OutgoingVoice
Preview

Operator guide

Voice stack

How agents run under policy

Campaign fields are the knobs; this page is the manual. It explains process, agent depth, prompts, guardrails, and tools — not a vendor or model catalog. For the public journey, see How it works · Compliance posture · Campaigns.

Call process

  1. Step 1

    Pre-dial gate

    Campaign enabled? Destination class allowed? Lawful-basis tags and consent artifact present? Not suppressed? Under frequency caps? If any check fails, nothing is placed — and the attempt is audited.

  2. Step 2

    Connect

    Telephony places the outbound leg with your caller ID. Ring, answer, busy, and no-answer are tracked. Media only flows after a live path is allowed by policy.

  3. Step 3

    Agent loop

    Speech in → understanding → language model with tools → speech out. Forced open (identity + AI disclosure + purpose), then dialogue until end, opt-out, or escalate.

  4. Step 4

    Close & review

    Disposition, transcript turns, optional recording, metrics (answer time, first AI audio, usage). Opt-outs write to suppressions; escalations become callback tickets.

Agent levels

Not every campaign needs the same “brain.” We run several agent tiers so you can trade speed, cost, and depth of reasoning without changing your campaign policy or compliance gate. Tier choice is an ops/product decision per use case — appointment confirm often wants fast and consistent; complex reschedule or edge-case triage can use a deeper tier.

Fast lane

Lightweight agent

  • Optimized for short, high-volume confirms and simple yes/no paths.
  • Lower latency to first spoken response; tighter reply length.
  • Ideal when the script is stable and tools are few (confirm, opt-out, end, escalate).

Balanced

Standard agent

  • Default for most appointment and reminder programs.
  • Enough context window to hold the open, callee reply, and multi-turn clarify.
  • Reliable tool use for opt-out, escalate-to-callback, and domain intents you allow.

Deep

Reasoning-heavy agent

  • More context and stronger multi-step reasoning when the conversation branches.
  • Better at following long guardrail lists and multi-document RAG snippets.
  • Higher latency and cost — reserve for edge cases or lower-volume, high-stakes queues.

Under the hood, language models differ in context length, reasoning ability, and speed. We select and swap models operationally; this UI does not surface vendor or model IDs. What you control is campaign policy, prompts, tools, and tier intent.

Speech understanding

Inbound audio is converted to text (partials when streaming is available, finals for the transcript). That text is what the agent reasons over and what you review later — not a black-box summary only.

  • Turn-taking — wait for end-of-utterance or barge-in cancel of TTS.
  • Noise & machines — voicemail / AMD paths can short-circuit to leave-message or hangup policy.
  • Audit — finals land on the call record with timing for QA.

Speech generation

Outbound audio is synthetic brand voice only — no clones of real people, candidates, or celebrities. Tiers differ in naturalness, latency, and how much you can customize delivery.

  • Brand voices — allowlisted voice profiles per org or campaign.
  • Customization — higher tiers: speaking rate, stability, style hints, locale.
  • Barge-in — cancel in-flight speech when the callee interrupts.
  • Forced open — identity + AI disclosure + purpose spoken before free dialogue.

Prompts — how they change agent behavior

Prompts are not one blob of marketing copy. They are layered so compliance and product rules stay stable while campaign language stays editable. You edit the knobs on Campaigns; this section is why those knobs matter on the line.

1. System / policy layer (platform)

Always-on instructions: you are an AI; never claim to be human; never collect full payment card data; honor opt-out immediately; escalate when asked for a person; stay inside allowed intents; do not invent appointments that are not in tools/context.

You are an outbound AI voice agent for {org}.
You must disclose that you are AI in the opening.
Never pretend to be a specific real person.
If the callee says stop calling / remove me, call tool opt_out then end politely.
If they ask for a human, call tool escalate_human.
Only discuss intents in: {intents_allowed}.
Refuse: {intents_blocked}.

2. Campaign open (forced)

Identity, AI disclosure, and purpose — configurable sample language per campaign, required flags in config. Example appointment open:

Hello, this is an AI assistant calling on behalf of {clinic_name}
to confirm your appointment on {date} at {time}.
You can confirm, ask to reschedule, opt out of these calls, or ask for a person.

3. Turn instructions (runtime)

Short reminders each turn or on tool results: keep answers brief for voice; one question at a time; confirm back critical facts; if uncertain, escalate rather than guess.

Callee just said: "{user_transcript}"
Respond in 1–2 short spoken sentences.
If they confirmed, acknowledge and offer one chance for questions, then end_call.
If unclear, ask a single clarifying question.

Guardrails

Guardrails are product and compliance controls that bound the agent — not optional prompt suggestions. Several layers stack:

Pre-dial (hard)

  • Campaign enabled / disabled
  • Destination class allowlist (wireless, landline, …)
  • Lawful-basis type + consent artifact requirement
  • Global / campaign suppression
  • Frequency caps (e.g. max attempts per 30 days)

In-call (hard tools)

  • opt_out → suppression write + polite close
  • escalate_human → callback ticket (MVP), not silent hangup
  • end_call → clean termination
  • Intent firewall: allowed vs blocked intents on the campaign

Disclosure & identity

  • AI disclosure required in the open
  • Org / brand identity required
  • Optional paid-for-by style disclaimers where campaigns need them
  • No voice clone of real individuals

Content & safety

  • No inventing medical/legal outcomes beyond scripted scope
  • No aggressive collections or upsell if intents_blocked says so
  • Escalate on abuse, distress, or “get me a person”
  • Keep answers short enough for spoken delivery

Tools, RAG, and grounded answers

The agent should not freestyle facts that belong in your systems. Tools and retrieval (RAG) ground the call in your data under policy.

Core tools (MVP)

  • opt_out — record do-not-call / suppression reason; end the conversation cleanly.
  • escalate_human — open a callback ticket with reason (callee request, confusion, policy). Staff owns the human leg.
  • end_call — successful or neutral close after confirm / message complete.

Later domain tools (examples): lookup_appointment, propose_reschedule_slots, send_sms_confirm — always behind campaign allowlists.

RAG-style grounding

Before or during the call, the orchestrator can inject short, retrieved snippets so the model answers from documents — not memory. Typical sources:

  • Campaign FAQ / “what we can say” snippets
  • Location hours, parking, prep instructions
  • Program-specific disclosure language
  • Appointment row for this callee (time, provider, location) from your CRM/export
[Retrieved context — do not invent beyond this]
Appointment: Tue Aug 12, 2026 · 2:30 PM
Location: North Clinic, 100 Main St
Prep: arrive 10 minutes early; bring insurance card
FAQ: Reschedule only via callback if within 24 hours.

Telephony vs intelligence

Phone network work (numbers, dial, hangup, media path, call events) is separate from conversation intelligence (understand speech, decide what to say, synthesize voice, run tools). That split is intentional:

  • Compliance and audit live in our control plane and orchestrator — not inside a third-party “agent builder” UI.
  • Speech and language components can be upgraded for latency or quality without rewriting campaign policy.
  • Call review always sees our transcript turns, dispositions, and metrics shape.

What “good” looks like on a live call

  1. Gate allowed the dial; lawful basis snapshot stored on the attempt.
  2. Open includes identity + AI disclosure + purpose before free-form chat.
  3. Callee can confirm, reschedule path, opt out, or get a human callback — without arguing with a robot persona.
  4. Transcript and outcome show up on Call review; opt-out hits suppressions; escalate hits the ticket board.
  5. Nothing on the line implied a clone of a real person or an unconsented wireless spray.

Not legal advice. Calling entities remain responsible for lawful basis, lists, and disclosures. Outgoing Voice enforces the policy you tag on campaigns.