Amazon Web Services has outlined a new workflow pattern that allows AI agents to recommend airline rebooking options while keeping reservation changes and payments under the control of deterministic software.
The approach combines Amazon Bedrock AgentCore with AWS Step Functions. In an airline disruption scenario, specialised AI agents can evaluate passenger preferences, seat availability, routes and other information to propose alternative flights. However, the agents are not permitted to directly change a reservation or issue compensation. Those actions take place only after conventional code validates the proposal.
AWS describes the principle as one where agents make proposals and deterministic code validates them. The architecture is intended to separate the probabilistic reasoning of generative AI from actions that can carry financial, operational or regulatory consequences.
In the airline example, a flight cancellation can trigger the workflow through Amazon EventBridge. Passenger information, including existing bookings, loyalty status and preferences, is then collected before AI agents begin processing affected travellers.
One agent identifies three potential rebooking options, while another prepares personalised communication related to compensation. AWS Lambda-based validation steps then check whether the proposed flights actually exist, have available seats, comply with fare rules and meet route requirements. Compensation eligibility is also calculated through deterministic rules rather than being left to an AI model.
Cases that pass validation can proceed through the workflow, while situations requiring further judgement can be routed to human employees. Step Functions also maintains an execution history, creating a record of proposals, validations and approvals that can be reviewed later.
The architecture is designed to address a wider challenge emerging as businesses give AI agents more responsibility. Generative AI models can reason across complex information, but their outputs remain non-deterministic and can contain incorrect or outdated information. AWS's design therefore places a verifiable software layer between an agent's recommendation and the final action.
The company said the same pattern can extend beyond airline operations to workflows where automated decisions involve financial or regulatory consequences. AWS also supports deterministic controls elsewhere in Bedrock AgentCore, including code-based evaluators that allow businesses to apply custom rules and checks without relying solely on another language model for validation.
The model reflects an emerging enterprise approach in which AI agents handle reasoning and recommendations while traditional software continues to control high-impact execution.
Disclaimer: This article may include information derived from interviews, press releases, public statements, research, company communications and other publicly available or third-party sources. Such material may be summarised, paraphrased or contextualised for journalistic and editorial purposes. All rights in third-party content remain with their respective owners.