then identified the incident type, facility and core question that needed to be answered.

Retrieve enterprise knowledge

A custom-deployed language model in Azure AI Foundry worked with a Retrieval-Augmented Generation pipeline. The model did not answer solely from general model knowledge: the RAG layer first retrieved the relevant company information and then assembled a location- and incident-specific answer from that material.

Bound the follow-up

The model structured the retrieved steps and responsible contacts into a clear action overview. Only retrieved company information and predefined follow-up actions remained available; where configured, a notification for selected contacts could be prepared.

1. The user described the incident and the location involved in natural language.
2. The custom model deployment processed the question as the entry point to the enterprise-grounded workflow.
3. The solution identified the incident type, facility and core question.
4. The RAG layer searched the internal knowledge source for relevant protocols and contact mappings.
5. The model structured the retrieved steps and responsible contacts into a clear action overview.
6. Only retrieved company information and predefined follow-up actions remained available.
7. Where configured, a notification for the selected contacts and an available channel was prepared.
8. The selected contacts could be notified by email, Microsoft Teams or SMS.

{
  "incidentType": "flooding",
  "facility": "7C",
  "protocols": ["relevant-protocol-reference"],
  "contacts": ["facility-response-contact"],
  "availableActions": ["notify_email", "notify_teams", "notify_sms"]
}

The output showed which information had been retrieved and which preconfigured actions were available. The model did not independently determine undocumented actions.

Confirmed core

• Azure AI Foundry: configuration and deployment of the custom-deployed model;
• Retrieval-Augmented Generation: grounding answers in the internal knowledge source;
• enterprise knowledge: surfacing protocols, facility data and contact mappings;
• controlled actions: notifying selected contacts through available communication channels.

Historical technical reconstruction

The exact historical service versions and connectors are no longer confirmed. The most plausible Azure architecture used Azure AI Search with embeddings for hybrid or vector retrieval and Azure Functions or Logic Apps for controlled actions. Notifications may have been connected through Microsoft Graph, Teams connectors or Azure Communication Services.

These components describe a technical reconstruction, not a separately confirmed product or version contract. No exact model version, database service or notification connector is therefore recorded as a public project fact.

The solution brought relevant information together, but it did not independently determine which undocumented operational action should be taken. Only retrieved company information and preconfigured follow-up actions remained available.

The assistant:

• grounded answers in retrieved company information;
• limited actions to preconfigured options;
• did not invent an undocumented procedure;
• did not activate an unconfirmed connector or escalation route;
• did not present an unconfirmed procedure as an available action.

At 5A Technologies, we build on experience across:

• designing the end-to-end architecture;
• configuring and deploying the model in Azure AI Foundry;
• setting up the RAG pipeline;
• structuring and indexing protocols, facilities and contact information;
• designing the question-and-answer logic;
• defining the possible follow-up actions;
• testing incident and location scenarios;
• technically handing the solution over to the intermediary.

The solution brought distributed emergency information together in one accessible, question-driven workflow. From one concrete situation, an employee could retrieve the relevant procedures, contacts and available actions without having to search multiple documents and contact lists manually.

The project was completed and handed over as an Azure solution to the intermediary, which could provide the package to the end organization’s IT department.

The public project description observes the following boundaries:

• it does not claim a global rollout or confirmed production use after the handover;
• it does not present an exact model version, database service or notification connector as a confirmed project fact;
• it does not claim fine-tuning;
• it states no proven savings, response-time improvement or adoption figure;
• it makes no confirmed compliance claim.

This project shows how a bounded AI solution:

• connects incident questions to controlled enterprise knowledge;
• links location context to applicable protocols and responsible contacts;
• gives retrieval and answer structuring distinct roles;
• limits follow-up actions to preconfigured options;
• combines technical handover with explicit boundaries around unconfirmed details and outcomes.