AI: Detecting, Predicting and Personalizing Peaks at Scale
Why AI changes the game
Until recently, designing a personalized peak per customer was economically impossible beyond a few dozen prospects. AI changed the equation:
- It detects peaks and endings in past interactions
- It predicts the ideal moment to trigger a peak
- It generates truly personalized peaks at scale
- It orchestrates endings without constant human intervention
With AI, a solo entrepreneur can deliver 10,000 personalized peaks per month — something previously reserved for 100-CSM teams.
Detecting peaks in past interactions
Conversation-by-conversation sentiment analysis
LLMs can analyze a transcript (call, chat, email) and identify:
- The emotional peak (positive or negative)
- The ending (final emotion)
- The distance between the two
Sample prompt to use with an LLM:
You are a customer experience analyst.
Here's the transcript of a sales call [TRANSCRIPT].
Analyze it using Kahneman's Peak-End Rule:
1. Identify the PEAK MOMENT (timestamp + excerpt)
2. Classify it: positive / negative / neutral
3. Perceived intensity: 1-10
4. Identify the ENDING (last 2 minutes)
5. Classify it: warm / cold / flat / abrupt
6. Give a MEMORY FORECAST (1-10)
7. Propose 3 concrete IMPROVEMENTS to the peak and the ending
Respond in JSON.
Aggregate analysis: where are your structural peaks?
Running 1,000 transcripts through this pipeline maps out:
- Recurring positive-peak moments (e.g., minute 18-22 when showing a specific figure)
- Recurring negative-peak moments (e.g., minute 35 when pricing is introduced)
- Types of endings that correlate with conversion
These insights reshape all sales training.
Predicting the next peak's timing
Propensity-to-act models
With a customer's usage history, a model predicts:
| Behavioral signal | What it indicates |
|---|---|
| 3 close-together logins | Strong intent — ideal peak moment |
| Sudden drop in usage | Churn risk — need for a retention peak |
| Milestone reached (e.g., 10 projects) | Moment to orchestrate a pride peak |
| Time since last contact | End of cycle to nurture |
Simple rule example
IF usage_30d > historical_avg * 1.5
AND last_positive_peak_days > 21
THEN trigger_personalized_peak(type="recognition", level=high)
This is event-based logic, but when enriched by an LLM that generates the peak content, personalization becomes industrial.
Generating personalized peaks
The LLM-generated peak
A handwritten personal peak takes 15 minutes. An LLM-generated peak with the right prompt takes 3 seconds and retains 90% of the effect — if the prompt injects real context.
Customer context:
- First name: Sacha
- Product used: [PRODUCT]
- Last measured result: 47 leads in 14 days
- History: started 3 months ago, 2 tickets resolved
- Tone preference: concrete and warm
Task: write a 4-5 sentence message that celebrates
the recent usage peak, mentions a specific (not generic)
detail, and suggests a next step.
Constraints:
- No marketing jargon
- No "we are thrilled to"
- No emojis except 1 max at the end
- Tone must read like a human who actually read the data
Example of a well-generated AI peak
"Sacha, 47 leads in 14 days — that's literally 3x the average on accounts with your profile. What stands out: 80% are from your LinkedIn campaign. If we double down on it next month, 100 is realistic. Can I block 20 minutes Wednesday to dig in?"
This message:
- Uses the customer's first name
- Cites a real figure
- Compares to a benchmark
- Proposes a concrete next step
- Doesn't read like a newsletter
Orchestrating endings at scale
The automated ending that doesn't feel automated
For the end of a contract, cycle, or project — AI can produce:
- A personalized results report
- A thank-you letter generated from history
- A personalized AI video (talking avatar with first name and figures)
- A roadmap of possible next steps
Guardrails to set
| Guardrail | Reason |
|---|---|
| Human validation of peaks above a threshold | A bad peak is worse than no peak |
| Content 100% based on verified data | Made-up claims destroy trust |
| Detection of negative signals (upset customer) | A celebratory peak sent to an angry customer = disaster |
| Per-customer rate limit | No more than 1 AI peak every 2 weeks |
Prompt engineering for peak-end
Meta-prompt structure
ROLE: you are [seller/founder persona], writing
personally to your customers.
GOAL: produce a peak message that celebrates
a specific customer accomplishment.
CONSTRAINTS:
- No more than 80 words
- Cite at least 1 figure from their history
- No marketing jargon
- Close, respectful tone, no forced familiarity
- End with an open question or open door
CUSTOMER DATA: [JSON injection]
AVAILABLE EVIDENCE: [real metrics]
FORBIDDEN:
- "We are thrilled", "Dear customer", "Feel free"
- Any compliment not backed by data
- Any mention of promo or upsell
Best practices
- Inject factual data rather than adjectives
- Ban generic marketing vocabulary (automatic wording check)
- Vary openings (not the same 3 phrases for 10,000 customers)
- A/B test two peak variants systematically
AI detection of negative peaks
Weak signals
An LLM continuously analyzes interactions and detects:
| Signal | Indicator |
|---|---|
| Frustration vocabulary | "still", "again", "disappointed" |
| Lengthening response times | Emerging disengagement |
| Cancellation questions | Imminent negative peak |
| Tone shifts | Casual → formal, warm → cold |
Triggering human intervention before the negative peak crystallizes prevents major churn.
AI-piloted recovery peak
IF negative_score_detected > 0.7
THEN
- notify CSM with summary
- pre-write a recovery message
- propose 3 compensatory gestures matched to severity
- DO NOT send automatically
Key principle: the AI prepares the recovery peak, but the human signs the gesture.
Ethical risks of AI × peak-end
Risk 1: personalization that feels like surveillance
"I see you've logged in 14 times this week" — sounds creepy. Stay data-aware without becoming intrusive.
Risk 2: the fabricated peak
Pretending an accomplishment that isn't one. Customers detect it quickly, trust is destroyed.
Risk 3: the manipulated ending
Using a warm ending to set up a hidden upsell. Rule: a sincere ending has no immediate commercial goal.
Risk 4: emotional deepfake
"Personal" CEO video fully generated, without the CEO's involvement. Legal and ethical gray zone — transparency is mandatory.
Use case: Amazon and end detection
Amazon has published several patents on journey-end detection:
- Delivery: end notification with a real photo of the dropped package
- Service: an automatic email 7 days after purchase to check satisfaction
- Prime: an annual summary of savings (personalized report = pride peak)
These endings transform transactions into memories — and explain part of the platform's structural loyalty.
Summary
AI scales the Peak-End Rule by making it operationally viable: automatic peak detection in past interactions, optimal timing prediction, generation of genuinely personalized messages through real-data injection. The real lever isn't the technology — it's the quality of the data you inject into prompts and the ethical discipline that prevents manipulation. In the next chapter, we'll see how to embed the Peak-End Rule in product strategy, retention, and viral growth.