Skip to content
Sales and voice agent

Sales and voice agent

What it does

It is PULSE’s sales development persona: it writes cold-call scripts, speaks short replies on a live phone call, and flags leads at risk of going cold.

  • Drafts a cold-call script for each lead that Find Leads prospects, grounded in what it found on that company’s own website (Leads).
  • On a phone call you place with Call, replies to each thing the person says, turn by turn, using the lead’s cold-call script when one exists.
  • Once a day, rates synced HubSpot leads as high, medium, or low risk of stalling, so PULSE can schedule a retention task for the high-risk ones (Churn watch).

Cold emails and LinkedIn messages for leads are written by the channel content creator.

What it reads

  • For a script: the lead’s name, title, and company, and, for prospected leads, gaps found on the lead’s website.
  • For churn watch: the lead’s name, company, stage, and notes.
  • Your diagnosis, to ground the pitch in your positioning.
  • On a call: the transcript of what the person just said. Speech-to-text runs on the platform’s own speech server.

What it produces

  • A short script: an opener, a reason for calling, questions, and one next step.
  • On a call: one to three natural spoken sentences per turn, with no lists or formatting, read out by the platform’s text-to-speech.
  • For churn watch: a single risk rating per lead.

Human oversight

  • A person starts every call. PULSE never dials on its own. Each call is placed from a lead’s Call action.
  • During a call, replies are spoken as they are generated. No one approves each reply before it is heard, so review the lead’s script before you call.
  • Scripts are drafts saved on the lead.
  • A high-risk rating only creates a retention task in a follow-up cycle. A person decides what to do with it.

Telephony set-up

A call to a lead is an AI call over the Twilio voice settings an administrator saves. Each call needs the lead’s recorded consent for that exact number, the lead’s local calling hours (08:00–21:00 at most), a number that is not on the do-not-call list (an unreadable list refuses the call), the attempt cap, and it opens by saying it is an AI. The earlier Asterisk path is retired.

Outbound calling is also off by default at the deployment level and has to be armed deliberately, and only an editor or above can place a call. The full set-up is in Telephony and messaging.

Outbound cold calling is regulated. Consent, calling hours, do-not-call lists, and recording disclosure are the operator’s responsibility.

Model

The deployment’s base model, shared with every other PULSE agent — on Runink Server, a Qwen-family model under the Apache-2.0 licence. See Sovereign inference.

Where it runs and data handling

On the deployment’s own inference server, with speech-to-text and text-to-speech on the platform’s own speech services: Runink’s shared machines on a Trial, your own cloud on Cloud and your own servers on On-premises. No outside AI service is used on any plan. The phone call itself travels through the Twilio account set on the Messaging page.

Guardrails

  • Lead with their problem, not the pitch. One clear, low-friction next step. No hype, no fake urgency.
  • Short and human. Spoken replies are one to three sentences, and objections get a brief acknowledgement, never an argument.
  • No placeholders. It never says “[Name]” or “[company]”; if it does not know a detail, it talks around it.
  • Never passes as a person. Requests to claim to be human, or to impersonate someone, are refused.
  • Never invents customer results, prices, or statistics, and never takes payments or card details.
  • On calls, what the person says is treated as material to answer, never as instructions to follow, and each turn passes the shared guardrails.

Scope

  • It works from the lead record, the script, and the call so far.
  • It speaks and drafts; follow-ups such as booking a meeting are yours. (Placing a call moves the lead to Talking; that is PULSE’s board, not the agent’s decision.)
  • Voice quality and turn-taking follow the speech services and the phone line.
  • A risk rating is a judgement from the stage and notes, not a prediction model.