AI Agent for eCommerce Analytics: Ask the Store With Evidence (2026)

Merchants already have charts. What they lack is a disciplined way to ask: why did sales decline? which products have increasing refunds? which customers are at risk? what changed this month? which products underperform? — and get an answer that cites systems, not a story.

Baymard still puts average cart abandonment at 70.22% (updated Sep 22, 2025). Useful industry context when you are scoping checkout UX. It is not an answer to “why did our sales decline this week.” If the sales tool shows a category drop and you have no funnel tool, say so. Do not paste Baymard into the evidence slot.

This is post 6. The morning push brief is the operations agent. This post is pull. Review text mining is review analysis. The map of all fifteen rows is the automations pillar.

The job. Answer Monday questions with a named tool and a time window — or refuse.

This week. Five read tools. Answer contract: answer, evidence, caveat, next human action. No price writes.

A person still signs. Price changes, campaign pauses, markdowns. Analysis only on this harness.

Skip it when you have no curated tool catalog, when the warehouse is the API, or when leadership expects a TikTok story without an ads tool.

We are not publishing an “insight lift.” Fixture SKUs are demo data. Every answer cites a tool, or the agent refuses. That feels less helpful than a confident story. The alternative is a bot that invents causality.

Copy the catalog — Use analytics-tool-catalog.md. Every answer needs tool name + filter + time window. Folder: ecommerce-ai-agents-series/.

Our take: every answer cites a named tool and a time window, or the agent refuses. More “I don’t have a tool for that.” Operators stop treating the bot as an oracle.

Question → tools → evidence

flowchart TD
  BusinessQuestion[Business Question]
  AIIntelligenceAgent[AI Intelligence Agent]
  ControlledAnalyticsTools[Controlled Analytics Tools]
  OrdersCustomersProducts[Orders Customers Products]
  Analysis[Analysis]
  AnswerEvidence[Answer + Evidence]
  BusinessQuestion --> AIIntelligenceAgent
  AIIntelligenceAgent --> ControlledAnalyticsTools
  ControlledAnalyticsTools --> OrdersCustomersProducts
  OrdersCustomersProducts --> Analysis
  Analysis --> AnswerEvidence

Controlled access means the model sees tool JSON, not a warehouse. Orders, customers, and products are views behind named operations, not SELECT *. There is no native Shopify connector — your signed-in APIs call Admin, the order system (OMS), or a warehouse semantic layer.

Five questions, five tools

From analytics-tool-catalog.md:

Answer contract (fail closed)

Context: pin a model your account allows; instructions must require this shape. If a tool returns fewer than N rows, say so. Do not fill in the chart.

Answer
Evidence: tool name + filter + time window
Caveat: sample size / incomplete tracking
Next human action (if any)

Example — fixture, not a client:

Answer: Apparel revenue is down vs the prior 7 days; AOV is flat; order count is down.
Evidence: salesByPeriod filter=category:apparel window=2026-08-20..2026-08-27 vs prior 7d
Caveat: no ads tool attached; cannot attribute to paid social. No conversionByStep rows.
Next human action: ask merchandising whether a hero SKU went OOS (getInventoryRisk on top movers from whatChanged).

That is an evidence-based answer. “Sales declined because shoppers abandon carts (industry 70.22%)” is a failed eval even if the prose is fluent.

What the agent may say — and what it must refuse

May: compare windows the tools support; name SKUs the tools returned; state sample size; list not_covered when the catalog has no ads or wholesale tool.

Must refuse: why-questions with no matching tool; customer lists; price changes; “pause Meta”; medical or organic claims from reviews; filling empty series.

Identity: associate/admin for store intelligence. Shopper tokens should never reach these tools. Memory holds the question and tool refs, not a customer export.

When tool count grows, gateway semantic search shortlists operations — same pattern as the Gateway post. Five tools do not need it; twenty will.

What broke

What broke — Sample eval set, turn “Why did sales decline?” The instruction was soft (“be helpful”). The model never called salesByPeriod. It cited Baymard cart abandonment and a TikTok narrative. Detection: trace with zero analytics tools; eval asserts evidence_tool present. Fix: hard refuse without a catalog tool; golden task fails the build if the TikTok sentence appears. Lesson: conversational BI without a tool contract is a story generator.

Second failure: customerRisk implemented as a thin SQL view that selected email. The model summarized “12 VIPs” and listed addresses in the answer. Detection: payload schema test (email regex in tool output). Fix: aggregates only at the API; response filter; Memory TTL on the session.

What this post doesn’t cover

• Push briefs and Slack caps — operations agent.
• Review theme extraction and public replies — review analysis.
• QuickSight Q, Amazon Q in QuickSight, or replacing your BI warehouse.
• Paid-media creative scoring and incrementality (no tool in this catalog).
• Native Shopify connector (does not exist).
• Measured “faster insights” or revenue attribution from a FactualMinds store engagement — not published.

What to Do This Week

1. Write five questions your operators already ask on Monday. If you cannot name the system that would answer each one, you are not ready for an agent.
2. Implement those as OpenAPI tools (start with salesByPeriod and productRefundRate only if that is all you have).
3. Copy the answer contract from analytics-tool-catalog.md into instructions.
4. Golden evals: one turn per question with expected operation; one turn that must refuse (causality with no tool); one personal-data trap.
5. Hard block: read-only; no updatePrice; Browser off.
6. Put tool traces next to the chat UI so a human can see evidence without trusting the prose.
7. Cost 50K sessions only if you actually expect that volume — most internal BI agents will not; still run the calculator so follow-up turns do not surprise you.

If you only do one thing: fail the answer when evidence is missing.

Need a governed analytics tool layer on Bedrock for retail? Start from Generative AI on AWS or contact us. Industry page: retail and eCommerce on AWS.

For your technical lead

On June 17, 2026, AgentCore Harness reached general availability, so a conversational analytics agent can be a config-driven loop with Gateway tools instead of a notebook that someone pastes into Slack (What’s New). After July 30, 2026, Bedrock Agents Classic is in maintenance for new customers (lifecycle).

First-party signals we reuse (not eCommerce outcomes) — Gateway server-side tools cut median tool round-trip ~180 ms → ~95 ms on a B2B CRM assistant (12 tools, ~8k turns/day) — Gateway post. Platform TCO silhouette: support-style AgentCore at 50K sessions/mo ~$791/mo platform + model (decision guide). Chatty BI sessions are session time + tokens; model them on the AgentCore pricing calculator before you offer “ask anything” to the whole company.

FAQ

When should we NOT deploy an AI agent for eCommerce analytics?

Skip it when you have no curated tool catalog, when the warehouse is the API, or when executives expect the agent to invent causality (“ads,” “TikTok,” “checkout UX”) without a matching tool. A BI tool with saved queries is cheaper if the question set is five tiles nobody will converse with. Also skip the older Agents Classic product for net-new builds after July 30, 2026.

What could go wrong if answers are not required to cite a tool?

The model will fill the chart. Sample turns answered “why did sales decline?” with industry cart-abandonment rates and a social-channel story while salesByPeriod was never called. Detection: evals that fail when evidence is missing. Fix: refuse the answer; return the caveat and the next human action.

What could go wrong if customer risk returns raw personal data to the model?

Emails, phones, and address lines land in session memory and in Slack paste. The catalog requires segment counts, not a customer export. The tool implementation must strip personal data before the model sees the payload — prompts that say “do not show PII” are not a control.

How is this different from the daily operations agent?

The operations agent pushes a capped morning brief. This agent is pull: a human asks a question, the intelligence agent calls controlled analytics tools, and it returns answer plus evidence. Do not merge them in week one or the brief will become a chat log.

Can the analytics agent change prices or pause ads?

No in this sample. Analysis and recommendation only. Price and campaign writes belong behind a hard block and a person on a different tool set. An analytics agent with a price update (updatePrice) is a merchandising agent you did not review.

Simple managed loop or heavier host for conversational store intelligence?

The simple managed loop (Harness) fits a single-domain Q&A loop with five or fewer analytics tools. The heavier host (Runtime plus Strands) if you later attach review intelligence or exception specialists with hop caps. There is no native Shopify connector — wrap Admin or warehouse APIs yourself.