Personalization at Scale: AI-Curated Customer Journeys in E-commerce

1) What “AI-Curated Journeys” Actually Means

Strip away the hype. An AI-curated journey is simply this: for each customer and moment, you decide the next best action (NBA)—what to say, where to say it, and how to assemble content—based on signals that predict value for the customer and revenue for the business. AI enters where rules end: ranking choices, forecasting outcomes, and learning which path performs best under uncertainty.

Two clarifications:

• AI ≠ magic. It’s math plus data. If your events are messy and your content is monolithic, models can’t help much. Fix the basics first.
• Personalization ≠ coupons. The most powerful “personalization” in many programs is control (skip/swap/pause) and clarity (progress-to-perk), not a net-new discount. See our take in Loyalty Program Optimization.

2) The Three-Layer Architecture: Data → Decisioning → Delivery

Layer 1 — Data (make it actionable, not exhaustive)

You need a short, reliable list of events and profile properties—not a data lake. For Shopify brands, start with: Viewed Product, Added to Cart, Started Checkout, Placed Order, Shipment Update, plus subscription and loyalty events (if applicable). Add two zero-party fields you’ll use immediately—primary_goal and variant_pref—via a one-question prompt (Digioh).

Layer 2 — Decisioning (rules + models)

Combine deterministic rules (“don’t send cart SMS if an email went 20 minutes ago”) with probabilistic models (“what’s the conversion lift if we recommend bundle A vs. B?”). Keep a human-readable policy that the team can reason about, and let models rank choices within those constraints.

Layer 3 — Delivery (content + channels)

Build messages from modular blocks (hero, proof, recommendation, progress header) that can swap by segment, goal, or model output. Orchestrate channels so the customer hears one voice. Email carries depth, SMS carries clarity/control, push handles moments, and the site reflects the same decision.

{
  "profile": {
    "id": "abcd123",
    "email": "sam@brand.com",
    "phone": "+1..",
    "primary_goal": "Hydration",
    "variant_pref": "Vanilla 32oz",
    "loyalty": {"points": 820, "tier": "Silver"},
    "subscription": {"active": true, "next_charge": "2025-10-18"}
  },
  "events": [
    {"type":"ViewedProduct","sku":"EL-32-VAN","ts":"2025-10-03T15:20Z"},
    {"type":"AddedToCart","sku":"EL-32-VAN","ts":"2025-10-03T15:25Z"}
  ],
  "nba_context": {"channel":"email","window":"T+24h","journey":"Activation"}
}
      

3) Data That Matters: Events, Identity, and Zero-Party Signals

Events: the verbs of your system

• Commerce: Viewed Product, Added to Cart, Started Checkout, Placed Order, Refund Issued
• Fulfillment: Label Created → Out for Delivery → Delivered (pipe to Malomo for branded tracking)
• Loyalty: Points Earned / Redeemed, Tier Changed (Yotpo)
• Subscription: Upcoming Charge, Skipped, Swapped, Paused, Cancel Initiated / Saved (Recharge)

Identity resolution: the art of less chaos

Bind email, phone, and push token to one profile as early as possible (e.g., at first email capture). Store channel preferences (channel_pref, quiet_hours) and respect them rigidly. Consistent IDs let your models learn reliably and your journeys speak with one voice.

Zero-party data (ZPD): two questions you’ll use immediately

• primary_goal (Hydration | Performance | Calm …)
• variant_pref (size/flavor/shade)

Ask one of them in your post-purchase quickstart; ask the other on PDP or via email in week one. Use Digioh to launch quickly and map properties into Klaviyo. For the strategy behind ZPD, read our primer: Zero-Party Data 101.

4) Decisioning: From Rules to Models (Propensity, Recommenders, Bandits)

Rules: the guardrails

• Respect quiet hours (e.g., 8pm–8am local). SMS or push outside that window only for customer-requested events (delivery today, upcoming charge).
• Recency-gating: if push fires, delay the matching SMS/email 15–30 minutes.
• Frequency caps: promotional SMS ≤1/48h; email cadence mapped to engagement bands.
• Consent preferences: theme (“deals|drops|updates”) controls which journeys can fire.

Propensity & churn models: who is most likely to act?

Use simple, interpretable models first (logistic regression / gradient boosted trees) or your platform’s built-ins (Klaviyo’s predicted churn/next order date) to rank who should receive which treatment. You don’t need deep learning to decide whether to send a “subscribe & save” invite vs. an add-on prompt—start with clean features: recency, frequency, monetary value, event counts, and ZPD.

Recommenders: what should they see next?

• Content-based: “Because you liked {category} and your goal = Hydration, show {hydration set}.”
• Collaborative: “People who bought X also bought Y.” (Rebuy makes this practical.)
• Hybrid: use content filters as guardrails (“only hydration SKUs”), then rank by collaborative scores or popularity.

Multi-armed bandits: which creative wins today?

Classic A/B testing wastes traffic on losers. Bandits allocate more traffic to winners as results emerge (Thompson Sampling/UCB). Use bandits to choose between two subject framings or two offer framings within your NBA constraints. Keep exploration live even when you have a leader; tastes change.

if (channel_pref == "sms_only" && now in quiet_hours) defer
if (recent_send_within == 15m) defer
if (subscription_active && days_to_next_charge <= 3) NBA = "subscription_control"
else if (propensity_subscribe > 0.6) NBA = "subscribe_invite"
else NBA = "goal_based_cross_sell"
content = rank_recommendations(goal=primary_goal, variant=variant_pref)
deliver(NBA, content)
      

5) Content Assembly: Modular Blocks, Tokens, and Guardrails

You can’t personalize monoliths. Break emails into swappable modules: hero, proof (UGC), recommendation, loyalty progress, subscription control, and footer. Each module gets rules on when it appears and how it populates.

Example: Klaviyo conditional

{% if person.primary_goal == "Hydration" %}
  {% include 'module_hydration_set' %}
{% elsif person.primary_goal == "Performance" %}
  {% include 'module_performance_set' %}
{% else %}
  {% include 'module_best_sellers' %}
{% endif %}
    

Loyalty progress header

Sync points_to_next_reward and tier_name from Yotpo to Klaviyo and place a single line at the top of every lifecycle email: You’re {{ points_to_next_reward }} points from $10 off — add any of these to unlock it. This one line increases click-through without training discount behavior.

Brand voice + AI copy

Large language models can generate on-brand variations quickly, but keep a “style card” (tone, banned phrases, compliant disclosures) and manually QA anything that touches compliance. Use LLMs for micro-edits, not free-form offer logic.

6) Channel Choreography: Email, SMS, Push, Site—One Conversation

Personalization dies when channels compete. Build a simple choreography and enforce it mechanically:

• Email carries why and proof—comparisons, UGC, bundles.
• SMS carries what now—1-tap control or a single recommendation.
• Push handles moments—delivery today, restock now, reserve window.
• Site reflects the same NBA—PDP modules and cart prompts via Rebuy; branded tracking via Malomo.

Use the orchestration backbone in our Holiday Retention Calendar to pace peaks and keep holdouts intact.

7) Journey Playbooks: Onboarding, Activation, Expansion, Replenishment, Reactivation

Onboarding (Days 0–14)

• NBA: teach first use, collect ZPD.
• Modules: quickstart tips + 1-click survey; support reply (Gorgias); loyal-progress line.
• SMS: “Your quickstart is here → {short_link}. Questions? Reply here.”

Activation (Days 7–30)

• NBA: cross-sell by goal, subscription invite if propensity high.
• Modules: goal-based set (Rebuy) + variant-matched UGC; progress header; subscribe tile (Recharge).

Expansion (Days 30–60)

• NBA: bundle upgrade; VIP early access for top tiers.
• Modules: tier benefits, early access window; “complete the set.”

Replenishment (Ongoing)

• NBA: reorder or snooze (SMS); upsell add-on; keep cadence aligned to cadence_intent.

Reactivation (Lapsed)

• NBA: value first—“what you loved”—then small perk (points boost), not a permanent code.

For the email modules that power these journeys, start with 10 Email Automation Workflows.

8) Privacy, Consent, and Brand Safety (AI Guardrails)

• Consent scope: honor channel preferences and themes. “Deals only” means deals only.
• Data minimization: capture less, use more. If a field won’t change the next touch, don’t ask.
• Auto-suppress sensitive inferences: do not infer protected attributes; restrict models to commerce signals and explicit ZPD.
• Human-in-the-loop: require manual approval for new offer logic; lock critical copy (compliance, legal).
• Explainability: keep a simple policy doc (“NBA rules”) so the team can audit decisions.

9) Measuring Lift: Holdouts, Uplift, and Finance-Ready Dashboards

If you can’t prove lift, you’re guessing. Keep persistent holdouts and measure the right leading indicators. A simple framework:

• Holdouts: 10–20% at the message or flow level.
• KPIs: Second-purchase rate (30D), reorder interval, add-on attach, subscription save rate, RPR (revenue per recipient) vs. control.
• Uplift modeling (optional): estimate true incremental impact when targeting differs across cohorts.

Weekly dashboard

• NBA adoption (how often the decision engine picked each action)
• RPR by journey and by bandit variant (with holdout)
• Engagement health (opens/clicks/taps; time between touches)—framework here: Engagement as a Leading Indicator.

10) Platform Notes: Why We Orchestrate on Klaviyo (Top Partner)

We choose Klaviyo because it puts data, decisioning, and delivery close enough to behave like one brain: profile properties and events stream in from Shopify, Yotpo, Recharge, Rebuy, Attentive, and Malomo; segments and conditional blocks let us encode NBA rules; and built-in predictions (e.g., next order date) give lightweight propensity signals without custom ML. When brands outgrow heuristics, we keep Klaviyo for orchestration and feed it scores from specialized models—but the team still works from one playbook.

11) Case Examples (Anonymized) With the Math

Case A — Beauty (AOV $39): “Goal-based activation”

• Problem: Low second-purchase (30D 14%), generic recommendations.
• Intervention: Captured primary_goal via Digioh (Day-2); swapped activation modules by goal; added bandit on hero framing (proof-first vs. offer-first).
• Result (6 weeks, holdout-adjusted): 30D second-purchase +23%; RPR +31%; best bandit variant was proof-first.

Case B — Supplements (AOV $44): “NBA favors control”

• Problem: High churn at day-30; aggressive couponing.
• Intervention: If subscription_active and days_to_next_charge ≤3 → push/SMS with skip/swap/pause (Recharge) + add-on tile; else → email goal-based bundle.
• Result: save rate +24%; add-on attach +29%; coupon usage on repeat orders −33%.

Case C — Apparel (AOV $68): “VIP as status you can feel”

• Problem: Tiers were labels; launches messy.
• Intervention: Yotpo tier perks reframed as experiences (stock guarantee, early access window); NBA = send early access push only to top tier.
• Result: VIP share 28% → 41%; launch sell-through in VIP window covered 58% of week-one inventory; complaint rate −35%.

12) 90-Day Implementation Roadmap

Phase 1 (Weeks 1–3) — Foundation

• Verify core events from Shopify + partners (Yotpo, Recharge, Rebuy, Attentive, Malomo) into Klaviyo.
• Launch a 1-question ZPD prompt (Digioh) for primary_goal; map to profile.
• Document NBA policy (rules & guardrails); add recency gating, quiet hours, and frequency caps.

Phase 2 (Weeks 4–6) — Journeys & Modules

• Build modular email templates (hero, proof, recommendation, progress header, subscribe control).
• Turn on onboarding, activation, replenishment, and reactivation with holdouts.
• Wire SMS/push nudges carefully (no double-taps); route replies to Gorgias.

Phase 3 (Weeks 7–9) — Decisioning

• Implement simple propensity rules (Klaviyo predictions + behavioral features) to pick “subscribe invite” vs. “cross-sell.”
• Add a bandit test on one creative element per journey (subject framing, module order).

Phase 4 (Weeks 10–12) — Scale & Prove

• Publish weekly dashboard: NBA adoption, RPR by journey/variant, second-purchase, reorder interval, save rate, add-on attach, engagement health.
• Kill low performers, double down on winners. Keep exploration on.

13) Troubleshooting Cookbook

Symptom: “We’re sending more, revenue is flat.”

Likely causes: channels competing; recency gating missing; no NBA policy.

Fix: implement recency gating; pick one NBA per moment; suppress redundant sends; add holdouts to verify.

Symptom: “Personalization made unsubscribes spike.”

Likely causes: ignored theme preferences; over-targeted SMS; creepy inferences.

Fix: honor themes rigidly; add SMS Snooze; restrict models to commerce + explicit ZPD.

Symptom: “Models recommend items we don’t have.”

Likely causes: stale inventory feed; missing guardrails.

Fix: filter by in-stock; fall back to category bestsellers; refresh feed hourly on peak days.

14) Templates & Resources

• 10 Email Automation Workflows to Boost Retention
• Zero-Party Data 101
• Holiday Retention Calendar
• Loyalty Program Optimization & Management
• How Recharge Powers Subscription Retention
• Engagement as a Leading Indicator
• Work with Sticky Digital · Request a retention audit