Product Case · Agentic AI Design
A major airline's contact center saw inbound volume spike 400–600% within 90 minutes of every disruption — with almost no way to absorb it but more phone agents.
Self-directed case study exercise, not a real client engagement — a Product Analyst design exploration of agentic AI architecture for airline IROPS (irregular operations) recovery, using a fictional airline (SkyWave) and illustrative figures throughout.
During weather delays, mechanical issues, or crew shortages, an all-human contact model couldn't keep up with the surge — hold times stretched to 47 minutes and customers defected to self-help. I designed an end-to-end agentic AI system: discovery, decision logic, systems integration, and conversation behavior, all the way through to a test plan.
Executive presentation (PDF)
Agentic workflow blueprint
Who
Airline CX operations and contact-center leadership
Manage customer contact volume during flight disruptions.
What
Absorb the surge without just adding headcount
Resolve routine cases autonomously, and know exactly when to hand off to a human.
Why
70% target containment
The design objective for the system — not a measured result from a live deployment.
Two things that were broken
Hold times spiked to 47 minutes
A 400–600% surge in contacts within 60–90 minutes overwhelmed an all-human model, pushing customers to defect to self-help channels.
Compensation costs nobody was managing
Unmanaged customer expectations during disruption drove significant, avoidable compensation payouts.
What we designed
- Mapped the problem before the solutionStakeholder map, current-state pain points, and KPIs (containment rate, CSAT, handle time, rebooking speed).
- Specified the agent's decision logicA workflow blueprint of intents, actions, and escalation paths — with guardrails and human-in-the-loop triggers.
- Designed how it connects to real systemsIntegration architecture across flight inventory, PNR, CRM, and ticketing, triggered by live flight-status events.
- Specified how the agent behavesAn intent library, sample conversation flows, and tone/persona guidelines — not a generic chatbot script.
- Built a test plan before building the agentA scenario matrix (cancellation, delay, missed connection, weather event) with success thresholds for each.
The agent isn't designed to handle everything — it's designed to know exactly when to hand off to a human.
What it was designed to move
400–600% surge
Contact volume within 60–90 minutes
the spike the system had to be designed to absorb
5 deliverables
Discovery through test plan
workflow blueprint, integration architecture, agent behavior spec, and more
70% / 4.2 / 55%
Containment, CSAT, and cost-per-contact targets
design objectives, not measured outcomes