Personalization & Conversational Commerce

Scenario Abstraction

The product knows a non-trivial amount about its user — what they've bought, watched, asked about, struggled with, signed up for. The classical personalization stack (collaborative filtering, segment-based campaigns) only goes so far. LLMs make it possible to:

• Reason over an individual user's history in natural language,
• Generate content tailored to that user,
• Converse with them about choices.

The scenario is broader than "chatbot." It spans personalized recommendation explanations, dynamic onboarding, conversational shopping, individualized education, and adaptive in-app guidance.

Solution Shape

1. User profile — a structured representation of the user: identity, preferences, history, current context (in-session intent, device, time of day).
2. Retrieval over user history — pull relevant past events (orders, sessions, support tickets) as context.
3. Catalog / capability retrieval — what products / lessons / actions are available right now.
4. Personalization decision — choose what to recommend, suggest, or talk about next. This may be classical (ranker) or LLM (planner-style).
5. Generation — LLM produces the user-facing artifact: a recommendation list with personal reasons, an explanation, a conversational reply, a lesson plan.
6. Action affordance — the output includes ways to act (buy now, add to plan, enroll, schedule).
7. Feedback loop — clicks, dwell, refusals, conversions feed back into both the ranker and the prompt eval set.

The non-LLM ranker is rarely fully replaced by the LLM; usually the ranker still selects candidates and the LLM does explanation, dialog, and re-ranking within a small candidate set.

Key Building Blocks

• User profile service with privacy controls and per-tenant isolation.
• Catalog / capability index — embeddings + structured attributes.
• Candidate ranker — classical recommender as the prefilter.
• LLM personalizer — works on candidates plus user context.
• Conversation manager — state, slot filling, handoff to a deterministic checkout / enrollment flow when ready.
• Guardrails — what the LLM may and may not promise (price, stock, eligibility).

Concrete Cases

• Conversational shopping assistant. "I need a jacket for a wedding in Tokyo in October" — clarify, recommend from catalog, explain choices, support follow-ups, hand off to checkout.
• Personalized recommendation explanations. Same recommendations as today's homepage, but each tile carries a one-sentence reason: "because you liked X and the author of Y mentioned this in an interview."
• Onboarding adaptation. First-run flow that adapts to what the user already knows (asked during signup) and which features matter for their stated goal.
• Adaptive learning paths. Tutor that picks the next lesson / exercise based on the student's recent mistakes, with conversational hints.
• In-app help that knows the user's state. "I see you're on the billing page with a failed payment — want to fix the card on file?"
• Personal finance coaching. Given transactions + goals, suggest budgeting moves; conversational, not a static report.
• Personalized news / content feed. LLM-curated daily digest with a written intro and "why you're seeing this."
• Travel planning. Dialog → preferences → itinerary draft → bookable elements.
• Healthcare care plan companion. Daily check-ins, medication reminders, escalation rules.
• B2B sales assistant for buyers. A guided buyer-side assistant that helps prospects spec configurations and pricing without a human rep on simple deals.

Similar Scenarios

• In-product copilots — narrower scope (one tool, one user), same personalization shape.
• Loyalty program nudges — generated per-user but typically one-shot, not conversational.
• HR self-service — personalized to employee's role / country / tenure.
• Concierge for premium customers — same engine, higher autonomy and broader catalog of actions.

Pitfalls & Evaluation

• Personalization theater. Generating "because you liked X" reasons that aren't true. The LLM must be given the actual signal it's allowed to cite; otherwise it confabulates.
• Cold start. New users have no history; the system must degrade gracefully to a strong default, not produce confidently wrong personalization.
• Privacy boundaries. What the LLM is allowed to reveal back to the user (about themselves) must be explicit. Mentioning sensitive inferences ("we noticed you've been browsing late at night") is creepy and sometimes illegal.
• Stuck-in-chat. Some users want a list, not a conversation. Always offer the structured alternative.
• Promise containment. The LLM must never quote price, stock, or eligibility unless those come from a tool call; otherwise users will be misled.

Useful metrics: per-user conversion / engagement uplift vs control, explanation usefulness ratings, conversation length to first successful action, abandonment rate, downstream support contact rate (good personalization reduces "what is this?" tickets).