Superintelligence for Burger Restaurants: Order Agent With Clear Limits

Superintelligence for Burger Restaurants: Order Agent With Clear Limits

Superintelligence for burger restaurants in practice: an order agent with a live menu, peak-hour rules, pay-on-delivery handling and human handoff.

•

Want to apply this in your company?

Book a call with Juan →

An order agent for a burger restaurant is an AI assistant that takes orders over WhatsApp or your website, checks the current menu, applies your peak-hour rules, and records the payment method the customer chose, including payment made at delivery. The value for the owner lies less in the sophistication of the model and more in its limits: the agent sells only what is on the menu, accepts only what the kitchen can produce in that time slot, and passes any change outside those rules to a person. When the word superintelligence comes up in this context, the useful question is which rules the system follows, not which promise the product makes.

Key takeaways

  • The agent reads a menu that you maintain, not a loose text it has to guess from.
  • Peak-hour rules, such as an order cap per time window, preparation time and out-of-stock items, are written down and applied the same way every time.
  • Pay-on-delivery needs an explicit rule covering Pix, card machine and change, and the agent should not invent alternatives.
  • Anything outside the rules goes to a member of your team, with the order summary already organized.
  • Before serving real customers, the flow is tested with simulated orders and a success criterion agreed on before the pilot.

What an order agent does and does not do

In practice, the agent performs four tasks: identify the customer, build the order from the menu, confirm the address and payment method, and give an estimated time. It does not set prices, grant discounts, promise a time the kitchen cannot meet, or change an order already sent to production without approval. This separation may seem obvious, but it is exactly what stops a friendly conversation from turning into a promise the operation cannot keep.

It is worth writing this boundary as a short list before any configuration. If your team cannot say in a few lines what the agent may and may not do, the system will improvise in edge cases, and that is where problems appear.

One menu as the single source of truth

The most common mistake is a scattered menu: an outdated spreadsheet, an old photo on Instagram and a block of text pasted into the prompt. When that happens, the agent offers a sandwich that was taken off the menu or an add-on that no longer exists. The alternative is to treat the menu as structured data, with item name, description, price, add-on options, meat doneness where relevant, ingredient restrictions, and an availability field that the kitchen or cashier updates the same day.

Two decisions help a great deal. The first is to name one person responsible for updates, because a menu without an owner goes stale. The second is to separate what the customer can order from what the team knows, such as internal recipes and suppliers. The agent should read only the first set. To structure this base, it is worth reviewing how to organize your data before automating, because the agent will only be as reliable as the records it checks.

Peak-hour rules without improvisation

Peak hours are where automation most tries to please and most often fails. An agent that accepts orders without limit on a Friday night makes promises the kitchen cannot keep. So the peak rule must be explicit and measurable inside the system. It usually involves three elements: a maximum number of orders per interval, an estimated preparation time by item type, and a standard message for when the queue is full.

A practical way to decide what happens in each situation:

  • If the queue is below the limit, the agent accepts the order and gives the estimated time with the agreed margin.
  • If the queue has reached the limit, the agent tells the customer the restaurant is full for that slot and offers the next available interval.
  • If an item runs out during the shift, the agent removes it from the options and explains the substitution to the customer.
  • If the restaurant needs to stop taking orders for any reason, a person pauses intake from the dashboard, and the agent reports that condition without explaining the cause.

The central point is that the agent does not decide on its own whether to raise the limit. That decision belongs to the shift manager, who can see the kitchen, the stock and the available couriers.

Pay on delivery and customer data

Pay on delivery is one of the most sensitive parts of the flow, because it mixes customer service, finance and logistics. The agent should record which method the customer chose from those the restaurant accepts, for example Pix on delivery, card at the courier's machine, or cash with change, and state what the restaurant actually accepts that day. It should not promise that the card machine works in every neighborhood or suggest that the customer pay through a payment link the restaurant has not set up.

The data side requires care under Brazil's LGPD. The general rule is to collect only what is needed to deliver the order: name, phone number, address and notes. A full card number should never pass through the conversation, and the history needs a defined retention period. If the agent uses an AI model hosted outside Brazil, also check the rules on international data transfer before sending personal information to that service.

Another technical point: instructions written as free text can be bypassed by malicious messages, for example a customer who tries to make the agent ignore the menu. OWASP describes this kind of risk as prompt injection, and the safest defense is to limit the agent's tools. It should be able only to read the menu, create a draft order and trigger a handoff, without the power to change prices or cancel orders on its own.

Changes, complaints and handoff to a person

Handoff is what separates a useful agent from a risky one. Some situations should go to a person immediately: a reported food allergy, a request for a discount or coupon, a complaint about a previous order, an address change after the courier has left, a customer who asks to speak with someone, and any case where the agent has no clear rule.

Handoff only works if the person receives context. The message to the team should include the assembled order, what the customer asked for in their own words, the reason for the transfer, and what the agent has already told the customer. That way, no one has to ask the customer to repeat everything. Each transfer also creates a record, which serves as the basis for reviewing the rules at the end of the week.

Illustrative scenario: on a Friday night, a customer builds a combo, asks to swap the fries for a salad, and then asks whether they can get a discount coupon for having waited. The agent confirms the swap because that option exists on the menu, but routes the discount request to the manager with the full summary. The customer is told that someone from the restaurant will reply, with no promise of any value.

How to test before going live

Testing should happen in a separate environment, with simulated orders and no real messages sent to the kitchen. A minimum script includes these cases:

  • A complete order with an item, an add-on and a payment method the restaurant accepts.
  • An order containing an item that ran out the same day.
  • An attempt to exceed the peak-hour limit, with an offer of another time slot.
  • An order with no address or with an address outside the delivery area.
  • A discount request, a complaint and a mention of an allergy, each with the expected handoff.
  • A message sent outside opening hours.
  • An attempt by the customer to change the rules through the conversation itself.

Each case should have its expected result written down before the test. That way, the review compares what happened with what should have happened, not with someone's impression of the conversation. To learn how to build this kind of evaluation, Anthropic has published guidance on agent evaluations that focuses on criteria and repeatable cases.

How we apply the method in practice

In our method, the work follows four stages: map the current process and its baseline, limit data and actions, pilot a real flow with logs and an agreed success metric, and then hand over the operation and decide whether the next step is to expand, adjust or stop. In a burger restaurant, mapping means following one order from WhatsApp to delivery and noting where the store loses time today. Limiting means defining what the agent can read and do. Piloting means starting with one shift or one channel, not all channels at once. Handing over means the restaurant team learns the records and the handoff rules.

For readers who want to understand the difference between this kind of agent and the broader idea of superintelligence, it is worth reading what ChatGPT is and is not in this debate. In a restaurant's operation, the useful distinction is more mundane: the system follows written rules, records what it did, and hands to a human whatever it cannot resolve.

If the flow uses the WhatsApp Cloud API, you also need to respect the 24-hour window for customer service messages and understand when an outbound message requires an approved template.

Frequently asked questions

Is superintelligence the same thing as an order agent? No. Superintelligence is a concept about systems that would outperform people at virtually all cognitive tasks, and it does not describe a product a restaurant can buy today. An order agent is a bounded system with a menu, rules and limited tools, and its value comes from how clearly those limits are defined.

Can the agent accept payment on delivery? It can record the payment method the customer chose, as long as it is one the restaurant accepts that day. The charge itself happens in the restaurant's process, through the card machine or the courier's Pix, and the agent should not promise payment methods that are not set up. Any exception, such as a partial payment or a specific change request, should go to a person.

When should the agent pass the conversation to a person? Whenever there is a reported allergy, a discount request, a complaint about a previous order, an address change after the courier has left, or any situation without a clear rule. The handoff should include the order summary and the reason, so the customer does not have to repeat the information.

Do I need to replace my ordering system to use an agent? Not always. In many cases the agent connects to the system the restaurant already uses, through simple integrations with restricted permissions. Before deciding, map the current flow, identify which data the agent needs to read, and start with a pilot that can be paused without harming the operation.

Sources