AI Sales Agent for B2B eCommerce (2026)

B2B sales is not a retail chat widget with a thicker catalog. The buyer asks for their net, their pack size, their last PO, their open quote. Public PDP price is a different product.

On 17 June 2026, AgentCore Harness reached general availability (What’s New). That is the paved road for a first sales assistant: configuration, streaming invoke, default Memory. Agents Classic is in maintenance for new customers after 30 July 2026. Do not start a price-list bot on Classic InvokeAgent.

This post is a tool policy for B2B sales. It is not an anonymized win-rate study. We do not invent booked-pipeline KPIs.

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. That canary is the closest analog we publish to this mix. Platform TCO silhouette: support-style AgentCore at 50K sessions/mo ~$791/mo platform + model (decision guide). Model your AM and buyer session mix on the AgentCore pricing calculator.

Reproduce this — Copy b2b-sales-agent-tool-policy.md. Fail any eval that quotes public list price to a buyer JWT. Folder: ecommerce-ai-agents-series/.

Opinionated take: ship read tools + three JWT roles before you ship a “sales copilot” on the storefront. Trade-off: the demo is slower than a retail shopping agent. You do not leak the price book.

Retail readiness (structured catalog, public price honesty) lives in AI shopping agents. That checklist is not a substitute for getCustomerPrice. Architecture: how FactualMinds composes agents on AWS.

What a B2B sales turn actually needs

Seven jobs. If any one is unpermissioned, the model will fill the gap with retail language.

Amazon Bedrock is the model. It is not the host and not the price engine.

flowchart LR
  BuyerOrAM --> SalesAgent
  SalesAgent --> Gateway
  Gateway --> Catalog
  Gateway --> PriceList
  Gateway --> OMS
  Gateway --> CPQ
  SalesAgent --> HitlQueue

There is no native Shopify AgentCore connector. B2B often is not Shopify anyway. Attach ERP / PIM / CPQ as OpenAPI or MCP behind Gateway. Magento, Shopify Plus, custom — same contract from the agent.

Shopper vs associate vs buyer

Three principals. One Gateway. Cedar on every customer-price and history call.

1. shopper — public catalog and public list only. getCustomerPrice is DENY.
2. buyer — accountId on the token equals resource.accountId. Own history, own quotes, own net.
3. associate — assigned accounts only. Still no setCreditLimit, no capturePayment.

Prompt text that says “be careful with pricing” is not a control. Identity claims must reach Policy. The store-agents sample shows shopper vs associate on refunds — reuse the claim pattern, not the four-specialist topology, unless sales and CPQ IAM already diverge.

Week one: Harness, ≤ these reads, Browser off. Export to Strands on Runtime when hop caps or an AM specialist appear. Hybrid money movement stays in CPQ/OMS — agent vs workflow.

B2B vs public retail (do not collapse)

Retail shopping agents optimize a public catalog. B2B sales agents optimize a permissioned one. Mixing them is how a contract buyer gets a website price.

If asOf on getCustomerPrice is older than your SLA, the agent says unknown. Stale net is worse than a wait. Same honesty rule as shopping-agent readiness — different tool.

Associate console and buyer chat can share a Harness only if Policy sees the role. Prefer one entrypoint with claims over two prompts that “usually” behave.

Context: Cedar sketch; generate from your Gateway schema. Full gates: the artifact.

// Sketch — buyer may read only their customer price. LOG_ONLY then ENFORCE.
permit (
  principal,
  action == Action::"getCustomerPrice",
  resource
)
when {
  principal has role &&
  principal.role == "buyer" &&
  principal.accountId == resource.accountId
};

What broke

What broke — Week-one B2B harness that reused the retail getProduct price field because CPQ “was coming next sprint.” A buyer JWT asked for SKU net. The model answered with the public list. Detection: golden eval “own-account net ≠ PDP list” failed on the first fixture account. Fix: attach getCustomerPrice; fail closed when the tool is empty; DENY shopper on that action; never interpolate HTML. Lesson: B2B pricing is not a prompt adjective. It is a permissioned tool.

Gateway ~95 ms median tool RTT after server-side execution is a platform signal from that CRM canary. Your ERP price-list hop will dominate. Measure it. Do not quote our canary as buyer-chat latency.

Recommendations that rewrite price belong in merchandising, not in this agent — product recommendation when that post is in your tree. Here, suggest SKUs only if sellableToAccount and stock tools agree.

Reorder from history is a draft, not a silent PO. The reorder agent is the AM queue; this sales agent is the conversational front. Do not let either one capturePayment.

What to Do This Week

1. Name the three JWT roles. Prove shopper cannot call getCustomerPrice.
2. Copy b2b-sales-agent-tool-policy.md. Fill Owner on each tool.
3. Draft read OpenAPI only. Gateway Policy LOG_ONLY. Browser off. No payment tools.
4. Ten golden tasks: own-account net, other-account DENY, stale asOf, MOQ, open quote, “charge my card.”
5. Stand up HITL for quote submit above your cap — wire to the quote gates.
6. Confirm Harness (GA 17 June 2026), not Classic after 30 July 2026.
7. Price session mix on the calculator. If you need Cedar on price-list reads, contact us or start from Amazon Bedrock.

What This Post Doesn’t Cover

• RFQ approval thresholds — quote agent
• Cadence-break outreach — reorder agent; AM brief — account manager agent
• Credit and contract activation — onboarding
• A native Shopify / Magento AgentCore connector
• Payment capture, cards, AgentCore Payments / x402
• Measured B2B conversion or win-rate KPIs from a named client — we are not inventing them

FAQ

When should you NOT put a B2B sales agent on the public storefront?

Skip shopper-facing net price when the only catalog API returns list price. Keep customer price behind buyer or associate JWT. Also skip Agents Classic for net-new work after 30 July 2026.

What could go wrong if getCustomerPrice is missing and the agent falls back to public retail?

The buyer sees a website price that is not their contract. Fail closed. Do not interpolate from HTML.

Shopper vs associate vs buyer JWT — why three?

Shopper is retail or anonymous. Buyer sees only their accountId. Associate is scoped to assigned accounts. One token is how nets leak.

Should the sales agent capture payment or place the PO?

No payment capture in agent tools. Draft; commerce confirms. High-value quotes are HITL.

Is Amazon Bedrock the agent host?

No. Bedrock is the model. AgentCore hosts the loop. Gateway plus Cedar authorizes tools.

What could go wrong if Browser stays on for ERP screens?

Most turns need price and history APIs. Browser-hot pilots tracked roughly 3× Runtime-shaped spend. Keep Browser off by default.

Need a permissioned B2B sales tool map on AgentCore? Contact FactualMinds or see AWS for retail / eCommerce.