The AI voice agent that lives in your Slack.

An AI voice agent is software that conducts spoken phone conversations using a stack of speech recognition (STT), large language model reasoning, and text-to-speech (TTS). Unlike old-school IVR (interactive voice response), modern AI voice agents handle open-ended natural language, can be interrupted mid-sentence, and produce voice output that's increasingly indistinguishable from a human speaker. The 2026 reference stack is Whisper for STT, a frontier LLM (GPT-5 or Claude) for reasoning, and a voice-clone provider like Cartesia or ElevenLabs for TTS, with a SIP carrier (Telnyx, Twilio) for the actual phone call.

Inbound flow: form submit on your site → sub-second Slack alert → Mavrick's voice agent dials the lead via Telnyx within 60 seconds → LiveKit room hosts the call → Whisper streams transcription → GPT-5 runs your approved discovery script → Cartesia 'Brandon' voice speaks responses → call ends with meeting booked or disposition logged → structured handoff card posted to your Slack sales channel with transcript, qualification notes, and CRM-record deep link. Outbound flow: similar except the agent dials from a target list instead of from form submits.

Carrier: Telnyx (~$1.50/mo per DID, $0.005/min US outbound — ~70% cheaper than Twilio at scale). Voice rooms + agent runtime: LiveKit Cloud. Speech-to-text: OpenAI Whisper (streaming, ~200ms latency). Reasoning: GPT-5 with your workspace-specific discovery script in the system prompt. Text-to-speech: Cartesia (default voice 'Brandon', voice_id 5cad89c9-d88a-4832-89fb-55f2f16d13d3, A/B tested winner on lead-completion rate). Compliance: per-state quiet-hours table seeded for 50 states + DC + 5 territories; National DNC + workspace DNC + consent verification fail-closed before any dial.

Cartesia is one of the leading voice-clone providers as of 2026. The default 'Brandon' voice is a neutral US-male tone trained on professional voice talent; on blind A/B against four alternatives it won on lead-completion rate. Most ears can't distinguish Cartesia output from a human speaker in a brief sales call. Custom voice clones (your own founder's voice, an existing sales rep, your brand's spokesperson) are available on Pilot+ tiers — you record 3-5 minutes of clean audio, Cartesia trains the clone, you A/B test it against the default before rollout.

Mavrick's voice agent identifies as AI when asked. The 'always pretend to be human' anti-pattern is explicitly banned by Mavrick's Constitution at /constitution. If the lead pushes back on AI, Mavrick offers to schedule a human callback, sends an SMS with a calendar link to the assigned rep, and posts the disposition to the Slack channel so your team can follow up personally. Honesty about being AI is a feature, not a bug — modern buyers respond better to transparent AI than to deception.

Cleared-hot is a one-time script approval, not a per-call approval. You approve the discovery script, the qualification criteria, the disposition logic, and the handoff rules once — typically during workspace setup. From then on, the voice agent executes consistently against the approved playbook on every call. The approval gate kicks in only when the agent wants to mutate something downstream — book a meeting on a specific rep's calendar, update a CRM stage, send a follow-up SMS. Those are one-click confirmations in Slack. The voice call itself runs against your pre-approved script without per-call approval friction.

Both. Outbound: you provide a target list (HubSpot view, Salesforce report, CSV upload, or a Mavrick-generated list from Apollo/Hunter integrations), set the dialing window, and Mavrick works the list — dialing, qualifying, dropping voicemails per a recorded template if no answer, respecting state quiet-hours, posting disposition per call. Inbound: form submit triggers an immediate dial within 60 seconds (see /for/speed-to-lead for the speed-to-lead-specific deep dive). Both use the same voice stack with workspace-configurable scripts per direction.

Mavrick blocks calls against four gates fail-closed before any dial: (1) National DNC list lookup, (2) state-specific DNC list for the lead's normalized state, (3) workspace-specific DNC list (your manually-suppressed numbers), (4) consent verification (opt-in record check against your CRM or webhook). On any gate miss, the dial is blocked, a forensic record is written, and the disposition surfaces in Slack so you can review the suppressed call. Quiet-hours: 50 states + DC + 5 territories seeded with local quiet-hours windows; calls outside the lead's local 8am-9pm window are deferred. You remain the controller of record under TCPA / state telemarketing law — review your own list legality.

Those are voice-AI platforms — you build the agent on top. You wire the LLM, write the script, integrate with your CRM, build your own Slack handoff layer, manage your own DNC compliance, set up your own concurrency caps. Mavrick is the AI coworker — voice is one capability, but it ships pre-wired with Slack-native routing, CRM handoff, compliance gates, state quiet-hours, cleared-hot approval architecture, and per-workspace concurrency caps. If you have an engineering team and a year of runway, building on a voice-AI platform gives you maximum control. If you're a marketing or sales team that wants the voice agent operating tomorrow, Mavrick is the assembled product.

Recruit (free): 10 missions to evaluate — roughly 5-8 outbound dials or ~15-20 inbound qualifications, including the per-minute Telnyx pass-through. Pilot ($50/mo): 20,000 credits ≈ 60-100 missions/mo. Enterprise: custom — dedicated DID provisioning, raised concurrency caps, 99.9% SLA, custom voice clones, signed DPA. See /pricing for the full breakdown. No per-seat fees on any tier. Carrier costs (Telnyx) are passed through at our rate, no markup.