Letting wholesale buyers order through their own AI assistant

Picture one of your restaurant customers on a Tuesday night, planning Thursday's prep. Instead of opening a laptop and logging into anything, she asks the AI assistant she already uses all day, Claude, ChatGPT, or Gemini: "What's on the fresh sheet this week?" The assistant reads your current sheet and tells her. She follows up: "What's my price on 5 cases of romaine?" and gets her number, the one on her account, with her negotiated override applied, not the list price. Then: "Order that plus 2 cases of salmon for Thursday." The assistant confirms the lines and the total, places the order on her net terms, and the confirmation email lands in her inbox exactly as if she had ordered on your storefront.

That is the whole pitch. Your wholesale customers can browse, check pricing, and order through the assistant they already talk to, and nothing about your operation has to change to make it possible.

No new app, one connection

The reason this works without a rollout project is that there is nothing to install. Your customer adds one connection URL to their assistant (a custom connector, in most clients a couple of fields in a settings screen), then signs in with the same email and password they use on your store. No separate account, no API key, no second password to forget.

From then on, the assistant acts as that person. A buyer sees exactly what they would see logged into the storefront: your published fresh sheet, their catalog, their own effective prices, their orders, their invoices and balance, their standing orders. Nothing more. A chef cannot see another restaurant's pricing, and a buyer at a different distributor who connects to their own supplier sees only that supplier's data. The permission rules are the same ones that govern the web portal, applied to a new front door, which is why the answers always match what the same login sees on the site. If you are still weighing whether to run wholesale ordering online at all, the portal fundamentals are covered in how wholesale B2B ordering works for food distributors; the assistant is a second way in, not a separate system.

For the buyer, the practical difference is that ordering becomes a conversation they can have anywhere. "What did I order last week?" "Add 2 cases of eggs to my weekly standing order." "Skip next week's delivery." "What do I owe you?" Each of those is a question their assistant can now answer or act on, with the same rules and limits as the storefront.

Your staff get the same door

The buyer side is the headline, but your team connects the same way with their staff logins, and the back office gains a fast lane:

Draft next week's fresh sheet from a rough note about what came off the truck. The assistant creates it as a draft; nothing publishes until a person says so.
Re-price a hundred products at once. Bulk updates are preview-then-confirm: the assistant shows the full resolved change list first and applies it only after an explicit confirmation. There is no bulk delete at all, so a misunderstood instruction cannot erase your catalog.
Import a spreadsheet of customers (or products) by pasting the data into the conversation and letting the assistant run the import.
Pull sales answers on demand: top customers this month, what a specific account ordered last week, revenue over a date range.
Enter carrier and tracking numbers on fulfilled orders without opening the admin panel.

Every action respects the staff member's role exactly as the admin panel does. A buyer login cannot invoke any staff capability, and a removed staff member loses access on their next call.

Where the assistant stops, on purpose

Handing an AI a door into your business only works if the boundaries are hard, so here they are, plainly:

The assistant never touches payment details. If a buyer pays by card, or wants to settle an open invoice, the assistant hands back a link into your normal web checkout. Card numbers never pass through the conversation.
Totals are always computed by the store from live catalog prices and the buyer's effective pricing. A price said in chat, by the buyer or by the assistant, is ignored. An assistant that hallucinates a discount cannot create one.
Orders placed this way are real orders. Same per-store numbering, same snapshot pricing, same confirmation email, same visibility in your admin, same audit trail. There is no separate "AI orders" pile to reconcile; your pick list does not know or care how the order arrived.

Invalid requests fail safely too: an order naming an unknown or unavailable product is rejected with the offending line called out, and no partial order is created.

Which assistants work

The connection uses MCP, an open standard most major AI clients now speak, so your customers bring whatever assistant they already have:

Claude works on all plans, including the free one. Buyers add the connection as a custom connector in Settings on the web.
ChatGPT needs a paid plan with developer mode enabled.
Gemini connects through the Gemini CLI and Gemini Enterprise; the consumer Gemini app does not support custom connectors.
Other MCP-compatible tools connect the same way with the same URL.

In practice that means the easiest thing to tell a curious customer is: "If you use Claude, it works today on the free plan; if you use ChatGPT, you need a paid plan with developer mode turned on."

What to do with this

You do not need to campaign for it. Turn it on, put the connection instructions in front of your accounts, and let the two or three chefs who already run their ordering lists through an assistant find it. They become the reference customers, the same way your first portal adopters did.

AI-assistant ordering is available on Minori Midori's Enterprise plan; the bring-your-own-AI page covers the full tool list for buyers and staff. If you want to watch an order travel from a chat message to your pick list, book a demo and ask to see it live.