irisbites

M2 · Free playbook

Install an AI Lead Responder
that replies in 30 seconds.

Small-business lead-to-conversion drops 80% when first response takes more than five minutes. This is the seven-day playbook we use on paid installs to eliminate that lag — across chat, email, DM, and SMS — written so you can run it yourself, free.

TL;DR

What it actually takes.

An AI lead responder is four things wired together: a chat-and-DM tool, a routing layer for email and SMS, a knowledge base, and a brain that drafts replies in your voice. The stack costs about $60 to $80/mo for most SMBs and takes roughly four hours of focused engineer work spread across seven days.

The hard part is not the platform. The hard part is the reply — what the AI says when a stranger asks a price question over Instagram on a Sunday night, what it does when the lead is angry, what it does when someone's testing whether you're real. Most failed lead responders are reply-quality failures, not tool failures.

What follows is the seven-day plan, the exact stack, honest tradeoffs between DIY and paid, the QA scenarios that tell you it's working, and the risks worth respecting before you turn the auto-reply on for real.

The 7-day plan

What you actually do, day by day.

Each day fits in a single focused block. Skipping the Day 5 anti-pattern audit is the single most common reason a lead responder install reads as AI to customers — don't skip Day 5.

Day 1

~60 min

Intake + KB confirmation

Write down the questions every new lead asks before they buy — pricing, scope, turnaround, who you serve, what you don't take on. If you already have an M1 receptionist, reuse that knowledge base; M2 reads the same Google Doc. The point of Day 1 is to make sure the AI replies from one source of truth, not three competing ones.

Day 2

~30 min

Pick your channels

List every place a stranger can message you: website chat widget, contact form, support email, Instagram DM, Facebook DM, WhatsApp, SMS to your business number. Decide which two or three you actually want auto-responded. Don't try to cover all seven on day one — start with the two that bring most leads and add the rest in week two.

Day 2–3

~45 min

Sign up for the tools

Tidio Starter for chat + DMs. Zapier Pro for the email + SMS routing. Twilio if SMS is in scope (one number, two minutes). An Anthropic API key for the deeper drafts. Optionally Resend or SendGrid if you want replies to land from your own domain without going to spam — verify SPF and DKIM the same day so DNS has time to propagate.

Day 3

~60 min

Configure Tidio Lyro

Paste your brand name. Connect the KB Google Doc. Paste the system prompt — including the anti-patterns blocklist. Install the chat widget script on your site. Wire OAuth to Instagram and Facebook if those are in scope. Send a test message from your own phone and confirm the reply lands in under 30 seconds.

Day 4

~90 min

Build the Zapier Lead Router

This is the biggest single piece of work. Six Zaps: Form Lead Router, Email Lead Router, WhatsApp router if approved, SMS router, Hot-Lead Owner Alert (SMS + Slack), and the 6pm Daily Digest. Each one classifies the inbound, drafts a reply via Claude, logs to your CRM or Google Sheet, and checks for emergency keywords. Test each path with a sample inbound before moving on.

Day 5

~60 min

Anti-pattern audit + QA

Submit three sample form inquiries per channel. Read every reply. If any contains a banned phrase from the blocklist, fix the prompt and re-run. Then run the ten QA scenarios from the install kit — happy-path booking, after-hours inquiry, hot lead with the word emergency, complaint, ambiguous lead, lead in a language you don't serve, spam form fill, lead asking for a refund (wrong channel), lead asking if it's a real person, and lead hanging up mid-thread.

Day 6–7

~15 min/day

Soft launch + watch

Turn the live channels on Monday morning. Watch every reply for the first day. Day 8 review every transcript. Day 14 review the leads that escalated and ask why. Day 30 pull the full report — total leads, percent auto-replied, percent escalated, average response time, hot-lead alert count. That report is what tells you M2 paid for itself.

The stack [Verified 2026-05-23]

Six tools. Most you can swap.

This is the default stack we use on paid installs. Tidio and Zapier are not affiliate picks — they're what we'd pick for our own business, which is exactly what we did before we sold this as a service. Pricing checked against vendor pages on 2026-05-23; verify again within seven days of a paid install start since SaaS pricing has been volatile.

Chat + DM responder

Tidio (Starter, $39/mo)

$39–$78/mo

Tidio Lyro handles chat widget + Instagram + Facebook DMs from one inbox. Good native AI, OAuth flows are quick, and the per-channel persona settings let you sound different on IG than on a corporate site without managing two tools.

Email + SMS router

Zapier Pro

$59/mo

Email replies need branching logic — classify the inbound, draft via Claude, log, check for hot-lead keywords, send. Zapier handles that in six Zaps. Make.com works too; Zapier is the safer pick for owners who'll touch the automations themselves.

Brain

Claude (Anthropic API)

~$10–$25/mo at SMB volume

The voice platform's default LLM produces AI-tasting replies. Routing through Claude with the anti-patterns blocklist enforces the human-voice rules. The cost per lead is single-digit cents. This is what makes Iris-built lead responses pass for the owner writing.

Phone number

Twilio

$1.15/mo per number + per-message

Only if you want SMS auto-replies. One number, two-minute provision, integrates with Zapier and Tidio. Skip if your leads never text you — most service businesses don't need it on Day 1.

Email deliverability

Resend or SendGrid

Free tier covers <3,000/mo

If replies need to send from your own domain — not a generic Tidio address — you'll need verified SPF + DKIM. Resend is the easier of the two for non-technical setups. Skip if you're happy with replies coming from the chat widget itself.

Knowledge base

Google Doc (shared with M1)

Free

If you already installed M1, do not create a second doc. M1 + M2 + M4 + M5 all read from one source of truth — that prevents your receptionist from answering one way and your lead responder from answering another.

All-in monthly

For a small business taking 30–200 leads/month, expect roughly $60–$80 all-in. Tidio is the floor; Zapier ops + Claude API cost is the variable.

The non-negotiable

Paste the anti-patterns blocklist into every channel.

“Thanks for reaching out! I'd be happy to assist you with your inquiry.” That sentence is in roughly nine out of ten AI lead responses we audit. Customers learn the cadence in a week and learn to ignore it in a month. Cold outbound has trained the entire internet to mute it on sight.

We catalogued ten categories of those tells with specific banned phrases and the human replacements we use instead. Before you turn any channel on, paste the full list into the Tidio system prompt and the Zapier Claude prompt. This single step does more for reply quality than picking a better model.

Read the full anti-patterns blocklist →

DIY or paid — honestly

When DIY makes sense. When it doesn't.

Most playbook PDFs end with “or just buy our thing.” This one is honest about when DIY is the better answer.

When DIY is the right call

You already run your own chat widget or email automations. You're comfortable in Zapier and Tidio dashboards. Your lead volume is under 50/week so a single missed reply during setup isn't catastrophic. You have ~4 hours of focused work over a week. Most owners who already have one or two AI tools running and want to wire a third belong here.

When Iris-Assist ($500) is the right call

You'd rather have a voice on the line guiding the clicks than read a checklist alone. You don't want strangers in your accounts but you want a real person beside you for ninety minutes. The middle path between DIY and a full install — you keep every credential, Iris keeps every step from going sideways.

When Iris Build Pilot ($997) is the right call

You'd rather pay for the install than learn Tidio + Zapier + Claude API + Twilio in one week. You want one channel live in seven days with a 30-day money-back guarantee. You're comfortable handing us limited-scope access to the accounts we install. Most owners over $200K revenue land here — the math on owner-hours-saved is brutal in our favor.

QA — the ten scenarios

Ten test leads before you flip the switch.

Submit each scenario via every channel you've wired. If any of the ten fails, fix the prompt or the KB and re-run. Don't go live until all ten pass on every channel.

  1. 01Happy-path inquiry — basic question with answer in the KB.
  2. 02After-hours inquiry — confirm tone shifts to “we'll be back at 9am” behaviour.
  3. 03Hot-lead trigger — message contains “emergency”, “urgent”, or “ASAP”. SMS + Slack to owner must fire.
  4. 04Complaint or angry tone — AI de-escalates and routes to a human.
  5. 05Pricing question NOT in the KB — AI escalates instead of guessing.
  6. 06Caller asks “Is this a real person?” — AI must answer honestly.
  7. 07Ambiguous lead — vague intent. AI should ask one clarifying question, not five.
  8. 08Wrong-channel request (refund, support, billing) — AI routes to the right team.
  9. 09Spam form fill — obvious bot content. AI should NOT reply, should mark + ignore.
  10. 10Lead in a language you don't serve — AI offers translation OR escalates politely.

Risks + safety rails

What to lock down before launch.

Channel sprawl. Turning on chat, email, DM, WhatsApp and SMS the same day means five places where a bad reply can embarrass you. Launch two channels in week one, add the rest one per week so you can actually read the transcripts.

Email deliverability. If your replies start landing in spam, every lead thinks you ghosted them. Verify SPF and DKIM on Day 2, then send three test replies to gmail, outlook, and yahoo addresses before turning the email Zap on.

Regulated industries. If you're in healthcare, legal, financial, or insurance, the AI must never diagnose, quote rates outside published ranges, or confirm whether someone is a client. Paste the relevant industry safety rails into the system prompt and run three adversarial test inquiries trying to get the AI to violate them.

Hot-lead misses. A real emergency that the classifier scores as “routine” is the worst possible failure mode. Lower the threshold for emergency keywords, and review every escalated alert in week one to make sure nothing slipped past.