AI voice agents for outbound sales
SamCallwell calls your leads with a natural voice, sticks to the script you approve, handles the unexpected with AI, and hands interested buyers to your team before they hang up.
Designed for 1 to 100 concurrent calls
The shift
Outbound sales depends on reps repeating the same opening all day. SamCallwell handles that work, so your team spends its time on leads who want to talk.
Anatomy of a call
Every call moves through the same traceable steps, from the moment a contact list is uploaded to the moment the record is saved.
01
Import a CSV of contacts and choose the script version, caller numbers, retry policy, and concurrency limit. Numbers on your do-not-call list are filtered out before dialing.
02
Each contact becomes a job on the outbound queue. Workers pick up jobs as capacity frees up, and locks make sure no one is called twice.
03
A worker dials through Twilio and tracks the call from dialing to ringing to connected, with every status update logged.
04
The opening line plays from pre-generated audio, so the lead hears a natural voice with no delay.
05
The lead's reply is transcribed to text so the conversation engine can decide what to say next.
06
Branching rules pick the next scripted line. If the lead goes off script, the LLM writes a reply within safety and latency limits, then steers back.
07
When the lead shows interest, the live call transfers to your sales number and the rep gets the context of the conversation.
08
The call record, transcript, events, and recording are stored and appear in the dashboard as soon as the call ends.
Live demo
Scripted lines play instantly. When a lead goes off script, the AI answers in the same voice, then steers back or hands off.
Sarah · NovaTech (example)
Outbound call · 00:00
This call
Scripted
Approved lines from a versioned script. The audio is generated ahead of time, so it plays instantly and sounds the same on every call.
AI reply
When the lead asks something the script doesn't cover, the language model writes a short reply in the same voice and guides the call back on track.
Handoff
Once interest is clear, the call transfers to a real person without the lead having to call back.
Hybrid voice engine
Known script lines are generated once and cached. Only truly new responses are generated during the call, which keeps playback fast and voice costs down.
Average wait before a line plays17%
Voice generation per call10%
Design target for a typical script
Find audio by script line, voice, and text hash.
Return a ready-to-play URL immediately.
Generate once, store the audio, and reuse it next time.
tts_cache/{voice_id}/{script_key}_{text_hash}.mp3Most lines start playing with no generation delay.
Each script line is generated once, not on every call.
The same line sounds the same on every call.
New script versions can be generated ahead of time.
Human handoff
SamCallwell listens for buying signals and transfers the live call to your sales team, with the context they need to pick up the conversation.
What your rep receives
Scale
Calls are queued and spread across workers that scale out horizontally. The system is designed to run 1 to 100 concurrent calls.
Concurrency limit
100 concurrent calls
Set per campaign. Calls beyond the limit wait in the queue.
Add workers to add capacity. Concurrency is set per environment.
Redis locks make sure each contact is dialed only once.
Jobs wait in the queue when capacity is full instead of failing.
Failed jobs retry with backoff, then move to a dead-letter queue for review.
Bring your own AI
Speech, voice, and language models are pluggable. Run them locally for privacy or use hosted APIs for convenience, and switch between them in configuration.
Writes replies when the conversation leaves the script.
Turns what the lead says into text.
Gives your agent a natural, consistent voice.
providers: llm: provider: ollama model: llama3.1 timeout_s: 8 stt: provider: whisper_local model: base tts: provider: elevenlabs voice_id: sales_female_01Change a provider in configuration. No code changes required.
Trust and compliance
SamCallwell is multi-tenant from the ground up. Every record, queue message, cache key, and file is scoped to a single company.
Pick a company to trace its data through every layer.
APIs and UI
company_id from the signed-in user's tokenQueues
company_id on every messageCache
company:{company_id}:… keysDatabases
company_id on every row and documentFile storage
companies/{company_id}/… pathsNumbers are checked against DNC lists before every dial, with an audit trail.
Recording consent is logged, and recording can be turned off where required.
Personal data is encrypted in transit and at rest.
Company Admin, Agent, and Viewer roles, each limited to their own company.
Firebase Authentication with verified tokens on every request.
Traffic is encrypted end to end through a secure tunnel and reverse proxy.
Observability
Metrics, dashboards, traces, and structured logs are part of the design from the start, so you can see what is happening on every call.
active_callstts_cache_hits_totalqueue_depthllm_latency_secondsHover or focus a chart to inspect
Prometheus counters and latencies for calls, queues, and providers.
Grafana dashboards for day-to-day KPIs.
OpenTelemetry spans from webhook to worker to provider.
Structured JSON logs you can follow by call ID.
Architecture
A container-first set of focused services. Each one does a single job and can be scaled or replaced on its own.
Edge
Core services
Messaging and state
Data
Voice and AI providers
Operations
Hover or focus a service to see what it connects to.
Roadmap
We are building toward a focused MVP first, then expanding into real-time conversation and richer agent tooling.
Get started
Tell us about your team and your outbound goals, and we will set up a walkthrough.