Lufthansa Group – 16M Automated Conversations Across Voice & Chat

Airline Reality

Weather, strikes, reroutes, crew timing, and baggage issues create multilingual surges across phone, web, and app. Passengers expect 24/7 answers that are consistent across channels and devices.

System Design

Unified intents for voice and chat; integrations for PNR, seat maps, vouchers, bag tracing, and disruption feeds; persistent context so a chat can pick up where a call left off. The assistant executes routine steps and passes complex re-accommodation to humans with full context.

Architecture at a Glance

1) Omnichannel ingress (telephony, web, app). 2) ASR for voice, NLU shared across channels. 3) Orchestration layer into PNR, inventory, disruption ops, vouchers, and payments. 4) PCI-scoped payment capture. 5) Analytics for AHT, containment, CSAT, and deflection.

Disruption Playbooks

When weather or ATC hits, the assistant doesn’t apologize and stall; it offers options: earlier/later departures, alternate routings, overnight vouchers if eligible, and callbacks. Read-backs confirm city pairs, dates, and fare implications.

Names, Codes, and Ambiguity

City names map to IATA and local variants; flight numbers confirm with carrier and date; names and PNRs read back with partial masking. If ambiguity persists, the assistant surfaces the top two valid options rather than guessing.

Payments, PCI, and Refunds

Payment capture remains inside a PCI-scoped microflow with DTMF masking on voice. Refunds and exchanges respect fare rules and settlement windows and are narrated clearly to reduce disputes and chargebacks.

Enterprise Outcomes

Around 16 million conversations per year are automated across voice and chat. AHT stabilizes during disruptions; self-service rises; and agents preserve time for complex re-accommodation and special assistance.

Quality Benchmarks (Indicative)

• Intent accuracy ≥ 90% on top flows • Named-entity precision high for cities/IATA and dates • Read-back compliance on critical fields • FCR improves as ambiguity falls • CSAT lifts when agent edits drop under 60–90 seconds.

Sample Dialog (Re-Accommodation)

Passenger: “My flight to FRA is delayed; what can I do?” • Assistant: “I see the delay. I can offer a 19:35 departure via MUC arriving 22:50, or stay on your current flight with a hotel voucher if you miss the connection. Which do you prefer?”

Accessibility & Languages

Bilingual prompts (DE/EN) with quick language switching; option to slow speech; SMS/email mirrors for hearing-impaired travelers; careful read-backs for names and dates to avoid mis-bookings.

Change Management

Roll out with high-volume intents; publish confirmation style guides; appoint contact-center champions; measure edit time and escalation reasons to refine language and flows where travelers stumble.

Pitfalls & Anti-Patterns

Guessing on ambiguous city names; hiding disruptions behind apologies; mixing payment digits with general dialog. Keep payment isolated; surface alternatives proactively; never assume date or station without a read-back.

Implementation Timeline (Indicative)

Weeks 0–2: governance and PCI scoping • 2–4: ASR/NER tuning for cities/IATA • 4–6: pilot buy/change/status • 6–10: add vouchers, bag tracing, disruption playbooks • 10+: expand languages, refine analytics, harden edge-case routing.

Mondial AI Approach

Large airline deployments show how multilingual assistants steady operations when language, integrations, and trust are designed carefully. At Mondial AI we build in that spirit—airline-grade NER for cities and IATA codes, RAG over live ops data, disruption playbooks, and PCI-scoped payments. In practice, we often push accuracy and first-contact resolution further—tightening confirmations and exposing alternatives sooner—so routine requests finish quickly and agents stay available for travelers who need human judgment.