How to Use AI to Write Retention Copy

AI-written retention copy performs 3-5x better than generic templates when done right. It performs worse than a good template when done wrong.

The difference is not the model. It is the prompt structure and the guardrails you put around what AI is allowed to produce.

The four things every prompt needs

Every effective retention copy prompt includes:

1. Specific customer usage data

Generic prompts produce generic output. Specific prompts produce specific output.

Bad: "Write a re-engagement email for a customer who has not logged in for 14 days."

Good: "Write a re-engagement email for Sarah, a marketing manager at a 40-person SaaS company. She signed up 8 months ago, used our Kanban view 4 times per week for 6 months, then stopped 3 weeks ago. She has not opened our last 2 emails. Her team has 5 seats. She last created a board titled 'Q4 campaigns.'"

The second prompt produces copy that references her actual behavior. Loss aversion kicks in. Open rates go from 4% to 22%.

2. Specific goal and CTA

"Retention email" is not a goal. Specify:

• Re-engage a drifting user → CTA is a specific in-product action
• Prevent churn on an at-risk account → CTA is a call with a CSM
• Recover a churned customer → CTA is a re-activation link with an offer
• Convert a trial user → CTA is upgrade with a specific plan

Every prompt should state the goal and the specific action you want the reader to take. Without this, AI produces polite non-actionable copy.

3. Voice constraints

AI defaults to a bland corporate voice. You have to force it away from that:

• "Under 120 words"
• "No em dashes, no 'moreover', no 'furthermore'"
• "Sound like a friend, not a marketing team"
• "Include one specific detail from the customer's account"
• "No rule-of-three lists"

These constraints remove the tells that AI-generated copy has. See the humanizer skill for the full list of AI writing patterns to avoid.

4. Context on your brand voice

Include 2-3 examples of previous emails you have written that worked. AI matches the style of examples faster than it matches abstract descriptions.

The workflow that produces good copy

Do not accept the first output. Run this loop:

1. Draft 1: AI writes the email from your prompt
2. Critique: Ask the AI to identify weaknesses in its own draft (this catches AI writing patterns better than you can)
3. Draft 2: AI rewrites based on the critique
4. Human edit: You spot-check the specifics (did AI hallucinate a feature name? get the tenure wrong?)
5. Humanize: Run through a pattern check for em dashes, "moreover", "in essence", and other AI tells

Total time: 5-10 minutes per email. Result: copy that performs like a human wrote it, personalized to the customer.

Where NOT to use AI

Three cases where AI-written copy hurts:

1. Very personal emotional beats. A message to a founder who just lost a co-founder, a message to a customer whose account just had a critical outage. Write these yourself.
2. Legal or compliance-critical copy. Cancellation terms, refund policies, data handling notices. AI can draft, humans must approve.
3. First-time templates for new sequences. Build the initial template with a human. Use AI to personalize the template thousands of times.

The tools to build this workflow

• Claude or GPT-4 API - the language model layer
• Customer.io, Loops, or Braze - to trigger emails based on behavior with the AI-generated content
• Segment or your data warehouse - to feed customer data into the prompts

Typical setup: 2-4 weeks of engineering to connect the pipeline. Ongoing cost: $50-500/month in API calls depending on volume.

The bigger retention strategy

See AI retention emails for the full sequence framework and how to write a cancellation email for a specific example.

To find out whether your email layer is your biggest retention gap, take the 60-second Churn Health Check.