A frontline agent shouldn't be scrolling through PDFs in front of a customer.
The airport-services arm of a global carrier needed its agents to answer faster than the customer could ask - we put the SOP library inside an agentic assistant.
The function runs everything airside: check-in, baggage, lounges, disruption handling, ramp services, ground dispatch, stand planning, aircraft cleaning. The operational manuals behind it evolve constantly - regulatory updates, network changes, new aircraft types, evolving customer-service policy. Frontline agents were caught in the middle. Customer asks a scenario question. Agent opens a PDF. Agent scrolls. Customer waits. Service drops. Multiply that by every touchpoint in a global hub operation and you have a structural drag on quality, speed, and compliance. The question: could a single assistant put the right SOP answer in the agent's hands inside the three seconds a customer will tolerate?
- 01
Build for scenario, not for keyword.
Airport agents don't ask 'what is the policy for X.' They ask 'this passenger has Y, with Z complication, and they're going to A - what now?' We built an agentic assistant on top of an LLM core that reasons across the scenario, looks up the relevant SOPs, and reconciles them into a single, citable answer.
- 02
Constrain to policy, refuse outside it.
Safety and compliance in aviation operations are not negotiable. We restricted the assistant to approved policy content and engineered explicit refusal for anything beyond it. The judgment call: where most teams optimize for breadth, we optimized for governed coverage. An assistant that won't go off-policy is worth more than one that might.
- 03
Ship for scale, day one.
Multilingual responses, sub-three-second latency, and a deployment pattern that worked across every frontline touchpoint. The pilot was the production: this became the first group-wide GenAI deployment in the carrier's operations.
Agents stopped scrolling and started answering. Service quality, consistency, and speed improved across every airport touchpoint where the assistant was deployed. Compliance went up because the answer source is now governed - when a policy changes, every agent gets the new answer the same day. Management gained a scalable channel for distributing operational updates across a global workforce. The unlock: a proven blueprint for further AI rollouts across the group's operations.
“In safety-critical operations, an AI assistant earns trust by what it refuses, not by what it produces. Build the constraints first; the capability follows.
Frontline teams stuck between customer-facing speed and policy-grade accuracy? We help operations leaders ship governed GenAI to the people who actually use it.
Let's talk
