Agentic AI in Production: How We Built a Chatbot That Guides Tourists and Saves Lives
By Motti Dadison · April 24, 2026 · 5 min read
We implemented an agentic AI platform for a client that serves two critical purposes: helping tourists explore Israel and providing real-time escape plans during armed conflict. Here's how we built it.
What Is Agentic AI?
Agentic AI refers to artificial intelligence systems that don't just answer questions — they plan, reason, use tools, and execute multi-step tasks autonomously . Unlike a traditional chatbot that responds to a single prompt, an AI agent can break a goal into sub-tasks, call external APIs, search the web, query databases, and loop back on its own outputs until the job is done.
The key properties that make an AI system "agentic" are:
Autonomy — operates with minimal human intervention
Tool use — can call APIs, run code, read files, search the web
Memory — maintains context across long sessions or multiple sessions
Planning — decomposes complex goals into ordered sub-steps
Reflection — evaluates its own outputs and corrects mistakes
The Architecture of an AI Agent
At its core, an AI agent is built around a Large Language Model (LLM) that acts as the "brain," surrounded by a set of tools and a memory layer. Here is a high-level view of the architecture we use at Softech:
Our Client's Challenge: Serving Tourists and Keeping Citizens Safe
One of our clients — a government-adjacent tech organisation in Israel — came to us with a dual mandate:
Build an AI assistant that helps tourists explore Israel — suggesting itineraries, attractions, restaurants, and real-time logistics.
Build an emergency escape-plan assistant that guides civilians during armed conflict — providing shelter locations, evacuation routes, and real-time alerts.
Both use-cases live in the same agentic platform, with the system switching context based on the situation. The result is a single, battle-tested AI backbone serving two very different — yet equally critical — audiences.
Use Case 1: Explore Israel — Tourism Chatbot
The tourism assistant acts as a personal travel concierge. A tourist types a free-form request like "I have 3 days in Israel, I love history and food, where should I go?" and the agent:
Queries a vector database of 10,000+ curated attractions and restaurants
Pulls live weather and event data via external APIs
Calls the Google Maps API to validate travel times between stops
Generates a day-by-day itinerary with map links, booking references, and opening hours
Remembers the user's preferences across sessions for personalised follow-ups
Use Case 2: Emergency Escape Plan During Conflict
When a security alert is triggered, the same platform shifts into emergency mode . The agent's priority changes entirely — from discovery to safety. It ingests real-time data from national alert systems (like Israel's Red Alert / Tzeva Adom ) and provides:
Shelter finder — nearest protected spaces based on the user's GPS
Evacuation routes — road-by-road guidance avoiding active zones
Family coordination — lets users share live location with family members
Real-time updates — push notifications as the situation evolves
Multi-language support — Hebrew, Arabic, English, Russian, French
How We Built It: Technical Stack
The platform is built cloud-native on GCP and AWS, with a strong Kubernetes backbone managed by our DevOps team. Here's the core stack:
Key Challenges and How We Solved Them
1. Latency Under Stress
During emergency events, the system can see a 100x spike in concurrent users . We solved this with auto-scaling Kubernetes node pools, pre-warmed Cloud Run instances, and a lightweight emergency-mode fallback that bypasses the full LLM pipeline when needed — delivering shelter coordinates in under 500ms even at peak load.
2. Accuracy of Life-Critical Information
Hallucinations are unacceptable when lives are at stake. We implemented a retrieval-augmented generation (RAG) pipeline where every shelter address and evacuation route is grounded in a government-verified database, with the LLM only responsible for natural language formatting — not the underlying facts.
3. Dual-Mode Context Switching
The same user might be exploring Tel Aviv restaurants at noon and need emergency guidance at midnight. We built a context router that monitors incoming alert feeds and can instantly switch the agent's system prompt, tool set, and response tone — without interrupting the user's session.
Results
🗺️ 40,000+ tourist itineraries generated in the first 3 months
⚡ <2 second average response time for emergency guidance
🌍 Supports 5 languages with native fluency
🛡️ 99.97% uptime maintained during active conflict periods
⭐ 4.8/5 user satisfaction score across both use-cases
Conclusion
Agentic AI is no longer a research concept — it is production infrastructure. This project demonstrates that a single, well-architected agent platform can serve radically different use-cases: delighting tourists one moment and protecting lives the next. The key is composable tooling, grounded retrieval, and a resilient cloud backbone .
At Softech, we specialise in designing and deploying agentic AI systems that are not just impressive demos but reliable, scalable products. If you're exploring how AI agents could transform your business or public service, let's talk .
Topics: Agentic AI, LLM, ChatBot, Israel, Emergency Response, RAG