
Superintelligence for Craft Supply Stores: Kits, Supplies and Workshops
How an AI agent can recommend kits and supplies and book workshops for a craft store, without promising stock the store has not confirmed.
Want to apply this in your company?
Book a call with Juan →In a craft supply store, an AI agent can answer questions about kits, supplies and workshops, suggest material combinations, and route the customer to a workshop signup. It does this by checking the information the store has already recorded, not by knowing what is on the shelf. When stock is not confirmed, the agent should say so clearly, offer an alternative, or hand the case to a person.
The term superintelligence appears often in market content, but in a store's daily operation, what works today is a limited system with tools, rules and records. It is worth separating those two ideas before deciding what goes into production.
Key takeaways
- The agent recommends kits and supplies based on a data source the store controls, not on assumptions.
- Every item carries an explicit status, such as available, reserved, on request or unavailable, and the agent only states what is marked as available.
- Workshops are booked only within seats, time slots and materials the team has released.
- Questions about restocking time, customization or exceptions go to a human agent, with the conversation history attached.
- The starting point is a pilot with a real flow, a record of every response, and an agreed metric defined before launch.
What the agent does and what it does not do
The agent's role is to reduce the queue of repetitive questions and give the customer a fast answer about what the store offers. It can explain what a embroidery kit contains, compare two types of fabric paint, state how long a pottery workshop lasts, and show the dates that still have seats. This work is predictable and can be described in rules.
The agent should not decide on its own what is in stock, should not negotiate discounts, should not promise delivery on a specific date without confirmation, and should not invent a technique or material that is not in the catalog. When a question falls outside that scope, the correct response is to acknowledge the limit and hand it off.
The line between what the agent can state and what depends on confirmation must be written into the design of the flow. In well-built implementations, this permissions list is as visible as the service script itself. Anthropic's material on agents stresses keeping flows simple and predictable, with well-defined tools, rather than giving a model broad freedom. For a store, this means few tools: look up the catalog, check the calendar, record interest, and transfer to a human.
How to recommend kits and supplies without promising stock
The most delicate question is availability. Many craft stores keep inventory in a spreadsheet, a point-of-sale system, an online store and a purchasing notebook, and these records do not always match. Before any agent is involved, the store has to choose one primary source and accept that it may be out of date at times.
A good practice is to work with statuses rather than exact quantities. An item can appear as available for immediate sale, available to order, reserved for a workshop, or unconfirmed since the last update. The agent answers based on that status and states the date of the last update when it matters.
Decide also what happens with composite kits. A canvas painting kit may depend on five items, and if one is missing, the set cannot be sold as a complete kit. The rule needs to be clear: the agent can suggest the items separately, offer the kit for later restocking, or route the customer to the team.
The store, not the agent, decides the status of each item. This responsibility belongs in the process, with a named person responsible for updating the critical items.
How to book workshops with real seats
Workshop booking looks simple, but it involves room capacity, quantity of materials, duration, instructor availability and cancellation rules. If the agent offers a seat that does not exist, the problem appears on the day of the class, with a frustrated customer and a team improvising. For this reason, the calendar the agent uses should be the same one the team uses, or be synchronized with it in a controlled way.
Before switching on booking, answer these questions:
- What is the maximum capacity of each class, and how many seats are already reserved?
- Are materials included in the workshop price, or does the customer need to buy something beforehand?
- Is there a minimum signup deadline, so the store can prepare the kits?
- What happens if the store cancels the workshop, and how is the customer informed?
- Who approves exceptions, such as a reduced class, a date change or a student with no prior experience?
If any of these answers is not defined, the agent should forward the request instead of confirming the seat.
Illustrative scenario: the question about the ceramics kit
Illustrative scenario: a customer asks on WhatsApp whether the cold clay kit suits children and whether the Saturday workshop still has seats. The agent checks the catalog, finds the kit description, and states that it is intended for children over twelve, without claiming anything beyond what is registered. It then checks the calendar, sees that the Saturday workshop has seats, but that the materials kit is marked as on request.
The correct response is not to confirm the enrollment. The agent explains that the seat appears to be available, that the materials must be confirmed by the team, and that a staff member will reply shortly. It records the interest, the intended date and the age stated, and passes the conversation along with that summary. If the team confirms, the enrollment is completed by a person or by a flow with approval.
This sequence avoids the promise that would cost the store the most: selling a seat whose material does not exist.
How to roll it out with a controlled pilot
In how we work, rollout follows four steps: map, limit, pilot and hand over. Mapping means understanding the current service, with the most common questions, the places where the team loses time, and the mistakes that already happen. Limiting means defining what the agent can look up, what it can record, and what it must never state. Piloting means running a real flow for a set period, with records and an agreed metric defined before the start. Handing over means transferring the operation to the team and deciding whether the flow will be expanded, adjusted or stopped.
A first plan can follow these steps:
- List the five most frequent questions about kits, supplies and workshops, with the answers the team already gives.
- Choose one source of truth for inventory and the calendar, even if imperfect, and record who updates it.
- Define the possible statuses for each item and the exact phrase the agent uses for each one.
- Build a set of test conversations in an isolated environment, including cases where stock is uncertain.
- Activate the pilot for one channel or time window, with daily review of conversations and a record of every handoff.
- At the end of the period, compare the agreed metric with the previous service and decide the next step.
To test the agent's behavior before exposing real customers, it helps to use simulated data and a separate environment. The article on mock tools and sandbox testing covers this step in detail.
Data, LGPD and WhatsApp messages
The agent will handle personal data: names, phone numbers, workshop interests, and sometimes photos of work or information about enrolled children. This requires care with purpose, data minimization and access. The store should collect only what the service needs, inform the customer how that information is used, and limit who can see the history. Brazil's National Data Protection Authority (ANPD) publishes guidance on the LGPD, and it is worth checking its official page before setting the data policy.
On WhatsApp, there are also channel rules. Replies within the customer service window are treated differently from outbound messages, such as workshop reminders. For reminders, the customer's permission is required and approved message templates must be used. See how this works in the 24-hour customer service window on WhatsApp API and in outbound WhatsApp messages with opt-in.
OWASP maintains a list of the most common risks in applications built on language models, such as instruction injection through customer messages. In a store, this can appear as someone trying to make the agent reveal internal prices or confirm a discount that does not exist. The defense starts with limiting tools and with the rule that the agent does not make commercial decisions on its own.
Finally, organized data makes a real difference. A catalog with standardized names, updated statuses and clear descriptions reduces errors more than any prompt adjustment. The topic is covered in data ready for superintelligence, and the difference between what marketing calls superintelligence and what a tool actually does is explained in is ChatGPT superintelligence?.
Frequently asked questions
Can the agent say with confidence whether a kit is available? It can only state what is recorded in the source the store has defined, with the date of the last update when that matters. If the status is uncertain, the correct response is to state the uncertainty and route the conversation to the team. No agent should confirm availability that a person or a reliable inventory system has not confirmed.
Do I need an integrated inventory system to start? Not necessarily. Many stores begin with a well-maintained spreadsheet, simple statuses and one person responsible for updating it at set times. Integration with the sales system can come later, once the pilot shows the flow works and which data is actually missing. Starting small reduces the risk of automating information that is not yet reliable.
Can the agent book workshops directly on WhatsApp? Yes, as long as the calendar it uses is the official one, the seats are up to date, and the cancellation and materials rules are defined. For workshops with special materials or classes with exceptions, the safer option is for the agent to record the request and a staff member to confirm it. The final confirmation should be clear to the customer, with the date, time and what they need to bring or buy.
How long does it take to know whether it is worth expanding the agent to other stores or channels? The decision depends on the metric agreed at the start, the volume of pilot conversations and the quality of the handoffs. The most useful step is to review, at the end of the period, which questions the agent resolved well, which required the team, and which led to errors. With those records in hand, it becomes simpler to decide whether the flow should be expanded, adjusted or stopped.