Square Just Made ChatGPT a Restaurant Ordering Channel

By Lazar Milicevic · Published July 2, 2026 · 9 min read

Your restaurant's POS handles Uber Eats, DoorDash, phone orders, and walk-ins. Now a customer asks ChatGPT "find me a decent burrito near Union Square that can deliver in 20 minutes," and the assistant needs somewhere to send that order. If your storefront isn't reachable by an agent, it doesn't exist in that conversation. Square's new ChatGPT app and Claude plugin close that gap for restaurants running on Square — with no integration work required on the merchant side, and reportedly waived processing fees on those agent-driven transactions.

This isn't a demo. It's the first mainstream example of what people have been calling "agentic commerce" landing in an actual vertical (food) with an actual payment rail (Square). If you build for SMBs, sell to SMBs, or operate one, the shape of the channel is worth understanding now — because the pattern is going to repeat outside restaurants very quickly.

What Square actually shipped

Square launched a ChatGPT app and a Claude integration that let a consumer discover a Square-powered restaurant, build an order, and pay — inside the AI assistant — without the merchant doing any technical work. The restaurant sees the order land in their normal Square dashboard, same as any other channel. Square is subsidizing the transaction economics for agent-originated orders, meaning the traditional card-processing fee isn't being passed to the merchant on these specific flows (verify current terms on Square's merchant page — subsidies change).

The pieces that matter:

• Zero merchant setup. If you're already on Square for Restaurants, your menu, hours, item modifiers, and payment credentials are already in the graph. Square publishes them into the agent channel on your behalf.
• Order lands in the existing POS. Kitchen printer, tickets, KDS — nothing changes. It's another inbound order.
• The consumer never leaves the chat. No handoff to a browser, no "click here to complete your order on the merchant site." ChatGPT and Claude collect the order, confirm, and pay.
• Fees are reportedly waived on these agent-originated transactions during launch.

That last point is the strategic tell. Square isn't trying to earn interchange on these orders — they're trying to become the default rail when an AI assistant needs to transact on behalf of a user. That's a land grab, and it will be temporary.

Why this is the first real agentic commerce channel

We've had years of "AI will book your travel" demos that never left the sandbox. What makes this different is boring plumbing, done right:

1. A structured merchant catalog already exists. Every Square restaurant has a normalized menu, modifiers, prices, and tax rules in Square's own database. The assistant doesn't have to scrape a PDF menu or guess.
2. Payment is solved inside the assistant. ChatGPT and Claude both have identity and a way to hold or invoke payment credentials. Square provides the accepting side. No redirect, no cart abandonment.
3. Fulfillment is a physical location the customer already trusts. Delivery/pickup are established motions.
4. The merchant surface is unchanged. No new dashboard, no new SLAs.

Compare to the "AI books your flight" pitch: three or four of those layers are still missing (airline catalog access, agent-authenticated payment, ticket issuance, exception handling). Food ordering with an existing POS provider is the shortest path from "AI can talk about it" to "AI can transact it."

What agent commerce means for the SMB stack

If you operate a small business and one of your discovery channels becomes "a user asks an assistant a question," a few of the assumptions in your marketing and ops stack break at once.

Two rows deserve a closer look. Discovery stops being about your homepage. It becomes about whether the assistant can find, describe, and trust your business — which means structured data, clean feeds, and being on a rail the assistant already integrates. Attribution goes dark unless the platform gives you referrer signals. Expect early agent channels to look like Google search circa 2013: you'll see orders, not the queries that caused them.

The MCP layer underneath

Both OpenAI (via ChatGPT apps) and Anthropic (via Claude's tool ecosystem, and increasingly through the Model Context Protocol) are converging on the same architectural pattern: the assistant is the client, and every capability — search a menu, place an order, refund a charge — is an MCP-style tool exposed by a server that the platform trusts.

The Model Context Protocol is Anthropic's open spec for how a model talks to external tools. A minimal MCP tool descriptor for a "place order" action looks something like this:

{
  "name": "place_order",
  "description": "Place a food order at a specific Square restaurant location.",
  "inputSchema": {
    "type": "object",
    "required": ["location_id", "items", "fulfillment"],
    "properties": {
      "location_id": { "type": "string" },
      "items": {
        "type": "array",
        "items": {
          "type": "object",
          "required": ["catalog_object_id", "quantity"],
          "properties": {
            "catalog_object_id": { "type": "string" },
            "quantity": { "type": "integer", "minimum": 1 },
            "modifiers": {
              "type": "array",
              "items": { "type": "string" }
            }
          }
        }
      },
      "fulfillment": {
        "type": "object",
        "required": ["type"],
        "properties": {
          "type": { "enum": ["PICKUP", "DELIVERY"] },
          "pickup_time": { "type": "string", "format": "date-time" }
        }
      }
    }
  }
}

Nothing exotic. But the policy around that tool is the interesting part: who's allowed to call it, what confirmation the assistant must show the user before invoking it, what the refund path looks like, and how the platform authenticates the merchant behind location_id. Square is doing that work for restaurants once, and every ChatGPT/Claude user gets it.

For anyone building SMB software, the takeaway is direct: your product's future distribution is a well-scoped MCP server, hosted somewhere a major assistant trusts. If you can describe your actions cleanly, validate inputs strictly, and handle idempotency, you're closer to being an agent channel than you think.

What restaurants should do this week

You don't need to build anything. You need to make sure your Square catalog is actually accurate, because now it's a customer-facing surface every time an assistant queries it. A few concrete checks:

• Modifiers. Every "no onion", "extra cheese", "make it a combo" needs to exist as a proper modifier in Square, not as a note in the item description. Agents read structured fields; they don't reliably interpret prose.
• Hours and pause states. If you 86 an item at 8pm on a slow Tuesday, mark it out-of-stock in the POS. An assistant confidently selling a customer a dish you can't make is a refund and a bad review.
• Prep times per fulfillment type. Pickup ETAs surface in the assistant. Wrong ETAs will bury you in "where's my order?" chats you never see.
• Tax and fees. Anything Square isn't computing (a local bag fee, a delivery surcharge) needs to be modeled properly or the customer sees a different total than they agreed to in the chat.
• Photos and short item descriptions. Assistants will use these to answer "what's good?" — a blank field is a lost recommendation.

A one-hour audit of your Square catalog will pay for itself the first time ChatGPT sends someone your way instead of the restaurant three doors down whose menu is a JPEG.

What this looks like beyond restaurants

Restaurants are the wedge. The pattern extends anywhere you already have (1) a structured catalog, (2) a payment rail the merchant trusts, and (3) simple fulfillment. That's a lot of SMB verticals:

• Local services. Booking a haircut, a dog groomer, a mobile mechanic through an assistant, backed by an existing scheduling platform (think Booksy, Vagaro, Acuity).
• Retail. Shopify merchants already have structured catalogs and Shop Pay for identity — the same pattern applies with slightly harder inventory constraints.
• B2B reorder. Any distributor whose customers place the same 20 SKUs every month via email. That's an assistant workflow with a receipt at the end.
• Home services quotes. Not full purchase, but "get me three quotes for gutter cleaning" is a lead-gen flow an agent can run against ServiceTitan/Housecall Pro-style systems.

The common thread: whoever owns the merchant's data and payment rail today has the shortest path to owning the agent channel tomorrow. Square just demonstrated the play. Expect Shopify, Toast, Stripe, Booksy, and the rest to answer in kind. If you're an SMB SaaS founder without a plan for this, your customers' next channel is being built by your competitors.

The uncomfortable questions

A few things this launch doesn't answer, and you should think about before assuming agent commerce is a pure win:

Who owns the customer? If ChatGPT drove the order, ChatGPT has the relationship. You have a receipt. Loyalty programs, retargeting, review requests — all of that is complicated when the customer never touched your storefront.

What happens when the agent is wrong? An assistant misreads a modifier and orders a peanut dish for a customer with an allergy. Who's liable — the platform, the merchant, or the user who confirmed the order? The consumer protection frameworks for agent-mediated commerce are not yet settled. Log everything.

What happens when fees come back? Square is subsidizing this now. When the promotional pricing ends, agent-originated orders will carry a fee. Whether that fee is competitive with DoorDash's 15-30% commission or closer to standard card processing will decide whether restaurants embrace the channel or block it.

Ranking and merchandising. When a user asks "best pizza near me," some restaurant gets recommended first. That ranking is a channel, and channels get monetized. Nobody has said what the paid-placement model inside these assistants will look like, but it's coming.

None of this is a reason not to be on the channel. It's a reason to be on it early, watch the data, and not build a business that assumes today's economics.

How BizFlowAI approaches this

We build MCP integrations for SMBs — the plumbing that lets your business be a first-class citizen inside an agent conversation instead of a footnote. For clients on Square, Shopify, HubSpot, or a custom stack, that usually means a scoped MCP server exposing the two or three actions their customers actually take (place an order, book a slot, request a quote), with idempotency, auth, and audit logging done properly the first time.

If you're an operator asking "when should we be reachable inside ChatGPT and Claude?", the honest answer is: as soon as the tools your customers use let you, and one week before your competitors figure it out. If that's a conversation worth having, book a discovery call and we'll walk through what a working agent channel looks like for your specific stack.

Work with BizFlowAI

If you'd rather have this built for you, that's what we do: production AI automation for solo founders and small teams — agents, integrations, and document pipelines that actually ship.

Book a free discovery call — 30 minutes, we map the highest-ROI automation in your workflow. No pitch deck, just engineering.

More guides like this on the BizFlowAI blog.