A telecom call is rarely a simple question. "My bill is wrong" may involve a plan change, prorating, roaming, device finance, tax, discount expiry, and an outage credit. A generic voice bot can sound natural while retrieving the wrong account or inventing a remedy.
A production telecom voice agent needs specialized speech, live account context, approved journey logic, and explicit authority.
Proof that scaled resolution is possible
Vodafone's FY2026 investor results report that SuperTOBi operated across all European markets with more than 70% end-to-end resolution and an eight-point NPS improvement. An earlier Vodafone description said TOBi handled about 15 million conversations per month across 15 markets.
The figures come from Vodafone and span different product generations. They establish scale, but a new buyer should define resolution consistently: the requested task completed correctly, written to the system, and not repeated within a defined period.
Journeys to automate first
Start with high-volume, reversible tasks:
- payment date and balance explanation
- usage and allowance checks
- appointment confirmation
- installation status
- outage confirmation and updates
- SIM activation and simple device setup
- address or contact updates after authentication
- add-on activation within preset rules
Leave disputed charges, vulnerable customers, cancellation, fraud, complex collections, and nonstandard compensation under human review until the system proves itself.
The specialized voice pipeline
- detect language, speech, and interruptions
- transcribe with telecom vocabulary and confidence
- authenticate through approved factors
- resolve the account, service, device, and active journey
- retrieve live billing, order, and network state
- interpret intent and fields with an SLM
- apply deterministic journey and authority rules
- generate a short grounded response
- confirm consequential details
- execute, log, and offer human transfer
The language model interprets. The billing and order systems remain the source of truth.
What domain ASR must recognize
Build tests for:
- account and phone numbers
- dates, currency, data quantities, and time windows
- plan, bundle, and promotion names
- device models and operating systems
- addresses and local place names
- network terms such as 5G, eSIM, roaming, and Wi-Fi calling
- multilingual and code-switched speech
- noisy mobile audio and speakerphone
Report critical-entity error separately from word error. Confirm identifiers and monetary commitments before action.
Ground answers in three contexts
Customer context: product, eligibility, permissions, history, consent, and current orders.
Commercial context: active price, promotion, terms, effective date, and jurisdiction.
Network context: outage, coverage, device compatibility, recent alarms, and estimated restoration.
RAG is appropriate for product and procedure content. APIs are appropriate for live balance, order, and network facts. Do not embed fast-changing account data into a static vector index.
Escalation is part of resolution
Transfer when identity fails, the customer disputes a charge, vulnerability is detected, fraud is suspected, cancellation intent is explicit, the policy does not support a remedy, or the customer requests a person.
The human should receive the transcript, verified identity state, extracted issue, actions already attempted, and retrieved evidence. A warm transfer that prevents repetition is a successful automated contribution even when containment is zero.
Metrics that resist gaming
Evaluation Scorecard
Verified task completion
requested action correctly finished.
Repeat-contact rate
same issue returns within 3 to 14 days.
Incorrect commitment rate
agent promises unsupported price, credit, or time.
Critical-entity accuracy
IDs, dates, amounts, plans, and devices.
Escalation recall
defined risk cases transferred.
Transfer completeness
human receives usable context.
| Metric | Definition |
|---|---|
| verified task completion | requested action correctly finished |
| repeat-contact rate | same issue returns within 3 to 14 days |
| incorrect commitment rate | agent promises unsupported price, credit, or time |
| critical-entity accuracy | IDs, dates, amounts, plans, and devices |
| escalation recall | defined risk cases transferred |
| transfer completeness | human receives usable context |
| customer effort | turns, repetition, and time to resolution |
| cost per durable resolution | total cost divided by non-repeated completed cases |
Do not optimize average handle time in isolation. A shorter first call followed by a second call increases cost and frustration.
A controlled rollout
Replay 1,000 to 5,000 historical conversations against the proposed journey. Create synthetic edge cases for rare vulnerability, fraud, and identity paths. Launch with one to five percent of traffic and daily sample review. Expand by intent, language, and customer segment only after stable performance.
Use a smaller model for intent, extraction, and routine response when it passes the test set. Route ambiguous conversations to a stronger model or a person. Track cost per verified outcome rather than cost per minute.
The conclusion
A telecom voice agent succeeds when it knows what the customer owns, what the network is doing, which commercial terms are current, and what it is authorized to change.
Natural speech is only the surface. Durable resolution comes from task-specific models, live systems, controlled journeys, and measurable handoffs.
Research note
Research is current through September 5, 2026. Vodafone figures are company-reported. Customer-service rules, recording, authentication, payment, and vulnerability requirements vary by market.
Continue the research
- The complete telecommunications AI model stack
- The telecom network-operations AI copilot
- RAG vs fine-tuning vs bespoke AI models
- Why an enterprise evaluation set becomes an AI moat
- Voice AI for reservations, orders and disruption
Building a Production-Ready System
Conscious Engines builds telecom voice AI that recognizes plan names, device models, account details, billing language, and network symptoms. The agent retrieves current policy and live customer state, then completes only explicitly permitted actions with confirmation and audit.