We’re putting $5.5M behind research and development of smaller, specialized AI models for real-world deployment — Read our manifesto →
    All posts

    [Case Study] 17 Million Queries at 97% Containment: Air India's AI.g Assistant

    How a rebuilt data foundation and generative assistant combined high-volume customer service with notifications, analytics, and operational visibility.

    Conscious Engines

    Industry: Aviation and travel
    Organization: Air India
    Use case: Generative customer service and enterprise digital operations
    Evidence basis: Air India first-party technology-transformation reporting
    Disclosure: This is an independent analysis by Conscious Engines. Results are company reported and the public account does not provide independent containment validation.

    1. Outcome at a Glance

    Air India reports that its AI.g assistant has handled more than 17 million customer queries, supports more than 18,000 sessions per day, and achieves 97% containment.

    Key Outcomes

    More than 17 million

    Customer queries

    Cumulative scale reported in the cited case.

    More than 18,000

    Daily sessions

    Operating volume reported in the cited case.

    97%

    Containment

    Company-defined and reported.

    About 1 million per day

    Notifications

    Messaging scale reported in the cited case.

    95.5%

    Notification delivery

    Reported delivery rate.

    MeasureReported resultEvidence note
    Customer queriesMore than 17 millionCumulative scale
    Daily sessionsMore than 18,000Operating volume
    Containment97%Company-defined and reported
    NotificationsAbout 1 million per dayMessaging scale
    Notification delivery95.5%Reported delivery rate
    Enterprise data sources45Data-platform breadth
    Data platformAbout 800 TBReported scale
    Management information1,500 KPIs and 300 dashboardsOperating visibility
    Internal usersMore than 1,000Analytics adoption
    Customer profilesAbout 80 millionData foundation

    The containment rate is unusually high. Without a published definition, it should be treated cautiously. Some organizations count a conversation as contained when no human transfer occurs, even if the customer abandons or returns later. Durable resolution is the stronger measure.

    2. The Operational Problem

    Airline customer service is event driven. Weather, delays, schedule changes, baggage issues, refunds, and airport disruptions can create sudden contact spikes. Customers want current, itinerary-specific answers across channels and time zones.

    A generic chatbot cannot solve this reliably without live booking, flight, policy, airport, and customer context. Static FAQs become outdated during disruption. Incorrect advice about check-in, refund, or rebooking can increase cost and damage trust.

    The public transformation account makes the data dependency visible. AI.g sits alongside a platform integrating 45 sources, 800 TB of data, 1,500 KPIs, and approximately 80 million customer profiles. The assistant's apparent intelligence depends on the availability and quality of that operational context.

    3. What Was Built

    AI.g is part of a wider digital core rather than a stand-alone chat interface.

    System at a Glance

    Customer identity

    Resolve profile, booking, and journey context.

    Operational data

    Supply flight, disruption, baggage, and notification state.

    Policy retrieval

    Ground answers in current fare and service rules.

    Conversational model

    Understand requests and explain options.

    Action tools

    Complete eligible changes or route them correctly.

    Notification platform

    Deliver proactive journey information at scale.

    LayerFunction
    Customer identityResolve profile, booking, and journey context
    Operational dataSupply flight, disruption, baggage, and notification state
    Policy retrievalGround answers in current fare and service rules
    Conversational modelUnderstand requests and explain options
    Action toolsComplete eligible changes or route them correctly
    Notification platformDeliver proactive journey information at scale
    AnalyticsMonitor demand, containment, failure, and operational causes

    The model stack should be routed by task. Lightweight classifiers can identify language and intent. Structured systems return flight or booking status. Retrieval supplies policy. Generative models explain complicated options. High-impact changes require authentication, confirmation, and deterministic transaction controls.

    Voice can extend the same architecture to calls, airports, and accessibility, using airline-tuned speech recognition for names, airport codes, dates, and booking identifiers.

    4. How It Reached Production

    Air India's case demonstrates that customer AI needs a digital foundation.

    Integrate real-time sources. During disruption, stale data is worse than no answer. Systems need freshness indicators and fallback behavior.

    Define containment rigorously. Count a request as resolved only when the objective is completed and no rapid repeat contact follows.

    Confirm high-impact actions. Rebooking, cancellation, payment, and refund changes should present the exact consequence before execution.

    Design for spikes. Models, data services, and contact-center handoffs must handle irregular peaks without severe latency.

    Measure by journey. Track resolution, transfer, repeat contact, complaint, latency, and cost for booking, baggage, disruption, refund, and loyalty separately.

    Proactive notification is part of the service model. One million messages per day can prevent contacts if they arrive accurately and on time, but a 95.5% delivery rate still leaves a meaningful unserved population at that scale.

    5. What Retail, Hospitality, and Travel Leaders Should Take Away

    Air India's 17 million-query scale shows that generative service can operate at airline volume. The deeper lesson is that the assistant was built on a consolidated operating and customer data layer.

    A production travel model stack combines speech-to-text, text-to-speech, multilingual voice and chat agents, policy RAG, booking and disruption tools, notification intelligence, and strict transaction controls. Smaller task models can handle the majority of predictable intents, lowering latency and cost.

    The primary metric is cost per durably resolved journey need, supported by containment definition, repeat contact, action accuracy, satisfaction, and disruption performance. A 97% headline is useful only when the denominator and outcome are transparent.

    Sources