Multi-Channel Customer Service

Solo founders, small SaaS teams, and local businesses all hit the same support wall: customers want fast replies, your team can't be on five channels at 11 PM. Flowly can handle the bulk — qualifying questions, answering FAQs, looking up order status — and escalate cleanly when a human is required.

What it does

Single agent listens on Telegram, WhatsApp, web chat, email
Inbound messages get PII-redacted at the gateway boundary (pre_gateway_dispatch hook)
Spam and bot traffic dropped before it costs LLM tokens
Common questions answered from a knowledge base (FAQ skill + memory)
Order status / account lookups via tools to your backend
Handoff to human via /escalate <reason> — agent stops, ticket fires to your CRM
Every conversation logged for compliance + training data

What you'll need

Connections set up: WhatsApp Business API, Telegram bot, web chat embed
A PII redactor plugin (covered in the PII redaction use case)
A spam-filter plugin with pre_gateway_dispatch hook
A handoff-router plugin that fires webhooks to your CRM (Zendesk, Intercom, custom)
Knowledge base content as a skill (markdown FAQs, product docs, policy excerpts)

Setup

1. Connect inbound channels

Each channel from the dashboard's Connections tab. WhatsApp requires the Business API (separate Meta verification — slowest part of the setup).

2. Build the safety pipeline

Two plugins fire before the agent sees a message:

python

# spam-filter/__init__.py — blocks obvious junk before LLM tokens get spent
from flowly.agent.hooks import SkipAction
import re

PATTERNS = [re.compile(p, re.I) for p in [
    r"crypto.*airdrop", r"bedava bitcoin", r"click here to claim",
    r"viagra", r"forex.*signal",
]]

def register(ctx):
    def filter_inbound(hook_ctx):
        text = (hook_ctx.event.content or "")
        for p in PATTERNS:
            if p.search(text):
                return SkipAction(reason=f"spam: {p.pattern}")
        return None
    ctx.register_hook("pre_gateway_dispatch", filter_inbound)

The PII redactor is identical in shape — see its dedicated use case.

3. Build a knowledge skill

Inside ~/.flowly/skills/support-faq/SKILL.md put your FAQs, return policy, hours, etc. Frontmatter description should be specific: "Use this skill when the customer asks about returns, shipping times, business hours, or account billing."

4. Configure the handoff plugin

Send to Flowly

Create a plugin called "handoff-router" with one slash command: /escalate <reason> When fired, it should: 1. POST to my Zendesk webhook with the recent conversation transcript and reason 2. Reply: "Thanks — a human will follow up within X hours. Reference: <ticket id>" 3. Add a note to memory tagged "escalated:<ticket_id>" with the conversation excerpt Use ZENDESK_WEBHOOK_URL from env.

5. Train the system prompt

In your dashboard's Agent settings, anchor the assistant's persona:

Send to Flowly

You are the support assistant for [Company]. Your tone is friendly but brief — answer the actual question, no preamble. ALWAYS check the support-faq skill before guessing. If the answer isn't there and isn't obvious, use /escalate. NEVER share refund amounts above $50, account credentials, or internal policy without explicit confirmation from a human. Use /escalate. Hours: 9–6 weekdays. Outside hours, lead with: "I'll do my best now; the team is available 9–6 if I'm wrong."

Tips

Audit before going live. Run the agent against ~50 historical support tickets and compare its replies to what the human said. Patch gaps before customers see them.
PII redaction is non-negotiable. Even for internal logs. Phone numbers and IDs in logs become a compliance problem when (not if) you get a data request.
Set token budgets. Customer questions are cheap individually but add up. Cap tokens per session and tools-per-turn to avoid runaway scenarios.
Watch the escalation rate. If /escalate fires more than 30%, the FAQ skill is missing content. If under 5%, the agent is probably overconfident — review a sample weekly.
Don't promise what you can't deliver. "I'll fix it now" then no human responds for 2 days is worse than "I'll get a human on this in 4 hours" delivered on. Match the agent's commitments to your reality.