Case study · Travel tech · Air cargo
How a global travel-tech provider replaced multi-screen cargo planning with one AI co-pilot, built with ShellKode FDEs.
Global travel and logistics SaaS provider · Air cargo
Too many options, too little time before departure. Now the planner just asks.
More from every flight
Higher capacity use and revenue per flight
Less time planning
Shorter planning effort and cycle time
Decisions on benchmarks
Consistent, benchmark-driven recommendations
People stay in charge
Every commercial action approved by a person and traceable
The challenge
The best load for each flight was buried in options, and found by hand.
The provider's software supports cargo operations for airlines. Close to departure, planners had to deal with unused reserved capacity, lower-value bookings and cargo that had to fly, checking several screens and relying on experience.
An early version of the capability existed, but the hardest part was still open. Hitting revenue and priority targets at the same time, within capacity, is beyond manual analysis. Nobody had decided what the AI should reason about and what should simply be calculated, and there was no way to test whether a recommendation was right. Closing that gap took Forward Deployed Engineers working inside the provider's teams.
Unused capacity found too late
Reserved space was spotted too close to departure to resell.
Manual, multi-screen planning
Every move depended on experience and several screens.
Trade-offs beyond human reach
The number of possible combinations made manual optimisation impractical.
Promises that can't be broken
Service commitments, priority cargo and operating limits all had to hold.
What we delivered
Agents that reason, tools that calculate, and planners who decide.
ShellKode FDEs embedded with the provider's architecture, engineering and QA teams so that every recommendation would be both right and trustworthy.
A proven approach to the hardest problem
The revenue and priority trade-off was broken into a set of strategies, and the agent picks one per flight.
A clear line between agent and tool
Architecture reviews and decision records set what Amazon Bedrock AgentCore agents reason about and what tools calculate.
A home for every business rule
A rule-placement framework decides whether each rule lives in code, configuration, knowledge or the prompt.
Quality you can test, again and again
Golden test scenarios, regression checks and a guardrail review keep every change honest.
The outcome
Planners start from ranked, explained options, not a wall of screens.
More value from every flight
Unused capacity surfaces early, and loads shift to higher-value cargo.
Faster planning
One conversation replaces several screens.
Consistent, benchmark-driven decisions
Recommendations are measured against revenue and priority benchmarks.
Full human control and auditability
Planners approve every action, and each one is re-checked and traced.
The stack
What it runs on.
- Agents
- Amazon Bedrock AgentCore
- Quality assurance
- Golden test scenarios, regression checks, guardrail review
What's next
The foundations are set. Next comes go-live.
A production-readiness assessment has set out the path to go-live.
- 01
Acting on the go-live recommendations
- 02
Running regression checks against the golden scenarios with every release
- 03
Extending the approach to more cargo workflows
Talk to us
Bring one decision your team makes by hand. Leave knowing what an agent should reason about.
30 minutes on one workflow, mapped to the agents, tools and guardrails behind it. No slides.