Ashfaque Rifaye

AT&T GenAI Virtual Assistant

Scaling enterprise conversational AI across voice, chat and messaging.

AT&T’s consumer business talks to customers across chat, voice, WhatsApp and RCS. The virtual assistant is the first line of that conversation, and it handles 1.5M+ customer interactions a month.

I own it end to end as product owner: conversation design, retrieval, the integrations that connect it to telephony and CRM, the KPI framework, and the roadmap that two development squads deliver.

Role
Product owner (AI Technical Business Analyst, PM)
Org
AT&T Consumer Technology & Experience
Period
Aug 2022 – present
Team
2 squads, 18 developers; 12-person cross-functional core team
Platform
Google CCAI · Dialogflow CX · RAG · REST APIs
Channels
Chat · Voice · WhatsApp · RCS
Monthly customer interactions
1.5M+
Containment rate
55%
Fewer live-agent escalations
28%
Estimated annual operational savings
$4.2M
Contents
  1. 01Context
  2. 02Problem
  3. 03Opportunity
  4. 04Product thesis
  5. 05Users and journey
  6. 06My role
  7. 07Key decisions
  8. 08Solution
  9. 09Technical architecture
  10. 10Delivery and execution
  11. 11Metrics
  12. 12Business impact
  13. 13What changed
  14. 14What I learned
01

Context

Customer care at telecom scale is a volume business. Every contact the assistant resolves is one a live agent doesn’t have to take, and every contact it can’t resolve still has to end well.

The assistant runs on Google Contact Center AI and Dialogflow, with generative AI and retrieval layered in. It serves AT&T’s consumer line of business across four channels, so one conversation model has to behave well in a voice call and in a WhatsApp thread.

02

Problem

Care was carrying high volumes of repetitive contacts across chat, voice, WhatsApp and RCS. Handle times were long, and live-agent escalations were expensive. Three things made it harder than adding a chatbot:

  • Answers had to be current and approved. Plans, policies and procedures change, and an assistant answering from stale content loses customer trust quickly.
  • Many issues need the customer’s account. Without integration into backend systems, the assistant can explain but not resolve.
  • Escalations started cold. When the assistant handed off, the agent needed context the assistant already had.
03

Opportunity

Two levers had direct cost impact: containment, the share of conversations resolved without a live agent, and handle time on the conversations that do reach one. At 1.5M+ interactions a month, a few points of containment or a few seconds of handle time move a seven-figure number.

A third lever protected the first two: quality. Containment bought with wrong answers returns as repeat contacts and lower satisfaction, so the goal was containment with guardrails, not containment at any cost.

04

Product thesis

An assistant earns containment when it understands the intent, answers from current and approved knowledge, can act on the account, and hands off with context when it can’t.

That put intent design, retrieval quality, integrations and graceful escalation ahead of model choice. The model matters, but it is one layer of several.

05

Users and journey

Four groups depend on the assistant, and each one judges it differently:

  • Customers on four channels want a resolution, not a conversation.
  • Live agents receive the escalations and need the context that came before them.
  • Care operations and leadership run the business on containment, handle time and satisfaction.
  • Business SMEs own the knowledge the assistant answers from.
Fig. 1.1The customer journey through the assistant
  1. 01AskCustomer starts on any channelChat, voice, WhatsApp or RCS
  2. 02UnderstandIntent recognized350+ intents, 1,200+ training phrases92% intent accuracy
  3. 03Answer or actKnowledge or account actionRetrieval over approved knowledge, or a backend API call
  4. 04ResolveContained in the assistant55% containment
  5. 05Hand offEscalated with contextThe agent receives the conversation so far32 sec lower handle time
Accuracy, containment and handle-time figures are program-level results from the KPI dashboards.
06

My role

I own the assistant end to end. In practice:

  • Product ownership and roadmap. I align the AI roadmap with CX objectives alongside US product owners, enterprise architects and business SMEs. That work secured a $2.5M budget for FY25 enhancements.
  • Conversation design. The intent model (350+ intents, 1,200+ training phrases) and the retrieval-backed answers behind it.
  • Integration architecture. How the AI layer connects to telephony infrastructure and backend CRMs through REST APIs.
  • KPI strategy. Power BI and SQL dashboards across 15+ metrics, including containment, CSAT, intent accuracy and fallback rate.
  • Delivery. Backlog grooming and sprint planning for two squads (18 developers) in SAFe, at 95% sprint-commitment reliability.
  • Documentation. Process flows, data-flow diagrams and BRDs for a 12-person cross-functional team across engineering, UX and operations.
07

Key decisions

These choices shaped the product. Each one traded something away.

  1. AGround answers in retrieved knowledge, not only scripted responses

    WhyScripted responses go stale as plans and policies change. Retrieval over approved knowledge keeps answers current.

    Trade-off managedRetrieval quality becomes a product surface to manage: content ownership, freshness and review.

  2. BRoute low-confidence answers to human review

    WhyA confident wrong answer costs more trust than a handoff. Flagging low-confidence responses keeps quality visible.

    Trade-off managedReview workload for the team, in exchange for a feedback loop that improves intents and knowledge.

  3. CTreat escalation as part of the product

    WhyNot every contact should be contained. Passing context to the agent through telephony and CRM integration makes the handoff faster.

    Trade-off managedMore integration scope and dependency on backend teams, repaid as 32 seconds lower handle time per interaction.

  4. DRead containment alongside quality guardrails

    WhyContainment alone rewards a bot that never lets go. Intent accuracy, fallback rate and CSAT keep it honest.

    Trade-off managedMore metrics to maintain, and containment targets that have to answer to quality.

  5. ETreat tokens and latency as product constraints

    WhyGenAI cost and response time grow with volume, so token usage and API cost are tracked with engineering next to response times.

    Trade-off managedCost and speed are weighed against response richness on every GenAI feature.

08

Solution

An omnichannel GenAI virtual assistant on Google CCAI and Dialogflow CX:

  • One conversation model, four channels: chat, voice, WhatsApp and RCS.
  • 350+ intents and 1,200+ training phrases, with generative responses grounded by retrieval over approved knowledge.
  • REST integrations to telephony infrastructure and backend CRMs for real-time context.
  • Human-in-the-loop review of low-confidence responses.
  • An AI performance analytics suite: real-time monitoring across 6 channels and 15+ KPIs.
09

Technical architecture

A simplified view of the request path, top to bottom, and the loop that improves it. Vendor internals are omitted.

Fig. 1.2AI architecture: request path and improvement loop
  1. 1
    CustomerConsumer line of business
    • Consumer customers
  2. 2
    ChannelOmnichannel entry
    • Chat
    • Voice (telephony)
    • WhatsApp
    • RCS
  3. 3
    Conversation layerGoogle CCAI · Dialogflow CX
    • Dialog and session state
    • Channel handling
  4. 4
    Intent and LLM processingUnderstanding and generation
    • 350+ intents
    • 1,200+ training phrases
    • Generative responses
  5. 5
    RAG and knowledgeGrounding
    • Approved knowledge sources
    • Retrieval
    • Grounded answer
  6. 6
    Backend APIsREST
    • Context lookup
    • Account actions
  7. 7
    CRM, telephony and enterprise systemsSystems of record
    • CRM
    • Telephony infrastructure
    • Enterprise systems
  8. 8
    Resolution or escalationOutcome
    • Resolved in the assistant · 55%
    • Live agent, with context
  1. 1Customer. Consumer customers reaching AT&T care.
  2. 2Channel. One conversation model serves all four channels.
  3. 3Conversation layer. Dialogflow CX manages the conversation, its state and channel specifics.
  4. 4Intent and LLM processing. Intent recognition runs at 92% accuracy; generative AI shapes the reply.
  5. 5RAG and knowledge. Answers are grounded in retrieved, approved knowledge instead of static scripts.
  6. 6Backend APIs. REST APIs enrich each conversation with real-time customer context.
  7. 7CRM, telephony and enterprise systems. The integration ecosystem I architected connects the AI layer to these systems.
  8. 8Resolution or escalation. Escalations fell 28%; context-rich handoffs cut handle time by 32 seconds.
10

Delivery and execution

Two development squads (18 developers) deliver the roadmap in SAFe. I own the backlog and sprint planning; the squads hold 95% sprint-commitment reliability, tracked in Jira and Confluence.

Fig. 1.3From roadmap to release
  1. 01AlignRoadmapTied to CX objectives with US product owners, architects and SMEs$2.5M FY25 budget
  2. 02SpecifyBacklogBRDs, process flows and data-flow diagrams
  3. 03BuildTwo squads18 developers in SAFe sprints95% sprint reliability
  4. 04ShipReleaseChat, voice, WhatsApp and RCS
  5. 05LearnMeasure15+ KPIs feed the next iteration

Stakeholders span engineering, UX and operations in India and product, architecture and business teams in the US. Process flows, data-flow diagrams and BRDs give that 12-person cross-functional team one shared picture.

11

Metrics

Containment is the headline, and it is always read with the guardrails that keep it honest.

Fig. 1.4Measurement: watch, move, prove
  1. WatchQuality guardrails
    Intent recognition accuracy
    92%
    Fallback rate
    CSAT
    KPIs monitored
    15+
  2. MoveOperational levers
    Containment
    55%
    Live-agent escalations
    −28%
    Average handle time
    −32s
  3. ProveBusiness outcomes
    Estimated annual savings
    $4.2M
    NPS points
    +8
    Monthly interactions
    1.5M+
Savings are an estimate of annual operational cost avoided.
12

Business impact

  • An estimated $4.2M a year in operational savings from higher containment and fewer escalations.
  • A $2.5M budget for FY25 enhancements, secured by tying the roadmap to CX objectives.
  • +8 NPS points from optimizations the KPI dashboards made visible.
  • AT&T Connection Award (2023) for the virtual assistant team, from the Consumer Technology & Experience organization.
13

What changed

Knowledge
BeforeScripted responses that needed manual upkeep
AfterRetrieval over approved knowledge grounds each answer
Escalation
BeforeAgents rebuilt context the assistant already had
AfterContext travels through telephony and CRM integration: 32 seconds saved per interaction
Quality
BeforeLow-confidence answers went unseen
AfterLow-confidence answers are flagged for human review
Measurement
BeforePerformance was hard to see across channels
AfterReal-time dashboards across 6 channels and 15+ KPIs

These rows summarize the program’s direction, not a single release.

14

What I learned

  • Containment is a lagging indicator. Intent accuracy and fallback rate show where it is heading, so I watch them first.
  • Retrieval quality is a product surface. An answer is only as good as the knowledge behind it, so content ownership belongs on the roadmap next to features.
  • Escalation is part of the experience. A good handoff saves seconds on every contact, and a contained conversation that should have escalated costs more later.
  • Guardrails build trust. Human review of low-confidence answers keeps quality visible to everyone who depends on the assistant.
Next case studyVerizon Digital CommerceUsing customer analytics, experimentation and product strategy to improve digital commerce journeys.Read the case study

Building an AI product, scaling an enterprise AI experience, or figuring out where AI actually creates value?

I work at the intersection of product strategy, AI systems and execution.

Let’s talkOther ways to reach me

Direct