Voice agent development for US teams that need calls to actually convert.

Honest answer: voice is the workload where same-zone on-call matters most, because a production incident means a customer is on the phone right now. Our Mohali team runs IST, which gives a native two-to-three-hour window with US Eastern late afternoon and a thinner window with US Pacific. For US voice deployments we run a dedicated US-hours pod out of our Frisco, TX office and a tech-lead-on-call rotation covering 9am to 6pm Central — not a junior support shift, the same senior engineers building your voice agent. Twice-weekly demos run in US business hours playing back real call traces. If your engagement requires 24/7 same-zone synchronous coverage on launch night, we will say so on the first call so you can pick a US-only voice consultancy instead.

Sub-1 second p95 turn latency end-to-end (user-stop-speaking to agent-start-speaking) is our standard target and the number we hit on the production insurance reference build. The breakdown: 200-300ms streaming STT first-token on Deepgram Nova-3, 300-500ms LLM first-token on Claude Sonnet or GPT-4o (lower with GPT-4o-mini or a fine-tuned 8B model), 150-200ms TTS first-byte on ElevenLabs Turbo or Cartesia. We measure latency per turn on every production call and ship it to your observability stack — Datadog, Honeycomb, or your SIEM. If your use case demands sub-500ms (telehealth, sales objection handling), we use Realtime APIs (OpenAI Realtime or Gemini Live) and tell you up front what tradeoffs come with that path.

Yes. Engagement controls are HIPAA-aligned. We sign BAAs before any PHI is shared, we pin LLM inference to a US region (us-east-1 or us-west-2) on Anthropic's HIPAA-eligible Claude or Azure OpenAI's HIPAA-eligible GPT-4o, Twilio is on its BAA programme, and Deepgram and ElevenLabs both offer HIPAA-eligible tiers. For clients with strict no-third-party-API requirements, we self-host Whisper for STT, an open TTS model for voice generation, and Llama 3 on vLLM for dialog management — the entire voice pipeline inside your VPC with zero inference data leaving your account. PHI redaction patterns cover SSN, MRN, insurance member ID, DOB, and the long tail of clinical identifiers; audit logs export to your SIEM with turn-level granularity.

Real-time, during the call. The dialog manager calls typed tools mid-conversation that hit Salesforce REST or HubSpot CRM API — opportunity creation, contact update, lead disposition, call-note write, calendar event create. We have shipped this integration on enough deployments that it is a configuration, not a discovery phase. For Salesforce specifically: OAuth 2.0 connected app with named credentials, JSON-schema-validated payloads, idempotency keys on every write so a network retry never creates duplicate records. For HubSpot: private app with scoped permissions, custom-object support, and association rules between Contact / Deal / Engagement. If your CRM is Zendesk, Pipedrive, or a custom system, the same pattern applies — we wire a typed tool, validate the payload, and write during the call rather than batching it.

TCPA, state two-party-consent recording rules, and CCPA disclosure handling are configured at the dialog-manager level. For TCPA: the agent reads the consent disclosure as the first turn, the agent will not place calls outside permitted hours per the called number's time zone, and the do-not-call suppression list is checked before dial. For state recording rules: the agent reads the recording disclosure in the opening turn for any state requiring two-party consent (California, Florida, Illinois, etc.), and recording is gated on opt-in confirmation. For CCPA: data-subject-request handling is wired into the agent so a caller can invoke their right to delete or access during the call. Compliance review with your legal team is built into week one — we will not ship an outbound campaign without it.

Yes — voice takeover audits are routine. Step one is reading the call recordings and traces, the latency telemetry, the refusal-rate and hand-off-success numbers, and the cost-per-call data. Step two is shipping the smallest valuable change to prove we understand the system — usually fixing barge-in handling, the escalation threshold, or the CRM write-back race condition that the previous vendor skipped. Step three is the longer-term rebuild plan if one is needed. Most voice takeovers we see did not need a full rewrite; they needed proper turn-detection, a confidence-based refusal layer, and a senior engineer on the build. We will be honest on the first call about which category your project lands in.

Most v1 voice agent engagements at Aiinfox land between $30,000 and $140,000 fixed-price for a focused build — an outbound campaign, an inbound deflection flow, or a HIPAA-aligned healthcare voice agent. Larger multi-quarter engagements with custom fine-tuning, bespoke evals, multi-language voices, and integration into a regulated platform typically reach $180,000 to $320,000. The cost difference versus a Bay Area or NYC voice consultancy lands roughly 30 to 50 percent lower on senior rates — but the headline is the engineer on your kickoff call writes your dialog manager, your CRM integration, and your eval suite through launch. No swap-out to a junior pool mid-engagement.

Outbound insurance voice (sub-1-second p95 agent saving 1,400 staff-hours per month and lifting renewal conversion by 28%), telco support (68% L1 deflection sustained over nine months on a 2M-subscriber bot — same dialog manager runs the voice version), EdTech adaptive voice (47% completion lift on Mockinto, the reference we ship ourselves), and HIPAA-aligned medical inquiry voice with citation accuracy as a release gate. Reference calls available under NDA. 50+ production systems shipped across 12 verticals — see the documented case studies for the engineering and business outcomes we can show publicly.