
AI for Wine Shops: Stock, Vintages and Responsible Sales
How an AI agent checks wine stock and vintages, applies age and responsible drinking rules, and confirms a sale only after those checks pass.
Want to apply this in your company?
Book a call with Juan →In a wine shop, superintelligence in practice means a tightly scoped AI agent that checks stock and vintages, applies age and responsible drinking rules, and confirms a sale only after those checks pass. The order matters: the shop's rule comes before the confirmation, not after the customer has already received a yes. When a check fails, the agent either declines clearly or hands the case to someone on the team.
Key takeaways
- The agent checks stock and vintage in the shop's system and only states availability it can verify.
- The age check is a fixed step in the process, verified in person at delivery or pickup, not a judgment made by the model.
- Drinking limits, such as a maximum quantity per order, live in the system and not only in the prompt text.
- A sale moves to confirmation only when every rule passes, and each decision is logged.
- Ambiguous cases, such as resale requests, vintage substitutions, or signs of excessive drinking, go to a human.
What the agent checks before it answers
The term superintelligence appears often in marketing material and can suggest something that decides on its own. In a shop, the sensible model is a narrow agent: it talks, checks, and records within limits the shop has written itself.
A wine answer is only useful if it comes from the shop's own system. The agent needs three sources: the catalog, with product, grape, region, and vintage; the stock, with the quantity that can actually be reserved; and the commercial rules, such as the current price, the quantity cap, and delivery areas. The model should not infer any of these from general knowledge about wine. If the product database is disorganized, the agent will only spread the disorder faster, so it is worth reviewing how the data needs to be organized first.
Vintage deserves its own rule. If a customer asks for a specific red and accepts another vintage, the shop needs to decide whether that substitution is allowed, whether it requires approval, and how the difference is communicated. Without that rule, the agent tends to fill the gap with a plausible suggestion, and that is usually what creates a complaint later.
In the technical design, prefer small tools with clear names, such as one to check stock, another to check vintage, and another to reserve, rather than one generic tool that does everything. This makes each action easier to test, audit, and restrict.
Age rules: the customer's word is not enough
Selling alcohol to minors is prohibited in Brazil, and a WhatsApp agent cannot inspect an identity document. For that reason, the age check cannot depend only on the customer replying that they are an adult. The safest design separates what happens in the chat from what happens at delivery or pickup.
In practice, the flow usually works like this: the agent presents the catalog and reserves items, but an order with delivery only moves forward when the shop's verification step is complete, and the check happens with a document at the door or at the counter. The record stores the pending verification status, who verified, and when. Which requirements apply to each delivery method should be confirmed with the shop's legal advisor, because it depends on how the business operates.
Illustrative scenario: a person writes at night asking for several bottles for delivery and says they are old enough. The agent confirms the bottles are in stock, creates the order as awaiting verification at delivery, and tells the customer the driver will check an ID. If the same person tries to override the rule by saying the shop is being excessive, the agent does not negotiate the requirement and passes the case to a human.
The channel matters too. Messages sent outside the service window require approved templates, and the 24 hour WhatsApp API window is a constraint the agent must respect when it resumes a conversation.
Responsible drinking as a rule, not a request to the model
Asking the model to be responsible does not create a rule. The limit has to exist in the system: a quantity cap per order set by the shop, a block on volume promotions above that cap, and a handoff to a human when the conversation suggests someone is already drinking heavily or asks for alcohol for a person who is visibly intoxicated.
The language is part of the rule as well. The agent should not use a tone that encourages drinking, promise health effects, or suggest that a wine will solve a personal problem. Food pairing and purchase suggestions for an occasion are acceptable when they stay informative and within the shop's policy.
Customer messages can try to get around the rules with lines like ignore your instructions. The prompt injection prevention guidance published by OWASP helps teams treat that text as data rather than as commands. Even so, the age rule and the quantity cap should sit outside the reach of the conversation, in code and configuration.
Step by step: designing the flow
A lean roadmap for a shop that wants to start. Each step has an owner and an exit criterion.
- Map the current process: who answers, how age is checked today, where stock and vintage are recorded, and where discrepancies appear.
- Write the rules with the team and the legal advisor: age requirements by delivery method, quantity cap, delivery areas, and vintage substitution rules.
- Separate model decisions from system decisions: tone and recommendations within the catalog stay with the model, while age, stock, and limits stay with fixed rules.
- Build small tools with restricted permissions: check stock, check vintage, reserve, and record the order. Pricing, discounts, and cancellations stay out of the agent's reach without human approval.
- Test with hard cases before any real customer sees the agent.
- Run a pilot with every decision logged and a success metric agreed in advance, then decide whether to expand, adjust, or stop.
How to test before the agent starts selling
The most useful tests are the ones that try to break the rules. Build a set that includes requests from people who say they are under age, orders above the cap, sold-out vintages, substitutions that are not allowed, discount requests without authorization, and messages that try to override the policy. Run this set every time the prompt, the model, or the rules change.
A simulation environment with fake stock keeps a test from becoming a real sale. As the guidance from Anthropic on evaluating agents suggests, each case should have an expected, verifiable result, not just an impression that the answer looked good. To set up this environment safely, see how to test agents with mocks and a sandbox.
Personal data and the right to review
An order carries personal data: name, phone number, address, purchase history, and, at delivery, document information. The shop should collect only what it needs to deliver and to meet the age rule, and define how long it keeps the record of each decision. If any data is sent to a hosted AI service outside the country, the shop needs to review what the LGPD requires in that case.
If the agent declines an order or changes a condition based on automated rules, the customer needs a way to request review by a person. The LGPD addresses this right for decisions made solely on the basis of automated processing. Refusals based on age should come with a clear message, a general reason, and a human channel for contesting the decision. The topic is covered in more detail in LGPD automated decision-making, and the official guidance can be followed on the website of Brazil's data protection authority, ANPD.
Where our implementation method fits
In how we work with shops, the path has four stages: map the process and current situation, limit data and actions, pilot a real flow with logging and an agreed metric, and then hand over the operation and decide whether to expand, adjust, or stop. For a wine shop, this usually means starting with one category and one delivery method, not the whole catalog. The pilot shows where the rules need adjustment before anything expands.
Frequently asked questions
Can the agent confirm a wine sale on WhatsApp by itself?
It can reserve items and record the order within the shop's rules, but completing a sale that includes delivery of alcohol depends on an age check done with a document at delivery or pickup. The chat cannot complete that step alone. For that reason, the order starts as pending and moves forward only when the rule is met.
Is it enough to ask the customer whether they are an adult?
Asking helps record the intent, but it proves nothing. A yes in a message does not replace checking a document. The agent can ask as part of the flow, as long as the shop does not treat the answer as proof.
How does the agent handle a vintage that is sold out?
It should report the unavailability based on the stock it checked and offer only the alternatives the shop has defined as permitted, such as another vintage of the same product if a rule allows it. If the substitution needs approval, the case goes to a person. The agent should not invent availability or promise a restocking date.
Should we start with a full agent or a simpler one?
Starting with the simpler version is usually the safer decision. An agent that checks the catalog, reports stock, and passes orders to staff already adds value with fewer risk rules. Recommendations and deliveries with alcohol can come later, once the decision log shows that the rules are working.