Air India cut refund times from 14 days to 4 hours with AI agents
After automating refunds, Air India set AI agents on email triage, name changes and agent assist — name changes fell from three days to 30 minutes.
What was changed
Air India's first agentic-AI deployment chose the hardest service workflow: refunds, a process spanning multiple validations, policy checks, backend-system interactions and cross-team dependencies. By orchestrating it end to end, the airline says refund turnaround fell from approximately 14 days to about four hours. That result became the platform for a September 2026 expansion into three further processes: automated email resolution, name-change automation, and a knowledge agent for service teams.
The email agent identifies multiple intents in a single customer email — refunds, baggage, loyalty and more — invokes specialised sub-agents, validates information, executes actions and consolidates one response from approved templates, sending automatically only above a 95% confidence threshold with human oversight otherwise. Name changes, which need eligibility checks and ticket reissuance often under imminent-travel pressure, went from about three days to 30 minutes.
A conversational knowledge agent gives service teams policies, fare rules and operational guidance in real time, cutting retrieval time and hold times during live interactions while improving response consistency, agent onboarding and first-contact resolution rates, according to the airline.
The programme runs on a cloud-first digital ecosystem spanning more than 140 enterprise systems, with more than 30 agentic AI initiatives underway, each evaluated against four pillars: increasing revenue, reducing cost, enhancing customer experience, and enabling new capabilities. The figures are the airline's own, reported through ET CIO.
Why it worked
The pilot targeted the workflow with the most systems and policy dependencies, so success proved the integration architecture.
Automated responses are gated by a 95% confidence threshold, keeping humans in the loop below it.
Approved templates and ambient learning from agent interventions keep quality and prompts improving.
A unified customer-data foundation built with Salesforce preceded the agentic layer, so agents act on one record rather than 140 systems.
What can be applied
Automate the ugliest workflow first: refunds touched every system and policy, so the pilot doubled as the integration layer every later agent reuses.
Aftermath
Air India says the foundation will keep supporting customer service as deployments scale across its operations, with more of the 30-plus agentic initiatives moving into production under the same four-pillar evaluation.