
AI for Delicatessen Stores: Orders, Expiry Dates and Restocking
An AI agent can log delicatessen orders, track product expiry dates and draft restocking suggestions for the owner to approve, with decisions recorded.
Want to apply this in your company?
Book a call with Juan →An AI agent in a delicatessen can log orders received over WhatsApp, track the expiry dates of products and prepare restocking suggestions for the owner to approve. The practical gain is taking out of people's memory and loose notebooks what currently depends on someone remembering, without letting the system decide on its own what to buy or throw away. Human review stays as the control point at every step.
Key takeaways
- An order must become a structured record before any confirmation goes back to the customer.
- Expiry tracking needs rules per product and per batch, not just a generic reminder.
- A restocking suggestion is a draft for approval, with the reasoning visible.
- Customer data enters the workflow with a defined purpose, in line with Brazil's LGPD data protection law.
- Start with one product or category, measure errors, and only then expand.
Why orders, expiry dates and restocking should be designed together
These three workflows share the same stock and often the same person. An order reserves product, expiry dates affect what is still sellable on the shelf, and restocking depends on both to decide how much to buy. When each one lives in a different spreadsheet, the owner finds the problem late, usually when a customer asks for something that has already expired or that never arrived.
Illustrative scenario: a delicatessen sells sliced cheeses, cold cuts and artisan breads. A customer orders two kilos of cheese for Saturday, but the oldest batch is close to its expiry date and the supplier delayed the week's delivery. Without connected records, the shop only notices the conflict on Friday night. With a structured workflow, the agent logs the order, flags the conflict and sends the decision to the person who can resolve it, whether by swapping the batch, offering another pickup date or negotiating with the customer.
How to structure orders before automating
Before putting an agent in front of customers, the shop needs to define what a complete order looks like. Usually that means the customer's name, phone number, product and quantity, pickup date or time window, delivery method and notes such as slicing or packaging. Without these fields, the conversation becomes a chain of messages that nobody can audit later.
The agent should collect this information and save the order with a clear status, such as received, awaiting review, confirmed or declined. It should not promise availability of a product that has not been checked against stock. When an order is customized, falls outside the catalog or involves an unusually large quantity, the case goes to a team member with a ready summary, so the customer does not have to repeat everything.
It also helps to define what the customer receives automatically and what needs a human reply. Order receipt confirmations and pickup reminders are usually safe to automate. Price changes, swapping an expiring product and exception requests stay with the owner.
Expiry control: rules per batch and a named person for each alert
Expiry control fails when an alert goes to everyone and therefore belongs to no one. The simplest approach is to record each delivery with its batch, quantity and expiry date, taken from the supplier or checked on arrival, and to set an alert window per category, since yogurt and canned goods do not need the same lead time.
The process that works day to day usually follows these steps:
- Record each delivery with batch, quantity and expiry date when available.
- Set, for each category, how far ahead of expiry the alert should fire.
- Assign a named person to each alert, not a shared general inbox.
- Log every manual removal with a reason: sold, lost, donated, returned or discarded.
- Do not let the agent offer customers a discount on near-expiry products without the owner's approval.
The decision to sell, promote or discard a product belongs to the shop and must follow the health regulations that apply, including guidance from the local health authority. The agent organizes the information and reminds the right person. The person responsible for the business makes the call.
Restocking requests: suggestions with reasoning and human approval
Restocking is where AI is most likely to look precise while being wrong. A suggested number with no explanation helps nobody trust it or correct it. So the agent should always show what led to the suggestion: current stock, recent sales, open orders, remaining shelf life and the supplier's minimum order.
A simple decision guide can shape the suggestion:
- If stock is above the minimum and no sales are expected, do not suggest buying.
- If the supplier has a minimum order, show that minimum before the suggested quantity.
- If remaining shelf life is short relative to sales speed, flag the risk of loss instead of suggesting a large batch.
- If sales history is too thin, mark the suggestion as weak and request a manual review.
- In every case, the order only goes to the supplier after a person approves it.
The draft can live in a simple dashboard or a shared spreadsheet. What matters is that the approval is recorded: who approved it, when, and with what changes. That trail is what lets you understand months later why a purchase was made and fix the rule if it keeps producing the same mistake.
If the shop already uses a purchasing agent, the logic for comparing quotes and routing orders for approval can be reused. See how this applies in AI agent for procurement.
Customer data, limits and security
Orders carry personal data: names, phone numbers, delivery addresses and sometimes dietary preferences. The LGPD requires that processing has a clear purpose and is limited to what is necessary. In practice, that means collecting only what the order needs, giving access only to the people handling it, and setting how long those records are kept. Brazil's national data protection authority, the ANPD, publishes guidance worth reviewing before you design the workflow.
Conversational agents carry a specific risk as well. Customer messages can contain instructions designed to make the system ignore its rules, such as asking for an improper discount or revealing internal information. Prompt injection is widely recognized as one of the main risks in applications built on language models, and the OWASP Top 10 for LLM applications is a useful reference for the technical team. In practice, the agent should have minimum permissions: it can read stock and record orders, but it cannot change prices, approve exceptions or access other customers' data.
For a broader governance view, the NIST AI Risk Management Framework organizes AI risk around mapping, measuring, managing and governing, which fits well with the pace of a small pilot.
How to run it as a pilot
The method we apply in AI projects has four steps: map the current process and its baseline, limit data and actions, pilot a real workflow with records and an agreed success metric, and then hand it over and decide whether to expand, adjust or stop. Applied to a delicatessen, the first step is measuring how the shop works today: how many orders get lost, how often a product expires without an alert, how long a restocking decision takes.
A reasonable pilot starts with one category, for example sliced cold cuts, and a single intake channel. The metric should be simple and agreed in advance: orders logged without rework, expiry alerts handled on time, restocking suggestions approved without major changes. At the end of the period, the shop decides whether to keep the workflow, adjust the rules or stop. Organized data is a precondition for any of these steps, because an agent cannot fix a messy product catalog.
For WhatsApp conversations, the 24-hour customer service window and the rules for outbound messages define what the agent may send outside a conversation the customer started. Pickup reminders, for example, require care with message templates and prior permission.
When the agent should hand the conversation to a person
Handing off to a human is part of the design, not a failure. The agent should escalate when a product has no confirmed stock, when the customer asks for a price or timing exception, when there is a complaint about an expired product, when the customer shows frustration, or when the order involves allergens or dietary restrictions. In each case, the order summary travels with the handoff, so the person does not need to ask everything again.
Every handoff should be logged. A history showing the decision, the reason and who made it is what the shop uses to improve the workflow later. Without that record, the team either trusts the agent or distrusts it without evidence.
For an example of the same reasoning in another retail business that takes orders, see superintelligence for chocolate shops, which covers orders and expiry with similar rules.
Frequently asked questions
Can the AI agent decide on its own what the delicatessen should buy? It is not advisable. The agent can calculate a suggestion based on stock, sales, expiry and minimum order, but a responsible person should approve the purchase. That approval needs to be recorded, with the reasoning shown at the moment of decision.
How should the system handle products close to their expiry date? It should alert the person responsible for the category ahead of a window the shop itself defines, not just display the date in a report. Any promotion or disposal depends on the owner's decision and on the applicable health rules. The agent records the removal and the reason, so the information helps with the next order.
Can orders be taken over WhatsApp while complying with the LGPD? Yes, as long as the shop defines the purpose of processing, collects only the necessary data, limits access and informs the customer clearly. Conversations should have a defined retention period and a channel where the customer can ask about their data. Before going live, review the ANPD guidance and, if the tool processes data outside Brazil, check the requirements for international data transfer.
How long does it take to know whether the pilot worked? It depends on order volume and how often you restock, so there is no universal timeline. The most useful step is to define the metric and the observation period in advance, with short weekly reviews. At the end, the shop decides whether to expand, adjust the rules or close the pilot, based on the records rather than on anyone's impression of it.