Aviation Copilot
AI customer service for airlines, with rules the model cannot break.
Context
In aviation, customer service is regulated. Delays, cancellations and rebooking carry rights defined by rule. An AI that improvises becomes a legal liability.
Challenge
- Answers must follow regulation and airline policy, no exceptions.
- Demand spikes during operational disruptions.
- Passenger personal data in every conversation.
Solution
A chat and voice agent that talks to passengers but doesn’t decide alone: decisions about rights go through a rules engine the model cannot override. The knowledge base comes from public civil aviation regulations.
The model suggests. The rule decides.
Pick a scenario and watch the Copilot handle it: the model understands the request and proposes actions, but passenger rights go through a rules engine it cannot bypass.
- Ready
- Reasoning
- Planning
- Retrieving
- Running
- Done
- Waiting for the passenger.
Simplified rules from Brazil’s ANAC Resolution 400/2016, fictional airline and passengers. Not legal advice. The model proposal is simulated; the decision engine is the same as in the product.
Architecture
- 01Central orchestrator with scenario-specialist agents
- 02Decision engine with policy rules outside the LLM
- 03RAG over public civil aviation regulations
- 04Server-side voice: transcription, security, RAG and speech synthesis
- 05Real-time dashboard with KPIs, cost and satisfaction
Governance & security
- Five security layers: PII masking, anti-jailbreak, output guardrails and role-based access
- Tone personalization with an anti-bias guardrail
- Synthetic scenarios to test before operating
Current state
- In evaluation, on synthetic data
- Operating figures will be published once real data exists