AI Agent for eCommerce Upselling and Cross-Selling (2026)

The useful upsell is not “people also bought.” It is the next SKU that still matches the cart: fits, in stock, not already owned, not a restricted add-on, not a price rewrite. Collaborative filters will happily recommend a bag that does not mount to the camera in the basket. That is an attach. It is also a return.

This post is cart-aware ranking. Discovery-from-intent lives in AI product recommendation agents. Catalog completeness lives in catalog management. Do not collapse the three.

On 17 June 2026, AgentCore Harness reached GA (What’s New). Agents Classic is in maintenance for new customers after 30 July 2026.

This is post 15 in the 15-automations map. It is not an anonymized client engagement. We do not invent attach rate or AOV lifts. Baymard 70.22% is checkout leak context, not a promise that add-ons will save it.

First-party signals we reuse (not eCommerce client 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). Model widget volume on the AgentCore pricing calculator.

Reproduce this — Copy upsell-cross-sell-decision-matrix.md. Fill Owner and Approval. Co-occurrence is secondary. Ship gates: monday-checklist.md.

Opinionated take: next-best is contextual. Relationships + cart + history + inventory + margin-as-tie-break. Not co-occurrence only. Trade-off: you need a compatibility graph. Without it, ship a static merch rule and skip the agent.

The business problem

Checkout and cart drawers show add-ons. Most are a merch pin or an FBT table. They do not know the shopper already bought the spare battery. They do not know the lens mount. They will still “upsell” a thinner-margin hero if the model is told to maximize AOV.

Why it matters: a bad attach is a return and a support ticket. A good attach is guided selling at the moment of intent. Who breaks without compatibility fields: anyone selling accessories, parts, or sized goods.

How stores handle this today

Shopify apps, “complete the look” pixels, rules in the OMS, sales associates on the phone. Rules win when the graph is small and correct (this SKU → these three SKUs). Agents win when the cart is messy (two categories, a constraint in chat, a size already chosen). Hybrid: rules for the obvious edge, agent for the rest — agent vs workflow.

What the agent can automate

• Upsell: same family, higher capacity / warranty, still constraint-safe, cited.
• Cross-sell: compatible accessory or consumable, in stock, not already owned.
• Bundle candidate: structured set for a merchandiser to publish.

Margin and live promos are inputs, not the objective function. A high-margin unrelated SKU must not win.

What stays deterministic

• Compatibility and restricted flags.
• Inventory and price lookup (no writes).
• Promo engine eligibility for “this add-on is on a deal” — display only.
• Cart mutation policy (shopper confirm).

No price writes. If they want a cheaper add-on, that is personalized offers + HITL, not setPrice.

flowchart TD
  cart[Current cart]
  agent[Upsell agent]
  tools[Relationships inventory history margin promo]
  next[Next-best SKU]
  present[Present for confirm]
  skip[Skip or escalate]
  cart --> agent
  agent --> tools
  tools --> next
  next -->|compatible in stock| present
  next -->|already owned restricted unknown| skip

Where humans stay

Merch owns the relationship graph and published bundles. Finance owns margin floors as filters, not as a ranking god. CX handles “I already have this” when history tools are wrong. Nobody should need a human to suppress an OOS add-on if inventory is fresh.

Systems to access

Week-one allow-list:

Writes week one: none. Later addToCart with Cedar and confirm. Payment capture out of tools. Memory: this cart. Do not store a “always upsell” preference that overrides compatibility.

There is no native Shopify AgentCore connector.

Production implementation

Bedrock is the model layer, not the host. Harness or Runtime hosts the loop. Gateway + Cedar on cart writes. Browser and Code Interpreter off. Next.js is storefront or HITL, not the runtime. Strands does not provide Gateway, Identity, Policy, or microVMs.

Share catalog tools with the recommendation agent. Separate prompts so a discovery session cannot silently attach. Supervisor + hop cap if those specialists coexist (store-agents).

Gateway ~95 ms median on the CRM canary is not your cart API’s p95.

What broke — Cart-drawer prototype ranked add-ons from FBT plus a “maximize margin” instruction. A camera body received a high-margin lighting kit that used a different mount. Detection: golden “body in cart → compatible lens” failed; traces showed getFrequentlyBoughtTogether and getMargin, no getCompatibility. Recovery: compatibility required; margin tie-break only; fail evals if FBT is the only evidence. Lesson: attach rate without fit is a return rate you have not measured yet — and we are not inventing yours.

What to automate first

One accessory family with a real fits-graph. Goldens: compatible add-on, already-owned skip, OOS skip, margin-only SKU must not appear, price-cut request must not write.

If a three-line merch rule already maps body → lenses correctly, leave it. Agent the messy carts.

What to Do This Week

1. Pick one family. Confirm relationship fields exist in PIM. If not, catalog first.
2. Copy upsell-cross-sell-decision-matrix.md. Owner and Approval on auto-attach (default: no).
3. Tools: cart, relationships, inventory, history. No setPrice.
4. Harness; Policy LOG_ONLY; default-deny writes; Browser off.
5. Five goldens above. Alarm on unexpected tool mix.
6. Presentment only — no silent cart mutation.
7. Price on the AgentCore pricing calculator. Run monday-checklist.md.

Need Gateway on cart + PIM before an attach widget goes live? Contact us. Related: Amazon Bedrock consulting, Generative AI on AWS, AWS for retail / eCommerce.

What This Post Doesn’t Cover

• Session-intent recommendations — recommendation agent.
• Discount issuance — personalized offers.
• Homepage merchandising — dynamic merchandising.
• Measured attach rate, AOV, or conversion from a client engagement.
• Payment capture / x402.
• A native Shopify AgentCore connector.

FAQ

When should we NOT run an AI upsell or cross-sell agent?

Skip it when a deterministic accessory rule is already correct, when compatibility is missing from the catalog, or when leadership wants silent cart-add for margin. Missing “fits” data is a PIM ticket. Also skip Agents Classic after 30 July 2026.

What could go wrong if we rank add-ons only by frequently bought together?

Incompatible accessories, duplicates already owned, and OOS bestsellers. Co-occurrence is a tie-break after relationships, stock, and history pass.

What could go wrong if the upsell agent can write prices?

It will cut the add-on to close the attach. No setPrice, no issueDiscount. Offers are a separate HITL path.

How is this different from a product recommendation agent?

Recommendations start from session intent. Upsell starts from the current cart. Share tools, not one unbounded prompt.

Should the agent auto-attach the add-on to the cart?

Not in week one. Return a payload; the storefront confirms. Quiet attach of unrelated high-margin SKUs earns chargebacks.

Is there a native Shopify AgentCore connector for upsell widgets?

No. Expose PIM relationships and cart via OpenAPI or MCP behind Gateway.