E-Commerce Agents: Product Search, Support, and Merchandising
E-commerce is often the friendliest industry entry point for agents: catalogs are structured, intents are clear, and many actions are reversible. Risk is lower than clinical or trading domains, but not zero - bad pricing, over-promising, and refund abuse still hurt.
Use this cheatsheet to scope product search, support, and merchandising agents without turning the store into an unsupervised coupon printer.
How to Use This List
- Pick the job row that matches your funnel stage, then read the matching detail table.
- Prefer tools that read catalog truth over free-form product claims.
- Cap autonomy on money, inventory writes, and outbound campaigns.
- Revisit fraud and fairness when you enable discounts or account changes.
Quick Map by Funnel Stage
| Stage | Agent job | Primary tools | Autonomy default |
|---|---|---|---|
| Discover | Conversational product search | Catalog search, facets, stock | High for recommend; low for claims |
| Decide | Comparison / fit Q&A | Specs, reviews summary, size guides | Medium; cite sources |
| Purchase | Cart and checkout help | Cart APIs, payment status (read) | Low on payment changes |
| Post-purchase | Support and returns | Order lookup, RMA, shipping | Medium with $ caps |
| Grow | Merchandising and CRM assists | Promo rules, segment stats | Human approve for live pushes |
Product Search Agent Cheatsheet
| Capability | Good behavior | Failure mode |
|---|---|---|
| Natural language search | Maps to filters + ranked SKUs | Invents products not in catalog |
| Constraint handling | Respects budget, size, brand | Ignores hard constraints |
| Stock awareness | Hides or flags OOS | Sells ghosts |
| Personalization | Uses consented preferences | Creepy or policy-violating profiling |
| Explanation | Why this SKU matched | Generic hype copy |
| Multi-item baskets | Complements and constraints | Bundle prices wrong |
| Multimodal (photo) | Search by image via approved model | Wrong attribute extraction |
Search tool rules
| Do | Don't |
|---|---|
| Query the live catalog index | Rely on model memory for price/stock |
| Return stable product IDs | Hallucinate SKUs |
| Surface policy-safe attributes | Claim medical/performance outcomes you cannot support |
| Log query → result set for eval | Optimize only for clickbait titles |
Support Agent Cheatsheet
| Intent | Agent can often | Needs human / hard rules |
|---|---|---|
| Where is my order? | Track and explain | Fraud-flagged accounts |
| Return window | Cite policy + start RMA | Exceptions beyond policy |
| Damaged item | Collect evidence + open case | High-value claims |
| Cancel order | Cancel if still cancelable | After fulfillment cutover |
| Refund | Auto under $ threshold | Above threshold / chargebacks |
| Subscription change | Apply allowed plan ops | Retention discounts beyond matrix |
| Abuse / threats | Escalate | Safety workflows |
Support autonomy ladder
| Level | Examples | Guardrail |
|---|---|---|
| L1 FAQ | Shipping times, materials | Ground in policy KB |
| L2 Account-read | Order status | AuthN required |
| L3 Low-risk write | Coupon ≤ $X, RMA label | Caps + rate limits |
| L4 Money movement | Refunds, goodwill credits | Dual rules + fraud checks |
| L5 Policy exception | Out-of-window returns | Human only |
Merchandising Agent Cheatsheet
| Job | Agent assist | Human owns |
|---|---|---|
| Collection copy | Draft descriptions from attributes | Brand voice approval |
| Promo ideas | Suggest bundles from affinity data | Margin and brand fit |
| Onsite ranking experiments | Propose hypotheses + metrics | Experiment launch |
| Email/SMS campaigns | Draft variants | Send-time compliance + consent |
| Price recommendations | Analyze elasticity bands | Final price publish |
| UGC moderation assist | Flag policy violations | Edge-case takedowns |
Merchandising hard stops
- No unsupervised live price changes on whole catalog.
- No sending campaigns without consent and suppression lists.
- No fabricated scarcity ("only 2 left!") unless inventory API says so.
- No competitor-price claims without licensed data sources.
Cross-Cutting Risk Cheatsheet
| Risk | Signal | Control |
|---|---|---|
| Refund fraud | Repeat high-risk patterns | Risk scores gate auto-refund |
| Prompt injection via reviews | Tool reads UGC as instructions | Treat UGC as untrusted data |
| Over-discounting | Margin collapse | Promo matrix + approval |
| Accessibility/fairness | Steering by protected attributes | Policy on personalization features |
| Hallucinated specs | Returns and distrust | Catalog-grounded answers only |
| Agent loops on chat | Cost spikes | Max turns + handoff to human |
Architecture Fit (Lightweight)
| Pattern | When |
|---|---|
| Bounded ReAct | Shopper chat with search + cart tools |
| Plan-and-Execute | Multi-step concierge ("outfit for event under $200") |
| Multi-agent | Separate browse, support, and fraud-check specialists |
| Template workflow | Returns and cancellations with fixed stages |
KPI Cheatsheet
| Surface | Quality KPIs | Business KPIs |
|---|---|---|
| Search | Constraint satisfaction, zero-result rate | Conversion, AOV |
| Support | Containment with CSAT, recontact rate | Cost/contact, refund accuracy |
| Merchandising | Edit distance to approved copy, policy flags | Margin, campaign CTR (human-launched) |
FAQs
What is the best first e-commerce agent?
Authenticated order-status and policy Q&A with catalog-grounded answers and no money movement - or conversational search that only reads the catalog index.
Should the agent set prices?
It can recommend within rules. Publishing prices should go through merchandising systems with approvals and audit.
How do I stop product hallucination?
Every product claim must reference a catalog product ID returned by a tool. If the tool returns nothing, say so and refine search.
Can support agents auto-refund?
Yes within strict amount, velocity, and fraud-score caps, with sampling review. Unlimited auto-refund is a loss-prevention incident waiting to happen.
How is this different from a classic chatbot?
Agents can multi-step across tools (search, cart, RMA). That power requires stop conditions and allowlists classic FAQ trees never needed.
What about user-generated reviews in context?
Use them as data to summarize, never as instructions that change tools, policies, or discounts.
Do I need multi-agent for a store bot?
Not at first. Split specialists when fraud checks, merchandising writes, and shopper browse need different privileges.
How should size and fit agents behave?
Prefer size charts and structured attributes; avoid absolute guarantees; route high-uncertainty cases to human or clear return policy.
What evals matter for search agents?
Offline: constraint match, catalog validity, diversity. Online: conversion quality, not only clicks on flashy wrong items.
When is an e-commerce agent premature?
When catalog APIs are unreliable, auth is weak, refund rules are undefined, or leadership wants full autonomous pricing and payouts on day one.
How do international stores change the design?
Locale-specific policy, tax, shipping, language, and consent rules should be data-driven policies - not one global prompt hoping for the best.
Where do human agents still win?
Hostile customers, complex damage disputes, VIP exceptions, and anything that needs empathy plus policy exception authority.
Related
- Industry-Specific Use Cases Basics - retail vs bank support comparison
- Why Industry Context Changes an Agent's Risk Profile - risk levers still apply
- Financial Services Agents: Fraud Triage, Reporting, and Trading Support - fraud thinking that transfers
- Industry-Specific Use Cases Best Practices - adaptation practices
- Consumer-Facing Agent Products: Shopping, Travel, and Concierge Bots - consumer product lens
- Regulated-Industry Agent Checklist: Compliance, Audit, and Human Sign-Off - when stakes rise
Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana
@solana/web3.js+@solana/spl-token.