AI Inventory Agent: What Should We Reorder Today? (2026)

Monday, the aisle looks fine until someone asks what should we reorder today — before ads spend on a SKU that will miss, before a buyer duplicates an open PO, before excess sits another season.

The job. Stop the buyer rebuilding “what needs a PO” from a spreadsheet, last month’s velocity, and a promo calendar in someone’s head.

This week. Rank SKUs as reorder, wait, stockout risk, or excess from on-hand, reserved, inbound, lead time, and open POs. Draft the brief. Do not send the PO.

A person still signs. Every purchase order. Tokens that are not buyer or admin cannot create one.

Skip it when a reorder-point formula already posts the right qty to the preferred vendor and nobody is joining promo calendars or open POs by hand. Also skip it if you cannot name the buyer.

This is series part 22 of the 2026 field guide. It is the inventory row in 15 automations and task 2 in back-office automation. The PO itself is purchase assistance. Excess is the dead-stock post. Demo fixture SKU-TEE-BLU-M at 5 days of cover is a sample, not a client. Days of cover and SKUs below are fixtures. We are not publishing a turns-of-cover KPI from a FactualMinds store.

Copy the risk brief — Open inventory-reorder-risk-brief.md. Fill Owner and Approval. Ship gates: monday-checklist.md. Folder: ecommerce-ai-agents-series/.

Our take: the agent produces a risk brief, not a purchase order. That feels slower than auto-buy. The alternative is an unattended purchaser with a chat window.

Automate first

If days-of-cover vs a fixed reorder point already posts the right qty to a preferred vendor, keep the rule. Agents pay off when you must join velocity, lead time, open POs, seasonality, and “are we still advertising this.” A formula that ignores inbound is how you double-buy. A chatbot that ignores the formula is how you invent qty.

Four classes, one brief

From the artifact:

flowchart LR
  InventorySignals[Inventory Signals]
  AIAgent[AI Agent]
  RiskBrief[Risk Brief]
  HumanApproval[Human Approval]
  POOrWait[PO or Wait]
  InventorySignals --> AIAgent
  AIAgent --> RiskBrief
  RiskBrief --> HumanApproval
  HumanApproval --> POOrWait

Signals: on-hand vs reserved, 7/28-day velocity, lead times, open POs, seasonality or promo calendar if a tool exists, safety-stock policy. Missing signal → unknown, not a guess.

Week-one evals: (1) flash-sale day stays off class A unless 28-day velocity agrees, (2) inbound inside lead time forces B, (3) advertised SKU with slipping ASN is C even if cover looks fine on paper.

Production shape

Bedrock is the model. Harness or Runtime is the host. Gateway + Cedar is the choke point. Strands is an agent framework, not infra. Next.js is a HITL queue if you want one — not the loop. Browser off.

There is no native Shopify AgentCore connector. WMS, ERP, Shopify Admin — your OpenAPI.

Assumes AgentCore Gateway, Policy on writes. Context: Harness or Runtime; Python 3.12+ only if you also compile store-agents stubs.

# Excerpt — inventory Gateway target (not a live spec).
paths:
  /inventory/position:
    get: { operationId: getInventoryPosition }
  /inventory/velocity:
    get: { operationId: getSkuVelocity }
  /purchase-orders/open:
    get: { operationId: getOpenPurchaseOrders }
  /purchase-orders:
    post: { operationId: createPurchaseOrder }

createPurchaseOrder starts HITL even if Cedar would ALLOW.

Absolute latency will be WMS/ERP, not Gateway. Scheduled inventory jobs are cheap vs shopper chat; still keep Browser off so the bill does not look like a support agent. Model cost on the AgentCore pricing calculator.

What broke

What broke — Week 1 of the sample loop used 1-day velocity after a flash promo. Class A fired on a SKU that had 3 weeks of cover at 28-day rate and an ASN two days out. Detection: Gateway trace showed getSkuVelocity without getOpenPurchaseOrders; Policy LOG_ONLY would have ALLOW’d a send the Cedar file did not name. Fix: require 7- and 28-day windows; fail the brief if open POs were not read; DENY send until buyer JWT. Lesson: a spike is not a run rate. Auto-PO is how you buy last week’s campaign twice.

A second failure: reserved units counted as available. The brief said wait. Open orders ate the last units. Fix: position payload is on-hand, reserved, inbound — three numbers.

What to Do This Week

1. Clone inventory-reorder-risk-brief.md. Name the buyer.
2. If a reorder-point formula already works, keep it. Add an agent only where humans join three systems.
3. Draft read OpenAPI: position, velocity, open POs, lead time. No PO create tool yet.
4. Stand up Harness or Runtime; Gateway; Policy LOG_ONLY. Browser off.
5. Prove DENY: createPurchaseOrder on a shopper token and on an associate without buyer role.
6. Eval: flash-sale day must not become class A without 28-day context; inbound must force class B.
7. Model cost on the AgentCore pricing calculator. For Gateway Policy on purchasing, contact us or start from Amazon Bedrock and retail / eCommerce on AWS.

Full ship list: monday-checklist.md.

If You Only Do One Thing

Put Cedar on createPurchaseOrder and require open-PO + 28-day velocity on every class A row. The brief is the product. The send button is not.

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. Do not start inventory automation on Classic (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). Model your mix on the AgentCore pricing calculator. The ~180 ms → ~95 ms figure is the CRM platform canary. Use it to place tools on Gateway, not to SLA a warehouse API.

What This Post Doesn’t Cover

• Purchase order vendor selection, pack size, and finance caps.
• Excess / dead stock markdowns and liquidation.
• Statistical demand models owning the unit forecast — demand forecasting agent.
• AgentCore Browser over a vendor portal.
• Measured stockout or inventory-turn KPIs from a FactualMinds client — this sample does not invent them. We have no published agent case study that claims fill rate improved by a percentage.
• A native Shopify inventory app (it does not exist).

FAQ

When should we NOT use an AI agent for inventory reorder?

Skip it when a reorder-point formula already posts a correct qty to the preferred vendor and nobody is reading promo calendars or open POs by hand. Also skip it if you cannot name a buyer who must approve every PO. An agent without an approver is an unattended purchaser.

What could go wrong if the inventory agent auto-sends purchase orders?

A flash-sale day becomes the run rate, reserved units look like a stockout, and you stack a PO on inbound that lands next week. Draft the PO. A human with buying authority sends it. Cedar should DENY createPurchaseOrder for every token that is not buyer or admin.

What could go wrong if reserved inventory is treated as available?

You wait on a SKU that is already promised to open orders, then stock out while ads keep spending. Position tools must return on-hand, reserved, and inbound separately. The brief fails eval if it uses a single “stock” number.

Should this run on AgentCore Harness or Runtime?

Harness fits a scheduled risk job with a short read-tool list. Use Runtime plus Strands when inventory, demand, and purchasing are separate specialists with hop caps. Bedrock is the model. Harness or Runtime is the host. Agents Classic is the wrong net-new host after July 30, 2026.

Is there a native Shopify AgentCore inventory connector?

No. Attach Gateway OpenAPI or MCP to your WMS, ERP, or Shopify Admin inventory APIs. Shopify, Magento, and custom stacks look the same at the Gateway boundary.