An AI agent does the reading and drafting in inventory work. It pulls stock from your system, sets it against sales and lead times, and hands you a short list: reorder this, expedite that, recount this, clear that. You decide what goes out.
IHL Group research, reported by Chain Store Age on 10 September 2025, puts inventory distortion (the cost of out-of-stocks and overstocks) at $1.73 trillion a year, 6.5% of global retail sales. I read Shopify's, Oracle NetSuite's and Cin7's own documentation, and Cin7's pricing page, on 9 October 2026. For forecasting and the wider picture, see what AI does across inventory management.
Key takeaways
- An agent reads stock, sales and lead times and drafts orders and emails. Approvals and count changes stay with you.
- Shopify retired Stocky after 31 August 2026. Purchase orders, transfers and per-location stock remain.
- The NetSuite AI Connector Service serves NetSuite data to supported AI clients over MCP; Cin7 Core's API allows 60 calls a minute.
- Data freshness depends on the route in: event triggers, polling or a nightly CSV. Alert per location.
- 21 of 63 public AI agent incidents in my database were a wrong commitment or action, or a destructive action. Keep write-backs behind an approval.
What an inventory agent actually does (and what it does not decide)
The agent does the sums and the drafts. You keep the judgment.
| Job | What the agent does | What stays with you |
|---|---|---|
| Reorder checks | Computes cover and reorder points, drafts PO lines | Approving and creating each PO |
| Reconciliation | Lists count variances with the transactions behind them | Any change to a count |
| Excess stock | Builds a disposal packet per slow mover | The disposal route |
| Supplier follow-up | Drafts chasers for unconfirmed POs, flags late deliveries | The relationship |
An agent that drafts is cheap to trust. One that submits orders and edits counts by itself is a bigger decision, and I would not make it in month one.
Which inventory systems can an agent query?
An agent can query the system you already run, if it has an API or a file it can read on a schedule. If you are choosing from scratch, a spreadsheet is enough while the catalogue is small, and Shopify's stock data is the natural start if you already sell there (Shopify Flow is free on the Basic, Grow, Advanced and Plus plans). Cin7's plans start at $349 a month, with ForesightAI as a paid add-on. If you run NetSuite, you connect through the NetSuite AI Connector Service, which Oracle says is not a paid feature.
| System | What an agent can read | What it can write | Limit that matters (9 October 2026) |
|---|---|---|---|
| Google Sheets or Excel | Any column you keep | Cells and rows, with edit access | As current as the last edit |
| Shopify | Per-location quantities by state, plus orders | available, on_hand, reserved, damaged, safety_stock and quality_control, with write_inventory; never committed | 100 cost points a second on Shopify's standard rate-limit tier; Stocky gone after 31 August 2026 |
| NetSuite | Records, saved searches and SuiteQL via the NetSuite AI Connector Service | Create and update records (ns_createRecord, ns_updateRecord) within the role's permissions; SuiteQL is read-only | Remote MCP client with OAuth 2.0; no Administrator roles |
| Cin7 Core | Its data through the Cin7 Core API; ForesightAI's forecast sits in the app at Purchases > Smart Reorder | Check Cin7 Core's API docs before planning write-backs | 60 calls a minute |
| A WMS with no API | The scheduled CSV export | Nothing; a person applies changes | As old as the last export |
Access first. Find your row above and ask what credentials an outside tool gets. Fix that before comparing agents.
Shopify, after Stocky
Shopify's help center is blunt: "The Stocky app will no longer be available after August 31st, 2026." Merchants must move to Shopify's Inventory management features, which keep purchase orders and transfers. Stocky's forecasts and suggested orders go with it, so decide where yours come from: an app, an agent or your own spreadsheet.
A PO in Shopify admin can be saved as a draft. Once it is marked Ordered, you create an inventory transfer to track and receive the shipment. Shopify documents purchase orders as an admin workflow, so in my routine the agent drafts the PO lines and a person creates the PO.
Sidekick, Shopify's AI assistant in the admin, "presents changes for your review before applying them", the right default for any agent. The Shopify restock agent guide covers the store side.
NetSuite, with the AI Connector
NetSuite already does reorder math. Its Default Lead Time is 14 days unless changed, and NetSuite uses it to auto-calculate an item's reorder point when the item has no lead time of its own. Oracle's Advanced Inventory Management FAQ gives the shape: "the reorder point is based on (demand ÷ time × lead time)".
The first useful query is dull: which items still run on the 14-day default? On a supplier with a 4-week lead time, that reorder point covers two of the four weeks, before any safety stock.
NetSuite adopted the Model Context Protocol (MCP), the plug that lets an AI client talk to a business system. My MCP adoption tracker counts 40,965 servers in the official registry as of 9 October 2026, up 11,089 (+37%) in 30 days. Through the NetSuite AI Connector Service and the free MCP Standard Tools SuiteApp, supported AI clients get 14 tools for records, reports, saved searches and SuiteQL. Two conditions apply:
- The client supports remote MCP (protocol version 2025-06-18) with OAuth 2.0. Oracle names Claude (Pro plan or higher) and ChatGPT.
- The role is not an Administrator role, and it can see what the agent needs: queries respect its permissions.
The limits: the SuiteQL tools are read-only, and no tool can call external APIs. Of the 14 tools, Oracle's list shows two that change data: ns_createRecord, which needs the REST Web Services (Full) permission, and ns_updateRecord. The role bounds both. For a first agent I would keep both off with a view-only role.
Ask Oracle, NetSuite's conversational assistant, draws context from areas including inventory and is available to existing US and Canadian customers; the NetSuite Next rollout there began on 15 July 2026 (ERP Today).
Cin7 Core, with ForesightAI
ForesightAI, built into Cin7 Core and Cin7 Omni, predicts demand "up to 24 months ahead", and its "smart replenishment automation generates purchase orders and sets safety stock levels automatically". It is sold only as an add-on, on top of Cin7's plans, which run from $349 a month (Standard) to $1,199 (Advanced), in USD before tax for new customers, as listed on 9 October 2026.
If you pay for ForesightAI, do not build a second reorder engine. Review Smart Reorder's proposals yourself, and give an outside agent the work that sits outside Cin7, such as supplier confirmations in your inbox. The limit is 60 calls a minute (a 503 error says so). If an agent made one call per product, 400 products would take just under 7 minutes.
How real-time can an inventory agent be?
Can an AI agent give you real-time inventory and better accuracy? It can get close, and it can help. An agent works from what your system last recorded, and changes reach it in three ways:
- Event triggers. Shopify Flow's "Product variant inventory quantity changed" trigger fires on every inventory change: orders, manual adjustments, apps. For one location, the agent reads its available quantity from inventoryLevels.
- Polling. The agent asks on a schedule, inside each system's limits, or runs a NetSuite saved search through the connector.
- Nightly CSV. The agent reasons about yesterday: fine for a Monday reorder run, too slow for a stock-out before lunch.
Flow's Send HTTP Request action, which hands an event to an outside agent, needs the Grow, Advanced or Plus plan. Shopify's own template sends a Slack message when a variant is out of stock at a certain location. That is the multi-warehouse rule: alert per location. A warehouse with 90 units and an empty shop reads as 90 in stock.
On accuracy, counts get truer when people count shelves. An agent shrinks the delay between a discrepancy appearing and someone seeing it, from the next stocktake to the next weekly run.
Ask every vendor one question: how old is my data when the agent sees it? That answer is your real-time.
The weekly routine, worked through
This is the routine I would run first. It is an illustrative example: a hypothetical store, with SKU numbers invented for the arithmetic, not taken from a customer.
Picture a Shopify store with two locations (a warehouse and a shop), about 400 active variants, and suppliers on 2- to 4-week lead times. A scheduled agent run fires every Monday at 07:00; the manager reviews at 09:00.
- Read stock. on_hand, available, committed and incoming per location for every variant, via the Shopify GraphQL Admin API. It pages through variants: no query may exceed 1,000 cost points, and Shopify's standard rate-limit tier restores 100 points a second.
- Read sales. 8 weeks of orders per variant, as average weekly sales.
- Do the sums. Weeks of cover = available ÷ weekly sales. Reorder point = weekly sales × (lead time + safety stock, in weeks), the same shape as NetSuite's own calculation, which adds safety stock on top of demand × lead time. These sums run on total available across both locations, because purchase orders land at the warehouse.
- Check what is coming. Incoming quantities and inbound transfers. A PO appears as a transfer once it is marked Ordered and a transfer is created; the Admin API's inventoryTransfers query lists them.
- Sort the exceptions. (a) Below the reorder point, nothing incoming. (b) Incoming lands after the projected stock-out. (c) Overstock: over 26 weeks of cover, no sale in 60 days. (d) Count drift: on_hand changed with no order, receipt or transfer to explain it. (e) Zero available at one location while the other holds stock: draft a transfer.
- Draft, don't send. PO lines per supplier in the summary, for the manager to create in Shopify admin (where a PO can be saved as a draft), plus draft supplier emails.
- Report. A Slack or email summary: counts per bucket, the ten riskiest SKUs, and the draft PO lines and transfers waiting for a yes.
Three invented SKUs show the buckets at work.
| SKU | Weekly sales | Stock | Incoming | Bucket and action |
|---|---|---|---|---|
| A | 35 | 118 available | 0 | (a) Reorder 168 units |
| B | 12 | 30 available | 48, in 3 weeks | (b) Expedite |
| C | 0 in 60 days | 410 on hand | n/a | (c) Disposal packet |
SKU A: reorder point 35 × (3 weeks lead time + 1 week safety stock) = 140. Available is 118, nothing inbound. The order-up-to level is 8 weeks of cover counted on stock today, 280 units: 280 minus 118 is 162, rounded up to cases of 24, so 168 units (7 cases).
SKU B: 30 available at 12 a week is 2.5 weeks of cover, and the 48 incoming land in 3 weeks. The shelf sits empty for half a week. A rule that ignores incoming stock would order more; the agent flags it to expedite and drafts no new PO lines.
SKU C: no sales in 60 days and 410 on hand put it in overstock. It goes to the disposal packet and its reorders stop.
At 09:00 the manager approves, edits or rejects each draft and creates the approved POs in Shopify admin. The agent sends only the approved supplier emails and logs everything. In this setup it never changes a count, creates an order or sends anything without a yes.
Reconciliation on a schedule
An agent platform can run inventory reconciliation on a schedule. The input is a cycle-count sheet or scanner export: the agent matches each line to on_hand at that location and lists every variance above your threshold, with the orders, receipts and transfers behind it.
In Shopify, an app with write_inventory can set on_hand directly with inventorySetQuantities. I would give a reconciliation agent read access only and have it draft adjustments for a person to apply.
As of 6 October 2026, my AI agent incidents database logs 63 public AI agent incidents: 12 a wrong commitment or action, 9 a destructive action. That is one in three. Count changes and POs belong behind a human approval step.
Who signs off a count change today? If the answer is "whoever notices", settle that before any agent touches your counts.
Excess and dead stock: getting to a disposal decision
An AI agent for excess inventory builds the same decision packet every time, and you pick the route. Bucket (c) finds the candidates; the agent adds age, unit cost, sell-through and any open supplier return terms.
| Route | What the agent gathers |
|---|---|
| Return to supplier | Purchase date, return terms |
| Markdown | Sell-through, margin |
| Bundle | Fast movers to pair it with |
| Transfer | Sales by location |
| Liquidate | Storage cost, lot size |
| Write off | Book value |
SKU C sold nothing at either location, which rules out a transfer, so its packet leads with return terms, sell-through and margin.
Supplier follow-ups and shipment tracking
An inventory and shipment tracking agent watches two moments: a PO nobody confirmed and a delivery that did not arrive. When a supplier has not confirmed after the days you set, the agent drafts a polite chase for your approval. When an expected date passes with nothing received, it flags which SKUs stock out first. In Shopify, receiving shows up on the transfer created for an Ordered PO, so a transfer still not received after the date your supplier gave is the signal; with several warehouses, the agent checks that per location, the same rule as stock alerts.
It is the same pattern as an invoice chasing agent: friendly nudge, firmer nudge, then a person. If purchasing is the bigger job in your week, my guide to an AI agent for procurement covers the five purchasing jobs I would hand over first.
Build, buy or hire a development shop?
Vendor-native tools (Shopify Flow and Sidekick, Ask Oracle, ForesightAI) live inside the system, so access is solved on day one. Each is built around its own system's data, which suits a business that runs on one system.
A general agent platform suits work that crosses systems: stock in Shopify, supplier threads in email, lead times in a sheet. For a plain if-this-then-that alert, the free automation platform picker compares Zapier, Make, n8n and others.
A development shop building a custom agent makes sense when your data sits in a WMS with no API, your rules are unusual, or the agent must pass your own security review. You pay for the build and every change after it.
Where to start
Pick one task. My default is the weekly reorder check: the arithmetic is simple, and a wrong draft costs little when a person reviews it.
The free Can AI do this task checker takes a task you do by hand and shows how much of it an agent could take on, how often it comes around and what your time on it costs a month. It needs no sign-up.
Gravity is in public beta. Its agents are already built: you describe the task in plain words, with no workflow to build. The catalogue is still short, so do not assume it covers your exact task yet. Explore agents if you want in.
Frequently asked questions
What is the best inventory database an AI agent can query for stock levels, reorder points and supplier updates?
Start with the one you already run, if it has an API. If you are choosing, pick a spreadsheet for a small catalogue, Shopify if you already sell there, Cin7 on plans from $349 a month (ForesightAI costs extra), or NetSuite with the NetSuite AI Connector Service, which Oracle says is not a paid feature.
Can AI agents replace inventory management software?
No. The software is the system of record for counts, locations, orders and costs. An agent reads that record and drafts the next step.
Can an AI agent reconcile inventory on a schedule?
Yes. It compares counts with the system, lists variances above your threshold and drafts adjustments. Keep write-backs behind a person's approval.
How do AI agents handle multiple warehouses?
They check each location separately, because total stock hides an empty shop behind a full warehouse. Shopify reports quantities per location, and Shopify Flow's own template sends a Slack message when a variant is out of stock at one location.
What does an inventory AI agent cost?
The vendor costs come first: Shopify Flow is free on the Basic, Grow, Advanced and Plus plans, the NetSuite AI Connector Service is not a paid feature (your AI client may need a paid plan), and ForesightAI is a paid add-on on Cin7 plans from $349 a month. Gravity's plans: Basic is free (one agent, no card), Pro is $5 a month during the beta (up to three agents, usage included), and Max is $20 a month. The catalogue is still short, so do not assume it covers your task yet.
Does it work with Excel?
Yes, if the agent can reach the file. One row per SKU and location, with on-hand, reorder point, supplier and lead time, runs the same weekly routine.
Sources
- IHL Group research on inventory distortion, Chain Store Age, published 10 September 2025.
- Stocky, Shopify Help Center, checked 9 October 2026.
- Purchase orders, Shopify Help Center, checked 9 October 2026.
- Manage inventory quantities and states, Shopify.dev, checked 9 October 2026.
- GraphQL Admin API rate limits, Shopify.dev, checked 9 October 2026.
- inventoryTransfers query, Shopify GraphQL Admin API reference, checked 9 October 2026.
- Shopify Flow, Shopify Help Center, checked 9 October 2026.
- Product variant inventory quantity changed, Shopify Help Center, checked 9 October 2026.
- Sidekick, Shopify Help Center, checked 9 October 2026.
- Inventory management preferences, Oracle NetSuite Help Center, checked 9 October 2026.
- Advanced Inventory Management FAQ (reorder point and preferred stock level formulas), Oracle NetSuite Help Center, checked 9 October 2026.
- NetSuite AI Connector Service documentation, article 4160616848 and article 3200541651, Oracle NetSuite Help Center, checked 9 October 2026.
- Available Tools in the MCP Standard Tools SuiteApp, Oracle NetSuite Help Center, checked 9 October 2026.
- NetSuite Next and the Ask Oracle rollout, ERP Today, checked 9 October 2026.
- API Status Codes, Cin7 Core Help Center, checked 9 October 2026.
- ForesightAI, Cin7, checked 9 October 2026.
- ForesightAI in Cin7 Core, Cin7 Core Help Center, checked 9 October 2026.
- Cin7 pricing, Cin7, checked 9 October 2026.