Land & Expand — Agent Assist STT POC as the Entry Point
Land & Expand — Agent Assist STT POC as the Entry Point
Same process I'd run on any F500 pursuit — before building a single slide.
Q1 2026 earnings call (April 21, 2026) + Capital One AI blog. Real pain = what executives say publicly. Fairbank on $8.5B opex pressure, the 84-93% agent relevance disclosure, Brex + Discover integration urgency.
Researched actual leaders: Dugal (Card Tech, built Chat Concierge), Talreja (ex-Goldman CTO), Corey Lee (contact center VP), LaPrade (CRO). Generic targeting doesn't work at F500.
Used COF's disclosed Q1 2026 metrics — $8.5B noninterest expense, 55.57% efficiency ratio, $446B avg loans — plus KPMG, Calabrio, Gartner benchmarks. Every number sourced from the April 21 earnings call.
No product spec leading. Sub-500ms = live agent assist that works. Private VPC = PCI/GLBA unlocked. 30% lower WER = 84→93% pushed to 97%+. Context first.
IBM watsonx = warm intro to Dugal, not cold. Discover closing May 2026 = time-bounded urgency. Changes dynamic from 'roadmap item' to 'needs to happen now'.
The goal: land a 30-day agent assist STT POC — highest ROI, lowest friction, fastest to a measurable result. Every other use case expands from this wedge.
Two contact center stacks actively merging. Every month of delay = more dual-stack cost.
AI-driven handle time reduction is one of the few efficiency levers COF can pull right now.
Discover + Brex + Hopper running simultaneously. Voice AI standardization is immediate.
They publicly disclosed 84→93% search relevance. That IS the POC brief.
The buying committee. Real names, real roles, real reasons each one matters.
Built the gen AI agent servicing tool. Presented Chat Concierge at NVIDIA GTC. Owns the STT infrastructure decision.
Owns auto finance + Financial Services. Auto originations up 21% YoY. Chat Concierge for dealerships is his initiative.
Fairbank's chief of staff. $2.5M RSU for Discover integration work. Controls PCI/compliance vendor sign-off.
Owns all Card + Retail Bank servicing — US, LatAm, Philippines. Ex-BofA. AHT and cost/call are his daily metrics.
Ex-Facebook AI Research, ex-JP Morgan. Leads ML platform including NLP/conversational AI.
CISO since 2020. Cleared Veza — shows vendor POC comfort. Must sign off on any PCI-scoped audio vendor.
Ex-Goldman Sachs Asset Mgmt engineering. API-first. Owns tech stack for the Discover + Brex track.
Presented Q1 2026 financials. $8.5B noninterest expense, 55.57% efficiency ratio. Any deal >$5M needs his awareness.
Public quotes that map directly to Deepgram capabilities.
“These investments will continue to be reflected in the efficiency ratio, but they are also the engine that powers long-term growth.”
Q1 2026: $8.5B noninterest expense (up 43% YoY), efficiency ratio 55.57%. Handle time reduction across millions of calls is one of the few controllable efficiency levers mid-integration.
“Our proprietary gen AI agent servicing tool helps agents resolve questions faster — we saw a shift from 84% to 93% highly relevant search results.”
They built the agent assist layer. Garbage in, garbage out. Deepgram's telephony-specific models cut WER 30% vs Google.
“We made expected progress on the Discover integration and synergies.”
Tech conversion deadline: H1 2027. Brex ($4.5B, closed April 7) + Hopper added. Deepgram = one STT layer, one compliance review across all three brands.
“Domestic card charge-off rate for Q1 was 5.1%… delinquency rate was 3.7%, down 55 bps YoY.”
Capital One built $230M of allowance in Q1 citing 'heightened geopolitical uncertainty.' With 100M+ customers, real-time call QA is a board-level risk tool. Private VPC required.
Every feature tied to a measurable business result and competitive delta.
| Feature | What it does | Business outcome | vs. Google / ElevenLabs |
|---|---|---|---|
| Sub-500ms streaming STT (Nova-3 + Flux) | Real-time transcription fast enough for live agent assist to surface before the agent responds. | 1M+ calls/week × ~45s AHT = ~$35M+ annual labor savings. | Google STT ~800ms — too slow for live prompting. |
| 30% lower WER on telephony audio | Models trained on noisy, compressed, accented phone audio — US, LatAm, Philippines profiles. | 84→93% relevance proven w/ generic STT. Deepgram → 97%+. | Google degrades 15–20% on noisy telephony. |
| Private VPC & on-prem deployment | Deploys inside COF's existing AWS/GCP private cloud. Cardholder audio never touches shared infra. | PCI-DSS, GLBA, CFPB compliance maintained. Unlocks live card servicing. | Google STT: cloud-only — non-starter for live PII. |
| Post-call transcription at 100% volume | Every call transcribed — not 2–5% sampled. Feeds QA, compliance, delinquency sentiment. | 100% compliance vs current 2–5%. Material CFPB risk reduction. | Google: questionable FSI posture. Whisper: no managed service. |
| Bundled Voice Agent API (STT + LLM + TTS) | Single API replaces IVR for top 10 intents. Flux end-of-speech enables sub-1s automated chat. | Gartner: AI cuts CC labor $80B/yr by 2026. 20% containment = ~$10–15M/yr. | Google CCAI requires Dialogflow. ElevenLabs incomplete. |
Current customer state, when I ask for help on disputing a transaction on capitalone.com.
Flux turn-taking, Nova STT, Aura TTS, sentiment + supervisor cockpit on every call.
Sources: Q1 2026 earnings call, SEC filings, EE Times, Calabrio, KPMG, Gartner.
★ Conservative. Excludes FCR + after-call work reduction.
★ Plus immeasurable CFPB risk reduction.
★ Five9 reported 2–4x accuracy gain on alphanumeric.
★ Single compliance posture across 100M customers.
What to say to each stakeholder — their problem, our message, the proof.
Your agent servicing tool hit 93% relevance. The ceiling is the STT layer underneath.
Deepgram cuts WER 30% on telephony audio. Closes the remaining 7% gap with zero changes to your app.
Five9: 2–4x accuracy gain. IBM's first voice partner in watsonx Orchestrate.
Discover integration = two stacks, two compliance postures, two STT reviews — unless you standardize.
Deepgram = single API inside your existing private cloud. One integration across both brands.
AWS, GCP, Azure live. SOC 2 Type II. 140K+ concurrent calls in production.
1M+ calls/week. Every second of AHT is real money. Agent assist isn't fast enough for live use.
KPMG: lookup 1–3 min → 20–30 sec. At your volume, 45s AHT = $18M+ annually.
Sub-500ms latency — suggestions surface before agents need to respond.
Monitoring 2–5% of calls. Post-Discover: 100M+ customers, two brands, active CFPB scrutiny.
100% transcription inside your private VPC. Zero data egress. Every delinquency signal flagged real-time.
SOC 2 Type II. HIPAA BAA. PCI-ready VPC. Production-deployed in regulated FSI.
The only vendor that simultaneously satisfies latency, compliance, deployment, accuracy, and scale.
| Dimension | Deepgram | Google STT | ElevenLabs |
|---|---|---|---|
| Streaming latency | ✓ <500ms (lowest in market) | ✗ ~800ms | △ TTS <100ms / STT not FSI |
| On-prem / private VPC | ✓ AWS, GCP, self-hosted | ✗ Cloud-only | △ No on-prem path |
| PCI / GLBA compliance | ✓ SOC 2 II, PCI-ready, HIPAA BAA | ✗ Shared cloud — disqualifying | △ HIPAA Enterprise-tier only |
| Telephony audio accuracy | ✓ 30% lower WER; built for noise | ✗ Degrades 15–20% on phone audio | △ Consumer-focused |
| Concurrent call volume | ✓ 140K+ production proven | ✗ Shared quota limits apply | △ Enterprise caps; not published |
| Full voice agent stack | ✓ STT + LLM + TTS in one API | ✗ Requires Dialogflow/CCAI | △ Strong TTS only |
What they'll say and how to respond.
“We already built our own agent servicing tool — we don't need another AI vendor.”
→ Deepgram is the STT infrastructure under your tool, not a replacement. You're at 93%. The gap is STT accuracy on telephony. We get you to 97%+ with zero changes to your app.
“We use Google Cloud — we're standardized on their AI stack.”
→ Google STT runs on shared infra. Live cardholder audio under PCI/GLBA cannot route through a shared endpoint. Deepgram deploys inside your GCP private VPC — you stay on Google, we replace the STT layer.
“ElevenLabs has better voice quality for customer-facing interactions.”
→ ElevenLabs leads on TTS — but STT isn't their core product, and compliance is Enterprise-tier only with no on-prem. Deepgram Aura-2 won the 2025 Contact Center Tech Award for FSI environments.
“Timing is bad — we're in the middle of the Discover integration.”
→ The Discover integration is the best possible timing. Inserting Deepgram as a unified STT layer now = one vendor, one compliance posture, one integration across 100M customers — not two parallel efforts at 2x cost.
Single objective: book a meeting with Arjun Dugal's org to scope a 30-day agent assist STT POC.
No pitch. Reference IBM watsonx partnership: 'We're IBM's first voice partner — saw your Chat Concierge work at NVIDIA GTC.'
'Your team got agent servicing to 93% search relevance — we worked with Five9 on a similar challenge and moved WER 30%.'
AE sends intro email through IBM. Reference mutual contact re: voice infrastructure for agent assist roadmap.
Share one-pager: 'What happens to your 93% search relevance when we replace the STT layer.'
Reference his MIT CSAIL talk on NLP/Conversational AI. Building voice infra for FSI contact centers.
Technical insight: 'We co-designed models and hardware to get sub-500ms streaming. Curious how you're approaching this.'
Architecture one-pager: 'STT as infrastructure vs. a feature — how we approached it for FSI.'
Ask for 20-min architecture conversation. Not a demo.
Identify CCW/ICMI 2026 attendance. Pre-connect on LinkedIn: 'Running a roundtable on AI agent assist latency.'
Reference BofA background. '1M+ calls/week — how are you thinking about next-gen agent assist?'
Send ROI one-pager: '45 seconds of AHT reduction at your call volume — here's the math.'
Propose 30-min call. Handle time benchmark on telephony audio like LatAm/Philippines profiles.
Scope and close a 30-day agent assist STT POC. IVR, QA, Discover unification are post-POC expansion.
Personalized per stakeholder. Subject line is the hook. No deck links, no fluff.
Arjun — saw your Chat Concierge work at NVIDIA GTC and your team's disclosure that you moved search relevance from 84% to 93%. We just ran a benchmark on telephony audio with the same noisy, compressed, multi-accent profile as Capital One's contact centers. WER 30% lower than Google STT. On your call volume, that closes the remaining 7-point relevance gap. We're IBM's first voice partner in watsonx Orchestrate — your IBM team can vouch. Worth 20 minutes?
Corey — at the volume your team handles across the US, LatAm, and Philippines, 45 seconds of handle time reduction is roughly $18M annually in fully-loaded labor cost. That's the delta between generic STT and Deepgram's telephony models on your call profile. Your team built the agent assist layer. We replace the transcription engine underneath. Zero changes to your application.
Chris — before your security team sees a POC request from us, I want to put the compliance docs in front of you first. • SOC 2 Type II certified • Deployment inside your AWS/GCP private VPC — zero data egress • PCI-readiness documentation available • HIPAA BAA on file Cardholder audio never touches shared infrastructure. I'd rather you review the architecture before a POC is scoped than after.
Sanjiv — you've talked publicly about digital tools helping dealers open financing conversations earlier. Chat Concierge is live. The natural next step is voice. Auto originations up 21% YoY = more dealership conversations, more noisy floor audio that degrades generic STT by 15-20%. Deepgram's models are trained for high-noise, multi-speaker environments. The compliance path into COF infra is already cleared through Arjun Dugal's team.
| Timeline | SDR | AE | VP / Exec |
|---|---|---|---|
| Week 1–2 — OPEN | LinkedIn connect: Dugal (IBM), Bose (NVIDIA/MIT), Lee (BofA). Research recent COF AI hooks. | IBM watsonx warm intro request. Prepare telephony WER benchmark for Dugal. | No exec touch yet — let SDR warm contacts. |
| Week 3–4 — ENGAGE | Follow-up DMs with specific insight per target. Register for Lee's CCW/ICMI session. | First meeting with Dugal via IBM intro. Parallel technical chat with Bose. | VP Strat Sales peer contact with Dugal. Reference IBM/NVIDIA ecosystem. |
| Week 5–8 — QUALIFY | Nims pre-empt: compliance package proactively. Monitor LinkedIn for new content. | Security review with Nims. Talreja technical deep-dive. Scope POC terms with Dugal. | VP joins Dugal call peer-to-peer. Present ROI at VP level. |
| Week 9–12 — CLOSE POC | Set follow-ups, prep reference customer stories. | POC results to Lee + Yajnik. Close commercial proposal anchored to H1 2027. Begin procurement with LaPrade. | VP + Deepgram exec briefing for Yajnik. Reference Discover + Brex urgency. |