AI Reorder Agent for B2B Customers (2026)

Monday, a plant that buys gloves every ~30 days and goes quiet on day 38 is not a cart-abandonment email. It is a hypothesis: stock at their dock, your ATP, a price-list change, a credit hold, a quality ticket, a competitor — or they are on shutdown. The agent investigates. Sales decides whether anyone speaks.

The job. Detect a missed purchase cadence, investigate blockers, and recommend an internal sales action.

This week. Produce an AM queue row with evidence. contact_customer stays false. Do not place the reorder. Do not capture payment.

A person still signs. Whether anyone emails or calls the account. Credit hold, quality disputes, and strategic pauses stay human. If customer price cannot be matched to the account, the note says so — it does not quote the public catalog as net.

Skip it when you have no account-scoped order history, no assigned AM, or the “signal” is a retail abandoned-cart timer. Also skip auto-email until Cedar DENYs customer send by default. Skip it when stock for the usual SKUs is only a weekly CSV and the agent would be inventing ATP.

This is series part 29 of the 2026 field guide. Cadence windows in the artifact are fixtures. We are not publishing a reorder-rate lift from a FactualMinds client.

Copy the playbook — Walk b2b-reorder-attention-playbook.md. Every row needs evidence_tool and contact_customer: false until a human flips it. Folder: ecommerce-ai-agents-series/.

Our take: assist AMs. Do not spam customers. Trade-off: you will “miss” some easy emails. You will not torch accounts that paused on purpose.

The AM’s daily five: account manager agent. Own-account net and history: sales agent. Warehouse replenishment is a different post — replenishment agent. Retail-style winback is not this job — retention agent.

Detect, then investigate, then recommend

flowchart TD
  cadence[Cadence anomaly]
  inv[Investigate]
  rec[Recommended sales action]
  human[AM / BDR]
  cadence --> inv
  inv --> rec
  rec --> human
  inv --> stock[ATP]
  inv --> price[Customer price]
  inv --> acct[Account / credit]
  inv --> tix[Support tickets]
  inv --> quotes[Open quotes]

Detection is a job or a query: expected window vs last order date. The model does not “notice” a customer. A tool returns daysSinceLastOrder.

Investigation is mandatory before call_am:

• Inventory / lead time on the usual SKUs
• Permissioned customer price (not public retail)
• Account status / credit hold
• Open or expired quotes
• Open support or quality tickets

If a blocker exists, the action is credit_block or do_not_contact — not a friendly nudge.

Cadence is a statistic you own. Median days between last four POs, a contract-stated interval, or an AM-entered expected window. The model does not invent “they usually order monthly” from prose. If history is too thin, the row is watch or needs_data.

Do not merge this with warehouse replenishment. DC min/max and vendor POs are replenishment / inventory agents. A missed customer PO is not a license to raise a your PO. Do not merge with consumer winback either — retention is a different consent and channel problem.

Public retail price in the AM note is a fail. The note either cites getCustomerPrice for that accountId or says price unknown. Same permission as the sales agent.

Sales action, not a blast

There is no send_nudge_email in week one. If marketing insists later, that is a new write tool with Cedar, opt-in, and HITL. Default DENY.

Associate JWT only for the cadence list. A buyer token must not see other plants’ missed windows. Browser off. Bedrock is the model. Gateway plus Cedar is the boundary. No native Shopify connector — attach OMS history yourself.

Context: output row; demo account; not a client.

{
  "accountId": "ACC-DEMO-104",
  "signal": "missed_cadence",
  "windowDays": 30,
  "daysSinceLastOrder": 38,
  "evidence_tool": "getOrderHistory",
  "recommended_action": "call_am",
  "contact_customer": false
}

Rank by revenue-at-risk you already store, not model confidence. Cap the list so it can feed the AM dashboard.

What broke

What broke — A cadence job that called the ESP because “reorder is just cart recovery for wholesale.” Day 38 emails went to accounts on credit hold and to a buyer in a quality dispute. Detection: helpdesk tags vs send log; AMs forwarded angry replies. Fix: delete the send tool; require investigation checklist; contact_customer defaults false; credit hold → credit_block. Lesson: a 30-day window is a sales queue, not a lifecycle campaign.

Scheduled Harness is enough: one invoke, JSON rows, no shopper channel. If you later fan out to quote-revive and credit specialists in one turn, export to Strands. Until then, do not staff a supervisor for a cadence list. Browser stays off. A portal scrape of the ERP “customer 360” screen is not getOrderHistory.

The AM who owns the account owns the send. Territory rules stay in CRM assignment, not in a prompt that says “be reasonable about who you contact.” If assignment is empty, the row goes to a coverage queue — not to a shared ESP list.

What to Do This Week

1. Pick one cadence definition (e.g. median days between POs, last 4 orders). Write it down. Do not let the model invent it.
2. Copy b2b-reorder-attention-playbook.md.
3. Read tools only: history, ATP, customer price, tickets, quotes. Policy LOG_ONLY.
4. Goldens: quiet-but-healthy, credit hold, quality ticket, shopper JWT, public-price-in-note.
5. Feed the AM five — do not stand up a second Slack channel.
6. Harness scheduled invoke. Not Classic after July 30, 2026. Browser off.
7. Architecture review: contact · retail / eCommerce.

For your technical lead

On June 17, 2026, AgentCore Harness reached general availability (What’s New). Agents Classic is in maintenance for new customers after July 30, 2026. A scheduled InvokeHarness with read tools is enough for a morning exception list — do not put that job on Classic.

First-party signals we reuse (not client churn KPIs) — Gateway server-side tools cut median tool RTT ~180 ms → ~95 ms on a B2B CRM assistant (12 tools, ~8k turns/day) — Gateway post. This reorder mix is the same shape: many small account reads. Platform ~$791/mo at 50K sessions is a silhouette (decision guide). Schedule-once-per-day is cheap; a chatty “nudge copilot” is not — calculator. Gateway ~95 ms is still the CRM-canary platform number. Your OMS history query is the wall clock that matters.

What This Post Doesn’t Cover

• Permissioned storefront sales UX — sales agent
• Full AM metric card — account manager agent
• DC replenishment / vendor POs
• Consumer winback sequences
• Measured reorder-rate lifts from a FactualMinds client — we are not inventing them
• Payment capture

FAQ

When should you NOT run a B2B reorder agent?

Skip it when you have no account-scoped order history, no assigned AM, or the “signal” is a retail abandoned-cart timer. Also skip auto-email until Cedar DENYs customer send by default. Net-new loops use Harness or Runtime, not Agents Classic after July 30, 2026.

What could go wrong if the agent emails the customer on a missed 30-day window?

You ping accounts on credit hold, quality disputes, or strategic pauses. You also train buyers to ignore you. Week one output is an internal recommended_action with contact_customer false.

What could go wrong if reorder logic uses public retail price in the AM note?

The AM forwards a website price that is not the contract. Same leak as the sales agent. Notes must cite getCustomerPrice or say price unknown.

Is this the same as a warehouse replenishment agent?

No. Replenishment is your DC/vendor PO problem. This post is customer purchase cadence for a B2B account. Do not merge those tool sets.

Should a buyer JWT see other accounts that “usually order by now”?

No. Cadence jobs are associate-only. Buyer tools stay on own history and own catalog — see the sales-agent policy.

Can the agent place the reorder?

Not in week one. Recommend. The buyer or AM confirms in the commerce system. No payment capture. Browser off.

Need a cadence-to-AM queue without an ESP blast? Contact FactualMinds or Generative AI on AWS.