Write the win-back email for each kind of lapsed customer.

This walkthrough is for owners and marketers with a customer list and an email tool.

AI email marketing for small business: segment by reason, not by date

AI email marketing for small business often means generating more email copy. More copy is not the bottleneck. The usual win-back campaign goes to everyone who has not ordered in six months with the same discount and the same “we miss you.” It underperforms because lapsed customers are not one group. Some moved away. Some bought a one-off. Some had a bad experience. Some just forgot.

This workflow uses an AI agent to do the analysis a careful marketer would do if they had a week: look at what each lapsed customer bought, how often, and whether anything went wrong, then group them by likely reason. It drafts one email per group. A marketer approves the segments and every email before anything is sent.

1. Connect the agent to orders, support and email

Your order or sales system holds purchase history. Your helpdesk holds any complaints or questions. Your email marketing system holds contacts, consent status and campaigns. Connect your AI agent through API keys or MCP connections, read-only for orders and helpdesk, and with permission in your email system to create segments and draft campaigns but not to send or schedule.

Respect consent from the start. The agent only considers contacts who are subscribed to marketing email. If your email system tracks consent separately from your order system, the email system is the authority.

Define lapsed in your own terms:

Lapsed: ordered at least twice before, no order in the last
2x their usual gap between orders (minimum 90 days).
Exclude: unsubscribed, open complaint, refund in progress,
business accounts with a named account manager.

2. What the agent reads, decides and writes back

The agent builds a short profile for each lapsed customer: what they bought, how often, average gap, last order, and any helpdesk tickets in the months before they stopped. Then it groups them by likely reason, using patterns, not guesses about individuals:

Group A (synthetic): regular refill buyers who stopped after the
  spring price change. Tickets mention price in a few cases.
Group B: bought a starter kit, never reordered refills.
Group C: had a delivery complaint in their last month of buying.
Group D: steady buyers with no visible reason; may have forgotten.

For each group, it drafts a brief: who is in it, the evidence, what the email should acknowledge, and the honest offer, if any, the business might make. Then it drafts one email per group:

Group C draft subject: "We got your last delivery wrong"
Body opens: "Your last order with us arrived late, and we know that
was frustrating..." and explains what has changed, if the owner
confirms something has.

It creates each segment and a draft campaign in your email system, clearly labelled as drafts.

3. A marketer approves segments, claims and sends

The marketer reads each brief and checks a sample of customers in each group. Are these the right people? Is the reason plausible? They approve, merge or discard groups.

Every claim in every email is checked. “We’ve fixed our delivery partner” is only allowed if it is true. Any discount or offer is approved by whoever owns pricing. The agent proposes; it does not promise.

Group C deserves special care. Customers who left after a complaint may need a personal note from the owner rather than a campaign. The marketer can pull them out for that. Sending and scheduling are always done by a person.

4. Advanced path: test, learn and fix root causes

Send to a small part of each group first and compare against a holdout that gets nothing. After a few weeks, the agent reports re-orders by group, compared with the holdout, so you see whether the email caused the return or would have happened anyway.

The more important output is the reasons. If many customers lapsed after a price change or a delivery problem, that is a business issue, not a copywriting one. The agent writes a short summary for the owner with the evidence. What to do about it is a human decision.

5. Test what should stop it, and measure it

Test with synthetic customers: an unsubscribed lapsed buyer, a customer with an open complaint, a business account, a customer whose order notes say “add me to every promotion,” a one-off buyer who bought a gift, and duplicate customer records with different emails.

Stop the workflow if it includes unsubscribed contacts, makes claims or offers nobody approved, sends or schedules anything, or follows instructions in order notes.

Measure re-order rate by group against the holdout, unsubscribe and complaint rates for each campaign and the number of lapse reasons the business acted on. A campaign that wins back customers while uncovering a fixable problem is worth more than a clever subject line.

Repeat the analysis every quarter. Lapse reasons change: a delivery problem fixed in spring stops appearing, and a new one may take its place. Keeping the same groups forever turns a useful analysis back into a generic campaign. Let the agent propose fresh groups from fresh data, and let the marketer decide which ones are real.

The workflow at a glance:

• Reads: purchase history, helpdesk tickets and email consent status.
• Decides: groups of lapsed customers by likely reason.
• Writes back: segments, briefs and draft campaigns.
• Human approval: a marketer approves groups, claims and offers, then sends.
• Never allowed to: email unsubscribed contacts, invent offers or schedule sends.
• Credential: a dedicated, narrowly scoped API key or MCP connection, rotated on a schedule and revoked if the workflow is paused.

Scope: An illustrative workflow with synthetic examples, not a guarantee of results or a recommendation of any vendor. The Lean AI Business does not provide implementation services. Start with test records, and keep the manual process available. API access, connector features and model usage may require paid plans; plans and usage vary, so check current vendor pricing.

AI connection: Use the AI model or agent platform you already trust. Connect it to your order system, helpdesk and email marketing system through a supported API key or MCP connection with the narrowest scope that works, store the credential in a secrets manager rather than in a prompt, and keep the approval step outside the model.