Inbound and outbound assistants: campaigns and messaging

Inbound and outbound are different problems

An inbound assistant waits for someone who wants something. An outbound assistant starts the conversation, which changes almost everything.

Common outbound uses: appointment and delivery reminders, payment and renewal notices, onboarding nudges, surveys, alerts, re-engagement, and campaigns or mailings across email, SMS, messaging apps, push notifications and voice.

Consent and compliance principles

The details vary, but the principles are consistent:

• Get consent and keep a record of when, how and for what. Don't assume it.
• Honor opt-outs instantly and everywhere, and confirm them. Keep a suppression list that every sender checks.
• Identify yourself clearly: who is contacting, and why.
• Respect timing: local time zones, quiet hours and frequency caps.
• Disclose automation where required or expected, and never present a bot as a person.
• Stay within the channel's rules. For example, business-initiated messaging on some platforms uses pre-approved templates, and bulk text messaging may need sender registration. Check each provider's current policy.
• Authenticate email (for example with SPF, DKIM and DMARC), include a working unsubscribe, and keep lists clean.
• Don't reveal personal information before verifying you have the right person.

Examples of laws and rules that may apply include the US Telephone Consumer Protection Act and CAN-SPAM, EU data-protection and e-privacy rules, Brazil's LGPD and Canada's CASL, plus industry-specific rules. In 2024 the US FCC confirmed that AI-generated voices count as "artificial" voices under the TCPA, which is covered in the case study.

Designing a campaign

A campaign is a loop. The reply and the opt-out are part of the design, not edge cases.

1. Audience. Segment by need and consent. Exclude anyone who opted out, was contacted recently or is in an active support case.
2. Trigger. Event-based (an order shipped, a bill is due) usually outperforms and annoys less than blast schedules.
3. Message. Identify yourself, give the reason, offer value, and ask for one clear action. Make opting out easy.
4. Channel. Pick the one the customer prefers, and keep sensitive content out of insecure channels.
5. Throttle. Send in batches, watch complaints and opt-outs, and stop automatically if they spike.
6. Replies. Decide in advance how each reply is handled. Handle keywords (STOP, HELP) deterministically. Send everything else to the assistant, with the campaign context attached.
7. Measure outcomes and harms, not only opens and clicks.

Notice that the reply moves straight into the inbound assistant, with the booking already known.

Mailings and AI-written content

AI can personalize at scale, which raises the stakes for control:

• Approved building blocks. The model fills slots in approved templates and claims, rather than inventing offers, prices or promises.
• Guardrails on content (module 10): no invented discounts, no unapproved claims, brand and tone rules, required legal text and unsubscribe links.
• Human approval for campaigns, with samples reviewed across segments and languages.
• Preview and test sends, including rendering in common email clients.
• Logging: keep the prompt version, the data used and the final text sent.

Voice outreach specifics

• Identify yourself immediately, state the purpose, and make it easy to stop.
• Wrong person. Plan the script when someone else answers or the person is unavailable.
• Verify before disclosing anything personal.
• Voicemail and answering machines. Decide whether to leave a message and what it may contain.
• Call windows and retry cadence respect local time and limits.
• Recording and disclosure follow the rules of each region. See module 19 and module 20 for the speech side.

Architecture

Consent is a shared service that every sender checks. Replies flow back into the same conversation service.

A pre-send check is a small, high-value guardrail:

from datetime import datetime
from zoneinfo import ZoneInfo

def can_send(contact: dict, message: dict, now_utc: datetime, recent_count: int) -> tuple[bool, str]:
    """Decide whether a message may go out right now."""
    if not contact.get("consent", {}).get(message["channel"]):
        return False, "no consent for this channel"
    if contact.get("opted_out"):
        return False, "opted out"
    if not message.get("template_approved"):
        return False, "template not approved"
    local = now_utc.astimezone(ZoneInfo(contact["timezone"]))
    if not (9 <= local.hour < 20):                  # example quiet-hours window; set from your rules
        return False, "outside allowed hours"
    if recent_count >= message.get("max_per_week", 2):
        return False, "frequency cap reached"
    return True, "ok"

Log every decision, including the reason a message was held. That log is part of your compliance record.

Measure outcomes and harms

Open and click rates are increasingly unreliable signals, so weight replies, outcomes and complaints more heavily. And always attach guardrail metrics to experiments: a campaign that raises conversions while doubling opt-outs is not a win.