Voice AI Agents for Sales: The New Outbound Motion

What actually changed

Three things landed in the last twelve months: streaming speech-to-text at near-human accuracy, LLMs that can reason mid-turn without a full round-trip, and low-latency neural TTS that doesn't sound like a GPS. The composite result is a conversation that feels responsive rather than transactional.

The threshold that matters is round-trip latency: humans expect a response inside ~600ms. Any voice agent above that will get talked over and lose the call. Modern pipelines using services like Deepgram or the built-in speech APIs from OpenAI and Google routinely hit 300–500ms.

The voice agent stack

A production voice agent has five moving parts: telephony (Twilio, Vonage, Telnyx), a streaming ASR, an LLM policy layer, TTS, and a CRM/tool bridge. The choke point is usually the LLM: if you're waiting for a full completion, the pause is audible.

The workaround is streaming — begin synthesizing speech from the first tokens the LLM produces. It requires careful sentence-boundary logic but it's how modern agents feel alive rather than scripted.

• Telephony carrier with SIP trunking and call recording
• Streaming ASR (Deepgram, Whisper realtime, Google STT)
• LLM planner with strict tool schemas for CRM writes
• Neural TTS (ElevenLabs, Cartesia, Google WaveNet)
• Orchestrator that manages barge-in, silence, and handoff

Outbound vs inbound: different problems

Outbound voice AI is a compliance minefield. Between TCPA in the US, GDPR in the EU, and a rising wall of state-level laws, cold outbound calling with AI voices carries real legal exposure. Most teams we advise use voice AI outbound only for warm follow-up (opted-in leads, existing customer motions) and let humans handle true cold prospecting.

Inbound is the sweet spot. Missed calls are pure revenue leakage, and a voice agent that qualifies, books, and confirms 24/7 recovers 15–40% of that leakage in the first quarter — with no compliance risk since the caller initiated contact.

The SDR economics

A human SDR runs $75–120k fully loaded and makes 40–80 quality dials per day. A voice agent handles thousands of parallel conversations at variable cost of roughly $0.06–0.15 per minute. The naive math suggests replacing everyone — but it's the wrong math.

The right frame is coverage. Voice AI lets you follow up on leads within 60 seconds instead of six hours (Harvard Business Review's classic study found a 7x contact-rate difference at that threshold), qualify off-hours traffic, and free human SDRs for the top 20% of accounts where personal relationship still wins deals.

Deploying without embarrassing yourself

Two rules keep launches from becoming Twitter threads: always disclose the caller is AI at the top of the call, and always allow a one-word handoff to a human. Both are trivially easy to build and they eliminate 95% of the reputational risk.

Beyond that, treat the voice agent like a new hire. Give it a persona, a scorecard, and a manager (the ops team monitoring transcripts). Iterate weekly on the top 10 failed calls. Six weeks in, you'll have something that outperforms the median SDR on discovery consistency.

A realistic rollout sequence

Week one and two are discovery and instrumentation. You are not building yet — you are shadowing the people who do the work today, recording how they actually decide, and wiring the measurement so you can prove change later. Skipping this is the single most common reason voice ai agent projects stall: without a baseline, every subsequent conversation about impact becomes an argument about anecdotes.

Weeks three and four produce a suggest-only version. The system runs on live traffic, produces its recommended handling, and a human accepts, edits or rejects it. Every one of those judgements becomes a labelled example. You are simultaneously delivering value — the operator's draft is written for them — and building the evaluation set you will need to justify autonomy.

Weeks five and six graduate the highest-confidence, lowest-blast-radius slice to autonomous handling behind a kill switch, with the remainder still routed for review. From there, expansion is a data-driven negotiation: each category earns autonomy when its accuracy on held-out cases clears the threshold you agreed in week one. Most teams reach 40–70% autonomous handling within a quarter, and the remaining cases stay human because they should.

• Weeks 1–2: shadow the process, instrument the baseline
• Weeks 3–4: suggest-only on live traffic, capture every correction
• Weeks 5–6: autonomous on the safest slice, behind a kill switch
• Quarter 2: expand category by category on measured accuracy

Measurement that survives a board review

Activity metrics — messages handled, tasks executed, hours "saved" — are the metrics of a project that is about to be cancelled. They rise regardless of whether the work was any good. Replace them with outcome metrics measured against a holdout: a slice of traffic deliberately handled the old way so you always have a live control group rather than a historical one.

The four numbers we hold ourselves to on voice ai agent engagements are resolution or completion rate without human touch, quality as judged by a blind human review of a weekly sample, cycle time from trigger to resolved, and cost per case fully loaded including inference and engineering amortisation. Report all four together. Any one of them in isolation can be gamed, and the combination cannot.

Publish the numbers weekly to a channel that includes sceptics. Programmes die in silence, not in criticism — and the fastest way to earn the budget for phase two is a four-week chart that a CFO can read without a translator. Our AI agent engineering team ships this dashboard as a deliverable in week one, before the first line of production logic is written.

• Autonomous resolution rate against a live holdout
• Blind quality score on a weekly random sample
• Cycle time, median and 90th percentile
• Fully loaded cost per case, including inference
• Escalation reasons, grouped and trended

Frequently asked questions