
Superintelligence for Bars: Aligning Events, Drink Demand and Restocking
How to connect event bookings, expected drink consumption and bar stock with an AI agent that asks a manager for approval above a set spending limit.
Want to apply this in your company?
Book a call with Juan →Superintelligence at a bar, in practice, is an AI agent that cross-checks three sources: the event booking, the expected consumption based on past tab history, and the stock you actually counted. It works out what is missing, prepares a draft order, and moves forward only when the purchase stays within a spending limit the manager has written down. The decision to spend stays with a person. The AI makes that decision faster, easier to trace, and less dependent on someone's memory.
Key takeaways
- The useful AI for a bar is narrow. It cross-checks bookings, tabs and stock, and it does not replace the manager.
- The approval limit must be written down, with the item types and the total amount set by the manager personally.
- Confirmed events should feed the consumption forecast before you buy, not after the stock has already run out.
- Booking data should be reduced to the minimum needed, with attention to LGPD (Brazil's data protection law).
- Start with one event type and one drink family, log every suggestion and every approval, and only then expand.
What the term superintelligence means at a bar
The term appears often in marketing material, and it almost always promises more than real use requires. For someone running a bar, the useful question is different: what does the system do with the booking, the tab and the stock count? None of these tasks requires general intelligence. They require correct data, clear rules and a responsible person who owns the result.
It helps to treat this as a narrow agent working inside a defined process. The difference between narrow AI, general AI and superintelligence keeps you from buying expectations in place of a working solution. That distinction is explained in Narrow AI, General AI or Superintelligence: A Scale for Managers.
Which data the alignment needs
Before any forecast, the bar needs to know where each piece of information comes from. Without that, the agent will produce well-formatted numbers with no solid basis. The minimum data usually includes:
- The event booking: date, time, event type, estimated number of guests, the package sold and the person responsible for the group.
- The tab history from similar events: which drinks sold, in what proportion, and how much was left over.
- The actual stock count taken on the day, not just the balance shown in the system without verification.
- The recipe cards for each drink: how much of each ingredient goes into each serve and each preparation.
- The usual suppliers, with delivery lead time and minimum order quantity.
- The manager's rules: what can be reordered automatically as a draft and what needs a case-by-case decision.
Recipe cards are the piece most often missing. Without them, the system knows how many bottles left the bar but cannot translate a drink into quantities of ingredients. The path to building them is covered in Superintelligence for Restaurants: Recipe Cards and Purchasing Data.
Step by step to align bookings, consumption and restocking
The goal of the first cycle is to prove that the flow works on a real case. A simple sequence:
- Map the current process: who receives the booking, who decides the purchase, where the stock is recorded, and how much time passes between confirming the event and placing the supplier order.
- Choose a recurring event type, such as birthdays or nights with live music, so there is enough history to compare.
- Define the source of truth for each data point: bookings in the system the bar already uses, consumption from the tab, and stock from the weekly count or the count just before the event.
- Write the forecast in plain terms, for example: average consumption per guest at similar events, multiplied by the confirmed headcount and adjusted for what was left over or short last time.
- Build the item list with minimum stock levels, the usual supplier and the lead time.
- Record the manager's approval limit, by purchase type and by total amount.
- Have the agent generate only a draft order, with the reason for each item, and send nothing to the supplier.
- Keep a record of every suggestion, approval, rejection and gap between what was forecast and what was consumed.
- After the event, review the rule, not just the quantity ordered.
Where the manager's approval limit sits
The limit is a business rule, not a feeling the model has. The agent should operate within restricted permissions: read bookings and stock, calculate, draft orders and request approval. Sending an order, changing a price or cancelling a purchase stays outside its reach. This separation follows the principle of limiting actions and data that appears in authorization guidance from OWASP. A practical way to organize the decision:
- Restocking of fast-moving items, within minimum stock and with the usual supplier: the agent prepares the order and the manager approves with a quick confirmation.
- Extra purchase because of a confirmed event: the agent calculates and the manager approves before anything is sent.
- Changing supplier or adding an item that is not on the menu: always a human decision.
- Changing or cancelling an order already sent: a human decision, with the reason recorded.
- Price changes, promotions or discounts: outside the agent's scope.
- Doubts about a booking, such as an event with no payment confirmed or a headcount far from its history: the agent pauses and asks for review.
The recommended approach is to keep the flow deterministic where the rule is clear and use AI where judgment is needed, such as estimating consumption from an irregular pattern. The distinction between fixed-rule workflows and agents that decide their own steps is discussed in Anthropic's guidance on building effective agents, and it is worth applying here without overdoing it.
Illustrative scenario: Saturday event with a large group
Illustrative scenario: a bar receives a birthday booking for Saturday, with a drink package billed by quota and most guests already confirmed. On Thursday, the agent reads the booking, compares it with the consumption of a similar event the previous month, and notices that sparkling wine is below minimum stock. It drafts an order with the reason for each item and sees that the total exceeds the automatic approval limit the manager set, so it forwards the draft with an explanation.
The manager adjusts the quantity of one of the drinks, because they know that group tends to prefer another, and approves. After the event, the record shows leftover beer and a shortage of sparkling wine. The correction that goes into the rule is not simply ordering more sparkling wine. It is adjusting the drink mix for that package type. That is the real gain: the lesson is written down and does not depend on who happened to be on shift.
Privacy and LGPD for bookings
Bookings often include a name, a phone number and sometimes notes about dietary restrictions or allergies. For buying drinks, almost none of that is needed. The agent should receive only what the decision requires: event type, time and estimated headcount. The guest list should not go to the supplier or to any external tool without a clear reason and an adequate legal basis.
If the bar uses WhatsApp to confirm bookings or remind guests about the event, the platform's rules and people's consent must be respected. The national data protection authority (ANPD) publishes the main guidance on LGPD in Brazil, and checking it is the right starting point before deciding what the system may store and for how long. Logs also need access control, because they record the team's actions and decisions.
How to apply Juan Carlo's method
The work follows four stages: map, limit, pilot and hand over. At the bar, mapping means describing the real flow from booking to tab to purchase and measuring a baseline, such as how long the team takes to close an order and how often stock ran out during an event. Limiting means giving the agent read-only access to booking and stock data, with drafts in place of sending.
Piloting means choosing one event type and a success criterion agreed in advance, such as the gap between what was forecast and what was consumed, with every approval recorded. Handing over means giving the flow to the manager with instructions, logs and a clear decision: expand, adjust or stop. When the operation needs ongoing support, maintenance can be handled by Be Human. None of this depends on promised results. The pilot exists to show whether the rule works at this particular bar.
Frequently asked questions
Is superintelligence for bars available as a ready-made product? Not in the sense of an autonomous machine that runs the stock by itself. What exists are agents and spreadsheets with rules that cross-check data and ask for approval. Be wary of any material that promises total control without mentioning data, permissions and records.
Can the agent place the drink order on its own? Technically, an integration can send orders, but the recommended starting point is for the agent to only draft the order. Sending should be automated only after a pilot period, with value limits and item types clearly defined and approved by the manager.
What is the minimum data needed to start? You need the booking with event type and estimated headcount, the tab history from similar events, a reliable stock count and the list of suppliers with lead times. Recipe cards improve accuracy a lot, but they can be added in a second stage.
Do I need an integrated point-of-sale system to begin? Not necessarily. A simple export of tabs and a stock spreadsheet are enough to test the logic. Connecting to the bar's system makes sense after the pilot shows that the rule works and that the gain justifies the effort of connecting the tools.