
AI for Fish Shops: Weighing, Shelf Life and Daily Orders
A practical flow for fish shops to log weighing, shelf life and daily orders, with loss alerts and final calls left to the store manager.
Want to apply this in your company?
Book a call with Juan →In a fish shop, the setup that works today is a simple flow: every weighing, every batch with its arrival date and shelf-life limit, and every order for the day goes into one shared record. A system with rules and AI can organize that data, flag what is close to its limit or outside its usual pattern, and suggest quantities. The final call, whether to sell at a discount, prepare the product, discard it, cut the order, or contact the supplier, stays with the store manager. There is nothing futuristic about this arrangement: it is organized records, explicit rules, and human review.
Key takeaways
- Weighing, shelf life and orders need to share the same product and batch identifier, or no alert has anything to refer to.
- The store sets shelf-life limits by product type and preparation method. The AI never estimates them.
- The agent suggests and alerts. Any discard, discount or purchase needs a recorded confirmation from a person.
- Start with one counter or one product category, not the whole store.
- Every alert, suggestion and decision is logged, so the week's losses can be reviewed without relying on memory.
What AI does in a fish shop and what it does not
The useful part of AI here is handling messy information. It can read the counter clerk's handwritten notes, interpret a WhatsApp message from a customer asking for fish for the weekend, compare sales history by day of the week, and warn when a batch is approaching its limit. This work is repetitive, needs constant attention, and usually ends up scattered across a notebook, a spreadsheet and a phone.
What AI does not do is judge freshness by smell, appearance or texture, certify the sanitary condition of a piece, or replace the scale. The scale must be calibrated to metrological requirements, and sanitary inspection follows the rules of the municipal health authority. The system should reflect these limits in its design: it records what was reported and points out what does not match, without pretending to a certainty it does not have.
If the team is not yet clear on what the system can and cannot decide, start by mapping the process before choosing any tool. A good starting point is our article on data ready for superintelligence, which covers what to standardize before automating.
How to record weighing without relying on memory
Weighing is the foundation of everything else. If inbound and outbound weights are not recorded the same way, no shelf-life or loss alert will mean anything. A minimum routine usually follows these points:
- Each box or piece gets an identifier at intake, with species, batch, supplier, gross weight, date and time.
- Outbound records net weight, destination (sale, preparation, cleaning or discard) and who weighed it.
- Only weight taken on the store's own scale enters the record. Estimates by eye do not.
- The difference between inbound weight and the sum of outbound weights goes into a shrinkage report. It is never corrected silently.
Separating shrinkage from discard matters. Some loss is expected, such as cleaning waste, trimmings and ice. Another part is product that could not be sold. Mixing the two hides the problem the store needs to see.
Shelf life: a limit per product, set by the store
Shelf life depends on the product, how it was processed (whole, fillet, smoked, cooked), the storage temperature and the supplier's guidance. For that reason, the maximum limit for each category should be set by the store manager, reviewed with the supplier and aligned with local food safety rules. The system calculates the expiry date from the intake date and the registered limit, and triggers an alert with whatever lead time the team chooses.
The most common mistake is letting the AI invent limits based on generic examples. A wrong limit for one cut can lead to both unnecessary discards and sale of product that should not be sold, and neither shows up as an error in the system.
The store sets the limit. The AI only applies the limit that was registered and warns when a batch approaches it.
Illustrative scenario: a counter receives fillets early in the morning. The system lists the batches that reach their sales limit later that same day and suggests either an end-of-day offer or preparation for the store's own use. The manager reviews, chooses one of the options and records the decision with its reason. If the suggestion does not fit the day, the manager dismisses it, and that dismissal is also recorded.
Daily orders: forecasting from history, with confirmation
Orders arrive by WhatsApp, by phone and at the counter. An agent can gather these into one daily summary, compare it with the history for the same day of the week, and suggest a quantity range for purchasing or preparation. A range is more useful than a single number because it makes clear what is an estimate.
The purchase itself should not be made by the agent. It can draft the list for the supplier, but sending and confirming it stay with the responsible person. This design also prevents a misread message from becoming a real order.
There is a privacy point here. Customer names and phone numbers are personal data. Under Brazil's LGPD, the store should collect only what the flow needs, for the purpose it has stated, and review that set periodically. For suppliers, a list of products and quantities is usually enough, with no customer data. The ANPD is the official reference for guidance on this topic.
If orders come in through WhatsApp, it is also worth reading about AI for restaurants, with orders and questions on WhatsApp, which deals with similar problems of queues, confirmation and timing.
Loss alerts and the manager's final decision
An alert is only useful if the team knows what to do with it. A practical way to organize this is to separate the types of signal and the expected action for each:
- Attention alert: product within its limit but close to the sales cutoff. Action: the agent suggests an offer or preparation, and the manager decides.
- Segregation alert: temperature record outside the registered range, or missing cold-chain record. Action: the batch is set aside and the manager decides its destination.
- Discard alert: limit passed. Action: the discard is recorded with reason and responsible person, never erased.
- Weight divergence with no apparent cause. Action: refer for checking, with no automatic stock adjustment.
In every case, the last word belongs to a person. The NIST AI risk management framework emphasizes human oversight and documentation of how a system is used, and that logic applies well to a store.
Recording alerts and decisions also helps the team learn. Observability, as the OpenTelemetry concepts describe it in general terms, means keeping a record of what the system did and why, so it can be investigated later. In the store, that means knowing which alert was issued, who made the decision and what happened to the product.
How to put it into operation without promising miracles
The method we use at Juan Carlo has four steps: map, limit, pilot and hand over. Applied to a fish shop, it starts by observing who weighs, where notes are made, which spreadsheets and messages exist today, and where loss shows up most often. That map is the baseline.
Next, data and actions are limited. The agent reads what it needs, does not change stock, does not confirm purchases and does not answer customers outside written rules. Then a single flow is piloted, with full records and a success measure agreed before starting, such as the team's ability to explain the gap between inbound and outbound weight. At the end, the team decides whether to expand, adjust or stop. When the operation becomes ongoing, it can be run with managed support from Be Human (be-human.ai).
On the tool question, the choice depends on volume and on who will maintain the flow. Anyone who wants to understand how a WhatsApp agent is organized, with service windows and handoff to a human, can read about the 24-hour WhatsApp API window.
Frequently asked questions
Does AI replace the scale or food safety control? No. The scale must be calibrated, and inspection follows the rules of the municipal health authority. The system organizes records, points out divergences and warns about deadlines, but it does not replace measurement or sanitary assessment carried out by people.
Which customer data goes into the daily order and how does LGPD apply? Only the data needed to fulfill the order, such as a name and phone number for contact or delivery. The store should state the purpose, limit what it collects and review that set from time to time. For suppliers, a list of products and quantities is normally enough.
Do I need a new system to start? Not necessarily. A well-structured spreadsheet with a batch identifier, a limit per category and a decision log is enough to test the flow. Moving to a dedicated tool makes more sense when volume grows and the pilot shows which information is really missing.
How do I know whether the alerts are helping the store? The answer comes from the record. Review the alerts issued, the decisions taken and the reasons the manager dismissed some of them, every week. If alerts are always ignored for the same reason, the problem is in the rule and should be adjusted, not the team.