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] The $53.9 Million Contact-Center Case: TELUS and Agentic AI

    How TELUS expanded conversation analysis from 200,000 to more than 22 million calls a year and reported faster issue detection and large annual operating savings.

    Conscious Engines

    Industry: Telecommunications
    Organization: TELUS
    Use case: Contact-center automation and conversation intelligence
    Evidence basis: Google Cloud customer case using TELUS-reported results
    Disclosure: This is an independent analysis by Conscious Engines. The source is published by the cloud provider and the savings figure is company reported.

    1. Outcome at a Glance

    TELUS expanded automated conversation analysis from roughly 200,000 calls to more than 22 million calls per year. It reported that AI automated 30% of contact-center traffic, identified issues 87% faster, and produced $53.9 million in annual operational savings.

    Key Outcomes

    More than 22 million per year

    Calls analyzed after expansion

    Current reported scale.

    30%

    Contact-center traffic automated

    Company reported.

    87%

    Faster issue identification

    Company reported.

    $53.9 million

    Annual operational savings

    Company accounting, methodology not fully published reported in the cited case.

    MeasureReported resultEvidence note
    Calls analyzed before expansionAbout 200,000 per yearPrior scale
    Calls analyzed after expansionMore than 22 million per yearCurrent reported scale
    Contact-center traffic automated30%Company reported
    Faster issue identification87%Company reported
    Annual operational savings$53.9 millionCompany accounting, methodology not fully published

    The increase from 200,000 to 22 million represents more than a hundredfold expansion in analytic coverage. That changes what the organization can detect. Sampling can reveal common issues. Near-complete analysis can identify emerging failure patterns, segment them, and connect them to product or network causes.

    2. The Operational Problem

    Contact centers contain a high-volume record of what customers cannot solve. Manual quality teams can review only a small sample. Important patterns may remain invisible until complaint volume or churn increases.

    Automated conversation intelligence has to solve several linked tasks: transcribe accurately, separate speakers, identify intent and outcome, detect sentiment and compliance signals, summarize the interaction, and aggregate patterns across millions of calls. Automation adds another layer by answering or completing eligible requests.

    Telecom speech is difficult because calls contain product names, account identifiers, addresses, plan terms, accents, background noise, and emotional speech. A generic transcription can corrupt the exact entities needed for resolution.

    The economic opportunity sits both inside and outside the center. A model can reduce handle time and automate simple contacts. More importantly, it can identify the product, billing, network, or process defect causing repeated calls.

    3. What Was Built

    The Google Cloud account describes a platform that combines contact-center AI and Gemini Enterprise for customer experience.

    System at a Glance

    Telecom ASR

    Transcribe calls and preserve critical entities.

    Conversation models

    Classify intent, outcome, sentiment, and compliance.

    Agent assistance

    Retrieve guidance and recommend the next action.

    Self-service agents

    Resolve bounded requests through approved tools.

    Analytics

    Aggregate causes and detect emerging issues.

    Enterprise action

    Route findings to product, network, and operations teams.

    LayerRole
    Telecom ASRTranscribe calls and preserve critical entities
    Conversation modelsClassify intent, outcome, sentiment, and compliance
    Agent assistanceRetrieve guidance and recommend the next action
    Self-service agentsResolve bounded requests through approved tools
    AnalyticsAggregate causes and detect emerging issues
    Enterprise actionRoute findings to product, network, and operations teams

    At more than 22 million calls, model cascades become important. A small speech or classifier model can handle common traffic. High-uncertainty calls can route to stronger models or human review. Batch analytics can use a different cost and latency profile from live agent assist.

    The contact center should also maintain a claim-to-evidence chain. If a dashboard says a billing issue is growing, analysts need sample calls, transcript spans, affected products, and confidence, not only a generated explanation.

    4. How It Reached Production

    TELUS's scale suggests a transition from sampled analytics to an enterprise-wide signal layer.

    Evaluate transcription by business entity. Word error rate should be supplemented with accuracy for plan names, locations, prices, dates, and account details.

    Separate containment from resolution. An automated call is successful only if the customer's issue is solved without a rapid repeat contact or downstream correction.

    Close the cause loop. Analytics should create owner-assigned issues for product and network teams. Faster detection matters only if corrective action follows.

    Audit savings. The $53.9 million figure should be decomposed into labor capacity, avoided contacts, shorter handling, quality work, and other benefits. Finance should approve realization rules.

    Monitor customer cohorts. Accuracy and containment should be reported by language, accent, channel, issue, and vulnerability group to prevent aggregate results from hiding service gaps.

    5. What Telecommunications Leaders Should Take Away

    TELUS demonstrates that contact-center AI can become enterprise intelligence. The more than hundredfold expansion in analyzed calls can reveal the operational causes behind demand, while automation addresses suitable contacts directly.

    A production telecom speech and action stack combines domain ASR, real-time agent assist, a constrained voice agent, enterprise RAG, conversation analytics, and root-cause routing. Sensitive calls can run on private infrastructure, with task-specific models handling the bulk of traffic and larger models reserved for difficult cases.

    The core metric is cost per durably resolved customer issue. It should be paired with repeat contact, complaint, churn, agent effort, and model cost. The public $53.9 million figure establishes that the upside can be material, but each operator must prove the accounting on its own baseline.

    Sources