phone_outreach_automation v1.0.0

name: phone_outreach_automation description: Plan and run agent-placed phone calls to businesses, vendor-neutral: write call scripts, run the pre-call compliance gate (AI-identity disclosure, TCPA consent, call-recording law), place or schedule calls through any telephony provider, and log dispositions, transcripts, and outcomes. Triggers when an agent needs to call real businesses, build a call script, check whether an AI call is legal, set up recording consent, or keep a call-outcome ledger.

Phone Outreach Automation

Purpose

Agents place real phone calls: sales, scheduling, vendor outreach, follow-ups. This skill covers the full workflow without tying you to any one telephony API. Compliance first, script craft second, logging last. An agent that calls businesses needs all three or it is a liability with a phone line.

Workflow

1. Define the call

Write down, before touching anything else:

• The objective: one sentence. A call succeeds by earning the next step, not by closing on the spot.
• Who you are calling and what you know about them (name, role, business, why now).
• The single call-to-action: one concrete ask, e.g. a 15-minute meeting with two offered times and a stated agenda.

2. Run the pre-call compliance gate

Every call passes this gate. If it fails, the call does not happen.

• Consent. AI-generated voices count as "artificial voice" under the TCPA (FCC declaratory ruling, Feb 2024). Telemarketing or prerecorded-voice calls to consumers need prior express consent, and the FCC proposed (Aug 2024) that consent cover AI content specifically. B2B calls are not exempt from all rules; verify the call type against the checklist in references/compliance-guide.md before dialing.
• AI-identity disclosure. Say up front that the caller is an AI voice assistant working on behalf of the named business. California has required a spoken AI disclosure on covered calls since Jan 2025; the federal disclosure rule is still a proposal, but disclosing anyway is the conservative play and the trust-building one.
• Recording consent. Federal law is one-party consent; ten-plus states require all-party consent (CA, CT, FL, IL, MD, MA, MT, NH, PA, WA, plus edge cases). For any multi-state or unknown-jurisdiction call, announce recording at the start and give the person a chance to hang up. Full state notes in references/compliance-guide.md.
• Opt-out. Telemarketing calls need an interactive opt-out during the call (press a key to be removed). Honor "STOP" / "do not call" immediately: end the call, suppress the number.
• DNC scrub. Check the number against internal and applicable do-not-call lists before dialing.

Never bluff through this gate. When the jurisdiction is unclear, pick the stricter rule. When a question needs a lawyer, say so and stop.

3. Write the script

Six parts, in order. Keep the opener and reason under 20 seconds combined. Full worked template in references/script-template.md.

1. Opener. Name, business, one line of who you are. "Hi [name], this is [agent name], an AI assistant calling for [business]. Do you have two minutes?"
2. Reason for calling. Why this person, why now. Specific beats generic.
3. Discovery. Two to three open questions about their process and pain. The caller should listen more than talk; Gong's call research puts the best talk-to-listen ratio around 43:57.
4. Pitch. Two or three sentences tying what you heard to what you offer. No feature lists.
5. Objection handling. Objections are interest, not rejection. Acknowledge, reframe with a brief example, reinforce value, never argue. Pre-write answers for the five objections you will hear most.
6. Call to action. One specific ask with two concrete times. If no answer, leave a 20-second voicemail: name, business, one-line reason, callback number.

Branching beats a wall of text. For each section, write the "if they say X" branches the agent should actually expect.

4. Place the call

Any telephony provider works: voice-API platforms, an agent phone-call tool, or a human-assisted dialer. The skill does not care which. Required capabilities:

• a script or prompt the voice agent follows,
• automatic playback of the disclosure and recording notice at call start,
• keypress or voice opt-out detection during the call,
• call recording with consent handling, and a transcript afterward.

If the provider cannot do one of these, that gap is a compliance gap. Say so plainly before proceeding.

5. Log the outcome

Every call gets logged the same day, no exceptions. Use bin/call_log.py or match its schema. Minimum fields: timestamp, number called, business, person reached, disposition code, AI-disclosure confirmed (yes/no), recording consent (announced/confirmed/not recorded), opt-out requested (yes/no), notes, transcript path.

Standard disposition codes (full list in references/dispositions.md): no_answer, left_voicemail, interested, not_interested, wrong_number, callback_requested, escalation_required, disconnected, do_not_contact.

A do_not_contact disposition suppresses the number from all future dialing immediately. Dispositions drive retry logic and reporting, so tag honestly; a mis-tagged call corrupts every report downstream.

Output Contract

A completed call run produces:

• the script used (with disclosure and opt-out lines present),
• a call log (JSONL or equivalent) with one entry per attempt, dispositions applied,
• transcripts or recordings where consent allowed, stored with the log entries,
• a short summary: attempts, connects, positive outcomes, opt-outs, DNC additions.

Operating Rules

1. Compliance gate before every call. No exceptions, no "this one is quick."
2. Always disclose the AI identity at the start of the call. This is not optional branding, it is the rule that keeps the operation legal and trusted.
3. Never fake a human. No claiming to be a person, no hedging the disclosure.
4. One-party consent is the floor, not the ceiling: announce recording whenever jurisdictions are mixed or unknown.
5. Log every attempt the same day it happens. A call that is not logged did not happen.
6. do_not_contact means done. No retries, no "one more angle," no exceptions.
7. This skill teaches the workflow, not a vendor. Do not hard-code provider-specific APIs into scripts or runbooks; name the capability, not the product.
8. If you cannot verify a legal claim (a number, a date, a state list), cut it or label it as needing counsel. Unverifiable claims never ship as facts.