M4 · Free playbook
Existing customers ask the same fifteen questions over and over. M4 answers those in thirty seconds, twenty-four hours a day, in your voice, while routing the few that actually need a human to the right team. Seven-day install, same playbook we use on paid installs. Free.
TL;DR
An AI customer-service layer is four things wired together: a support tool with an AI module, a knowledge base, a classification + escalation tree, and a CSAT loop that scores the AI's resolutions. The stack costs about $25 to $300/mo depending on volume and takes roughly four hours of focused work across seven days.
The hard part is not the AI — Intercom Fin and Help Scout AI both work well out of the box. The hard part is the escalation tree. A bad reply to a new lead loses an opportunity. A bad reply to an existing customer churns a relationship and earns a one-star review. The safety rails matter more here than anywhere else in the Iris stack.
What follows is the seven-day plan, the exact tool decisions per volume, the honest tradeoffs between DIY and paid, the QA scenarios, and the risks worth respecting before you let an AI answer a complaint without a human in the loop.
The 7-day plan
Each day fits in a single focused block. The Day 4 escalation tree is the work — the rest is wiring.
Day 1
~90 min
Pull the last 30 days of support tickets. Tag each by category — order status, password reset, refund, complaint, billing, product question, account change, integration help, legal trigger, other. The category mix tells you which industry overlay to apply and what auto-resolve looks like for your business. If you already run M1 or M2, the KB is shared — don't make a second one.
Day 2
~60 min
Under 100 tickets/mo: Crisp Plus at $25/mo. 100–500 tickets/mo: Intercom Fin at $0.99/resolution (the cleanest pricing model). Already on Help Scout: stay there and turn AI Drafts on. Over 2,000 tickets/mo: stop reading this playbook and book a consult — you need Front + Intercom hybrid, not a 7-day install. HIPAA requires the BAA path; budget 1–3 business days for Intercom to countersign.
Day 3
~60 min
Sign up. Subscribe. Enable Fin (or AI Drafts on Help Scout, or AI Responder on Crisp). Connect the shared KB Google Doc as a knowledge source. Paste the support system prompt with the industry safety rails. Wire the custom actions you can actually support — check_order_status, lookup_account, issue_refund. Skip the ones you can't back with real data; a broken action is worse than no action.
Day 4
~90 min
Build the category taxonomy — 8 base categories plus 5 industry-specific from the overlay. Wire the escalation tree: legal-trigger keywords go to the owner's Slack within 60 seconds; refunds over your threshold queue for human review; complaints route to the team lead. The refund auto-resolve cap is the highest-stakes number in this install — write it in the system prompt, in the custom action config, and in the install doc. Three places.
Day 5
~45 min
Wire the CSAT survey (1-question Y/N is the simplest; 5-star is better signal but lower response rate). Configure the daily digest email — total tickets, % auto-resolved, top categories, escalations. Build the dashboard inside whichever tool you picked. Verify the digest hits the owner's inbox at 6pm local time.
Day 6
~60 min
Generate 5 sample replies. Zero anti-pattern hits required. Run all 10 base QA scenarios plus the 5 industry-specific ones from your overlay. Test escalation: submit a ticket with a legal keyword, confirm owner Slack fires within 60 seconds. Test the refund cap with a ticket under and a ticket over the threshold — the over-cap must queue, never auto-resolve.
Day 7
~15 min/day
Turn the AI on in queue mode for the first week — every AI reply goes to a human for one-click approval before sending. Day 7 review the queue: how often did the human edit before sending? Day 14 check CSAT; under 60% means retune. Day 30 pull the report — auto-resolution rate, CSAT, top categories, time saved. That tells you when (or if) to flip to autonomous mode.
The stack [Verified 2026-05-23]
M4 is the simplest stack of any module — the AI module is the work; the tooling around it is light. Intercom Fin's per-resolution pricing has been adjusted twice in the last year; re-verify against vendor pages within seven days of any paid install.
Support AI (under 100 tickets/mo)
$25/mo
Cheapest viable AI-on-support stack. Native AI responder is good enough at low volume. Don't overpay for Intercom when you're only handling 50 tickets a month — the per-resolution math doesn't make sense yet.
Support AI (100–500 tickets/mo)
$39/mo Starter + $0.99/resolution
Per-resolution pricing — $0.99 per ticket Fin actually resolves; nothing if it doesn't. Cleanest economics in the space and the best AI in 2026 for SMB support volumes. The default choice.
Support AI (existing Help Scout user)
$25–$50/user/mo
Don't migrate to a new tool to add AI. Help Scout Plus or Pro includes AI Drafts and AI Summarize. Tune the personality, connect the KB, ship. The migration cost from Help Scout to Intercom isn't worth it under 500 tickets/mo.
Knowledge base
Free
Same KB doc your receptionist and lead responder read from. If you make a second doc just for support, your support replies will drift away from your lead replies in a month. One source, no drift.
Escalation channel
Free (existing tool)
Legal triggers and complaints have to route to a human within 60 seconds. Slack is the default — use your team's existing workspace, channel #customer-support. If there's no Slack, email escalation to owner@business works; SMS is overkill until you scale past 500 tickets/mo.
CSAT survey
Included
All three tools have CSAT built in. Don't add Delighted or a separate survey tool — the native survey ties scores to specific tickets and to the AI vs human resolver, which is exactly the data you need.
All-in monthly
Volume drives cost. Under 100 tickets/mo: $25 all-in on Crisp. 100–500/mo: $50–$300 on Intercom Fin (the variable is resolution count, not headcount). Over 500/mo: more nuanced — book a consult before committing.
The non-negotiable
“I completely understand your frustration.” “Thank you for your patience.” “Rest assured.” Customers reading those phrases at 9pm on a Tuesday after a bad day immediately know they're talking to a script. In sales context that costs you a lead; in support context that costs you the relationship.
The replacement isn't flattery — it's specificity. “Got it — your order #1284 shipped Tuesday and was delivered Wednesday at 11:14am. The tracking link sent it to the wrong door. Here's how we fix that.” Paste the blocklist into the support system prompt under a NEVER USE section and audit five sample replies before going live.
Read the full anti-patterns blocklist →DIY or paid — honestly
Most playbook PDFs end with “or just buy our thing.” This one is honest about when DIY is the better answer.
You're handling fewer than 100 tickets/mo and you already know the categories. You're comfortable in your support tool's admin panel. You can spare four hours over a week. The stakes are bounded — a bad reply to a tier-1 ticket is recoverable as long as the escalation tree works for the tickets that matter.
You'd rather have a voice walking through the Intercom Fin setup than read the docs alone. You don't want strangers in your support tool but you want the prompts and the escalation tree built right the first time. Most owners on Help Scout who want to add AI properly land here — one call, ninety minutes, you keep every credential.
Your support burden is the bottleneck you're trying to fix and you'd rather pay for the install than learn the platform. Auto-resolving 60–80% of tier-1 tickets is worth thousands of dollars of owner-equivalent time per month. Most owners spending 5+ hours/week on support belong here — the math goes from minus-five-hours-a-week to plus-thousands-per-month inside thirty days.
QA — the ten scenarios
Submit each as a real test ticket. If any fails, fix the prompt or the KB and re-run. Don't go live until all ten pass — especially the escalation tests.
Risks + safety rails
Auto-refund fraud. The single biggest financial risk in M4. If the refund custom action doesn't cap at the threshold you set in intake, a single bad actor can drain your refund budget in an afternoon. Cap it in three places: the system prompt, the custom action config, and the install doc.
HIPAA + BAA. If you're healthcare-adjacent, do not process PHI before Intercom countersigns the BAA. Not for testing, not for “just one ticket”. Use synthetic test data during the 1–3 day wait. If the BAA blocks past five business days, switch to Crisp Enterprise (HIPAA add-on with a faster BAA workflow).
Escalation tree silence. If a complaint fires escalation and the Slack webhook is broken, the owner finds out from a one-star review three days later. Test the escalation tree every Monday morning for the first month — submit a synthetic legal-trigger ticket, confirm the alert fires.
Migration cost. If you're on Zendesk or Freshdesk and switching to Intercom, the customer history migration takes a full day and needs scheduled downtime. Don't start it on a Monday. Don't start it during a sale. Plan for a Wednesday-Thursday window with the team warned.
DIY the whole thing free. Or pay $500 and Iris is on a call walking through the escalation tree. Or pay $997 and we install M4 inside your support tool in seven days.
Free
Download the PDF
This playbook as a PDF you can print and check off.
Get the PDF →
$500
Iris-Assist
You do the clicks; Iris is on the line. One 90-min call. You keep every credential.
Book a session →
$997 + $97/mo
Iris Build Pilot
We install M4 in your support tool. Seven days. 30-day money-back guarantee.
See Pilot →