What connecting your ERP to Claude or ChatGPT actually means

Most companies already use ChatGPT or Claude. The problem is that the assistant knows nothing about the operation.

It cannot see current inventory. It does not know which SKU belongs to which supplier. It cannot create a sales order or update a price without somebody copying the answer back into the ERP.

Connecting the assistant to the ERP changes that. But only if the connection gives it real tools, clear permissions, and a controlled way to write data.

Otherwise it is still just a chat window.

A real example: upload the catalog and create the items

We worked through this with a customer whose inventory arrived in spreadsheets and whose orders were sometimes handwritten.

They uploaded an order photo to ChatGPT. ChatGPT checked the products against Convexity, matched the items that already existed, and created three missing items. The order became a sales order, with corrections made in plain language along the way. The next step they discussed was uploading updated Excel files to refresh stock balances.

The important part was not extracting text from a photo. Plenty of tools can do that.

The important part was that the assistant could check the ERP, use the existing records, create the missing ones, and save the result back into the operating system.

What you can do through Claude or ChatGPT

Once the ERP exposes controlled tools to the assistant, useful operations include:

• Upload a supplier catalog and create inventory items that do not exist.
• Match spreadsheet rows to existing SKUs and flag uncertain matches.
• Update stock balances from a new inventory file.
• Prepare bulk price changes by supplier, category, margin target, or uploaded price list.
• Check availability across variants and locations.
• Turn an email, PDF, spreadsheet, or photo into a draft sales order.
• Create a purchase-order draft for understocked items.
• Move an approved order into picking and fulfillment.
• Build an operational report from current ERP data.

These are not all equally risky. Looking up inventory is not the same as changing 5,000 prices. Creating a draft purchase order is not the same as sending it to a supplier.

The system should treat them differently.

Bulk changes need a preview

Take price updates.

You should be able to upload a vendor price list and say: match these products, update their costs, and keep a 35% gross margin. The assistant can prepare the changes quickly.

It should not silently publish them.

A good workflow shows what matched, what did not, the old value, the proposed value, and the rule that produced it. A person approves the batch. Then the ERP applies it and records who changed what.

The same rule applies to catalog imports. If the assistant is unsure whether two item names refer to the same SKU, it should ask. Guessing creates duplicate items, and duplicate items are expensive to clean up later.

Claude and ChatGPT are the interface, not the database

The ERP remains the system of record. Claude or ChatGPT is a way to operate it.

The connection is typically made through a standard such as MCP, which lets the assistant call specific tools. Those tools can be narrow: find an item, list stock by warehouse, create a draft order, update approved prices.

That boundary matters. The assistant should not receive unrestricted database access. It should only see the company data and actions allowed for that user.

For higher-risk actions, the flow should be simple:

1. Read the relevant ERP records.
2. Prepare the proposed change.
3. Validate item matches, quantities, prices, and required fields.
4. Ask for approval when the action crosses the company’s risk threshold.
5. Execute the change through an ERP tool.
6. Keep an audit trail and return any exceptions.

Start with one ugly workflow

Do not start with “we want AI in our ERP.” That does not mean anything.

Start with a workflow people hate doing. A 600-row catalog import. A weekly vendor price update. Orders copied from email into QuickBooks. Purchase recommendations that somebody turns into POs by hand.

Give the assistant the minimum tools needed to complete that workflow. Keep the first version draft-first. Measure the errors and the time saved. Then expand the permissions.

We have already connected Claude and ChatGPT to live ERP workflows for item creation, inventory updates, order entry, and purchasing tests. If you want to see what this would look like against your own process, book a working session. Bring the spreadsheet or order your team currently handles by hand.