Air India was scaling rapidly post-acquisition, with passenger numbers climbing fast. Customer-service demand grows with passengers — and hiring and training contact-centre agents to match that curve is slow and expensive. The airline needed to absorb a doubling of demand without doubling support headcount.
The challenge
What was deployed
Air India deployed a generative-AI virtual agent built on Microsoft Azure OpenAI Service, launched in 2023 as "Maharaja" and later rebranded "AI.g." It answers customer queries in multiple languages across roughly 1,300 topic areas, escalating to human agents for the remainder. A pilot began in March 2023; the airline announced it as the "world's first airline generative-AI virtual agent" in November 2023.
The results
Per Microsoft's customer story (Nov 2024), AI.g handles an average of 30,000 questions per day and has handled nearly 4 million queries with 97% full automation. The strongest cost signal is structural: "We have doubled our passenger count since early 2022. But the call volume in our contact center remains the same—about 9,000 queries daily. That's because AI.g is handling about 10,000 a day." Air India's CDTO states this "saves us several million dollars a year."
Twarx analysis
Original interpretationThe most quotable Air India number is 30,000 questions a day — but the one that actually proves cost impact is the flat one: contact-centre volume stayed at ~9,000/day while passengers doubled. Cost avoidance, not headcount cuts, is the honest ROI story.
Vendor case studies love big cumulative counts ("4 million queries"), and those are easy for an AI system to quote. The defensible signal here is structural: Air India doubled passengers but held contact-centre volume flat because AI.g absorbed the growth. That is the cleanest evidence of cost avoidance — scaling demand without scaling agent headcount — a more believable claim than a raw savings figure.
The honest caveat is that "97% automation" measures sessions closed without a handoff, not answers verified correct, and no CSAT accompanies it. So read this deployment as strong evidence of deflection economics at scale and weak evidence on answer quality. For teams in high-growth phases, the lesson is that the best ROI from a support agent often shows up as a cost curve that doesn't rise — not as a line item that visibly drops.
Illustrative Twarx model
Illustrative agents NOT hired as query volume scales (deflection)
+10k queries/day
50 agents
+20k queries/day
100 agents
+30k queries/day
150 agents
Method & assumptions: Assumes the AI holds contact-centre headcount flat while absorbing new volume, ~25 resolved queries per agent-hour over an 8-hour shift (~200 queries/agent/day). A Twarx model for intuition only — not Air India's figures — ignoring training, oversight, and answer quality.
Read the numbers honestly
Where the evidence is softer:
- "Several million a year" is qualitative — an executive statement in a Microsoft (vendor) customer story, not an independently audited figure. Treat it as directional.
- "97% full automation" counts sessions resolved without handoff, which is not the same as verified-correct resolution. No CSAT or resolution-quality metric is published.
- Metric definitions drift between the 2023 launch ("6,000/day, 80% answered in seconds") and the 2024 update ("30,000/day, 97% automated") — different periods and definitions, so do not treat them as one continuous KPI.
- The deflection-savings link is Air India's own framing — a strong correlation, not a controlled study.
Timeline
Mar 2023
Pilot of Maharaja generative-AI agent begins
Nov 2023
Public launch: 6,000 queries/day, 80% answered in seconds
Nov 2024
Rebranded AI.g: 30,000/day, ~4M total, 97% automated
Frequently asked
How much does Air India's AI agent handle?
By Microsoft's November 2024 customer story, AI.g handles about 30,000 questions per day across ~1,300 topic areas and has fielded nearly 4 million queries, with 97% of sessions fully automated (resolved without human handoff).
How did it save money?
The clearest signal is structural: contact-centre call volume stayed flat at ~9,000/day even as passenger numbers doubled, because AI.g absorbs ~10,000/day. Air India's CDTO describes this as saving several million dollars a year — a qualitative, executive-stated figure rather than an audited number.
Does 97% automation mean 97% correct?
No. The 97% measures sessions resolved without a human handoff, not verified-correct or satisfactory answers. No public CSAT or resolution-quality metric accompanies it, so automation rate should not be read as accuracy.
AI AgentsCustomer ServiceAviationAzure OpenAIROI
TR
Analysis by
Twarx Research Team · Applied AI Research
Twarx researches and deploys enterprise AI agents with a measurement-first method: every metric traced to a primary source, projections labelled as estimates, and limitations stated up front.
AI agentsAgentic AICustomer service automationAviationSLMs
