A service bot can answer when check-in starts. A production voice agent must understand the customer, find the live reservation or order, explain the applicable policy, offer only feasible choices, confirm important details, complete an authorized change, and leave an auditable record.
The gap between those systems is where most enterprise risk and value sit.
Scale is possible, but containment is not the whole result
Air India reports that its AI.g system has processed more than 17 million queries, serves more than 18,000 sessions each day, and contains 97% of contacts. In retail, IKEA parent Ingka reports 3.2 million interactions, 47% inquiry resolution, and approximately EUR 13 million in savings from Billie over 2021 to 2023.
These outcomes show that digital agents can handle large volumes. They do not prove that every contained journey ended correctly. A serious deployment should define containment as a verified business outcome, not the absence of a transfer.
The real-time voice stack
The real-time voice stack
Telephony and audio
The system needs low-latency streaming, echo cancellation, voice activity detection, interruption handling, and graceful recovery.
Domain ASR
Speech recognition must capture names, addresses, booking references, product codes, dates, times, flight or train numbers, hotel names, quantities, currency, and loyalty tiers.
Dialogue and policy model
A specialized language model identifies the journey, gathers required fields, asks a minimal clarifying question, and selects a permitted tool. It should know what it cannot do.
Enterprise retrieval
RAG returns the current fare rule, return policy, property information, disruption policy, baggage allowance, loyalty benefit, or accessibility procedure.
Transaction tools
APIs search inventory, price options, modify the booking, create a return, issue an approved voucher, send a payment link, or notify a human.
TTS and confirmation
Speech should sound clear rather than theatrical. Read back names, dates, routes, quantities, prices, refund amounts, and consequences.
1. Telephony and audio
The system needs low-latency streaming, echo cancellation, voice activity detection, interruption handling, and graceful recovery. It should support telephone bandwidth, background noise, and a user switching languages.
2. Domain ASR
Speech recognition must capture names, addresses, booking references, product codes, dates, times, flight or train numbers, hotel names, quantities, currency, and loyalty tiers. Evaluate these entities separately from word error rate.
3. Dialogue and policy model
A specialized language model identifies the journey, gathers required fields, asks a minimal clarifying question, and selects a permitted tool. It should know what it cannot do.
4. Enterprise retrieval
RAG returns the current fare rule, return policy, property information, disruption policy, baggage allowance, loyalty benefit, or accessibility procedure. Retrieval must filter by brand, market, product, date, channel, and customer status.
5. Transaction tools
APIs search inventory, price options, modify the booking, create a return, issue an approved voucher, send a payment link, or notify a human. The tool validates the action independently of the language model.
6. TTS and confirmation
Speech should sound clear rather than theatrical. Read back names, dates, routes, quantities, prices, refund amounts, and consequences. The final confirmation should state what changed and what did not.
Design every journey as a state machine
For a reservation change:
- Identify and authenticate the customer.
- Retrieve the exact reservation.
- Confirm which segment, guest, or item is changing.
- Retrieve governing policy and entitlement.
- Search live eligible alternatives.
- State price difference, restrictions, and availability.
- Obtain explicit confirmation.
- Execute through the system of record.
- Verify the returned transaction state.
- Send written confirmation and save the audit record.
The model manages natural language around the states. It cannot skip a required state because the customer sounds certain.
Disruption requires a different objective
During a major disruption, demand spikes while inventory and human attention become scarce. The agent should prioritize safety, vulnerable customers, time-critical journeys, and policy-defined eligibility. It needs fresh operations data and a degradation mode if a downstream system is stale.
Offer ranked feasible options with clear consequences. Do not promise a room, seat, refund, or delivery until the system of record confirms it. If no option exists, say so and preserve the customer's place in the escalation queue.
Air India says its broader communications platform delivers about 1 million notifications daily with a 95.5% delivery rate. This highlights another requirement: agent actions and proactive communications should share a consistent event source.
Trust should be visible in the interface
The 2026 Expedia Group survey found only 8% of 5,700 surveyed consumers comfortable with AI booking autonomously, while 68% preferred booking through a trusted brand.
Practical trust features include:
- Identify the agent as AI.
- Show or read the source of policy-relevant answers.
- State whether price and availability are live.
- Confirm sensitive fields and material consequences.
- Offer a human path without forcing repeated explanation.
- Send a written transaction record.
- Never request full payment credentials in an uncontrolled model context.
Metrics that prevent false success
| Layer | Metric |
|---|---|
| Audio | end-to-end latency, interruption recovery |
| ASR | booking, date, route, price, and identity accuracy |
| Intent | journey accuracy, safe abstention |
| Retrieval | policy recall, source freshness, citation precision |
| Tool use | valid call rate, unauthorized action rate |
| Journey | verified completion, transfer, repeat contact |
| Customer | complaint, cancellation, satisfaction, accessibility outcome |
| Economics | cost per correct resolution, incremental revenue or retained value |
Audit contained failures. Sample calls the system marked successful and confirm the system-of-record outcome, subsequent contact, refund, or complaint.
Multilingual design
Translate meaning, not only words. Preserve names, codes, prices, dates, and legal or policy terms. Evaluate native speakers on real journeys. A model may be fluent in a language but weak on a brand's local catalog or travel vocabulary.
Walmart's 44-language associate translation deployment shows the operating scale possible for language support, although associate communication is a different risk context from customer transactions.
A safe rollout
Start with one low-ambiguity journey and read-only system access. Add transactional tools after the model demonstrates critical-entity accuracy and policy retrieval. Cap transaction value, require confirmation, and maintain immediate rollback or correction paths.
Pilot across accents, languages, noisy environments, accessibility needs, angry callers, system latency, and partial outages. Include users who do not speak in concise commands.
The conclusion
The enterprise voice agent is a natural-language layer over a controlled transaction system. Its usefulness comes from handling the messy way people speak. Its safety comes from narrow authority, live data, deterministic validation, and explicit confirmation.
When those pieces work together, the agent can resolve real journeys rather than merely deflect contact.
Research note
Research is current through September 5, 2026. Deployment metrics are company reported. Recording, authentication, payment, refund, consumer, accessibility, loyalty, and travel obligations vary by jurisdiction and workflow.
Continue the research
- The complete retail, hospitality and travel AI stack
- Frontline AI for retail, hospitality and travel
- RAG vs fine-tuning vs bespoke AI models
- Why an enterprise evaluation set becomes an AI moat
- Telecom voice AI for customer service
Building a Production-Ready System
Conscious Engines builds voice AI for reservations, orders, and disruption as a natural-language layer over controlled transaction systems. Domain speech and policy retrieval handle the conversation, while explicit tools validate price, inventory, entitlement, payment boundaries, confirmation, and final system state.