
Superintelligence for Building Supply Stores: Orders and Deliveries
How an AI agent with stock rules and human confirmation can triage building supply orders, deliveries and material questions, with clear limits.
Want to apply this in your company?
Book a call with Juan →In a building supply store, what makes "superintelligence" useful in practice is a narrow AI agent, connected to your stock system and your house rules. It receives orders from contractors, reports delivery status and answers catalog questions. But it only confirms what the system confirms, and it hands off to a person anything that needs approval or technical judgment. This article shows how to design that flow with clear limits, records and a pilot with an agreed metric before any expansion.
Key takeaways
- A contractor order becomes a commitment only after the available stock is checked and the person who ordered gives explicit confirmation.
- A delivery answer needs status from your dispatch system, with a time window, a person on site and a delay notice.
- Material questions split into two kinds: technical data sheet information, which the agent answers while citing the source, and technical guidance, which goes to a person.
- Pricing, credit, discounts and product substitutions stay outside the model's reach. They come from your own rules and systems.
- A pilot with one channel, a few products and a full record of every decision shows where the flow works before you expand it.
Why the word superintelligence does not help you decide
The term shows up in generic content and creates the expectation that the system knows everything about your business. For someone running a store, the useful questions are different: which task will be automated, with what data, with which permissions, and who answers when the response is wrong. Current models are good at interpreting requests written in irregular ways, summarizing conversations and filling in fields. They do not know what is on the shelf unless a tool checks the stock.
The difference between narrow AI, built for a task with known data, and the promises of general capability is explained in Narrow AI, General AI or Superintelligence: A Scale for Managers. For a store, the practical choice is almost always the narrow one.
What the agent can do and what should stay with the team
A sensible design separates the tasks the agent executes from the tasks it only routes to people. In general, it can:
- Check availability and existing reservations in the store system.
- Draft an order with items, quantities and the job site address.
- Report the status of a delivery that is already registered, with the recorded time window.
- Answer questions about unit of sale, packaging and intended use, based on the registered technical data sheet.
- Open a case for damage, shortage or mismatch, with the order data and the photos sent by the customer.
The team keeps approving discounts, setting dates when the system does not provide one, recommending materials for structural loads, negotiating payment terms and deciding brand or specification substitutions. This boundary should be written into the agent's instructions and also enforced through technical permissions, not only stated in the prompt.
Contractor orders: from message to confirmation
A construction order usually arrives in short messages, with typos, popular product names and quantities in mixed units. The agent can organize that input, but it should not create a final order on its own. The recommended flow is this:
- Identify the customer, the job site and the person who will receive the material, and confirm the delivery address.
- List each item with quantity, unit, and whether an equivalent item is acceptable or only the requested brand.
- Check available stock after subtracting reservations from other orders. If an item is missing, report only the date the system already shows.
- Generate a draft order, with prices and terms taken from the commercial system. The model does not calculate amounts.
- Send the summary to the buyer or customer, and record the order only after explicit confirmation.
- Route to a person when the request involves credit, a discount, an item outside the catalog or a complaint. Purchasing cases that involve quotes and approval follow a similar logic, described in AI Agent for Procurement: Compare Supplier Quotes With Human Approval.
Deliveries: reliable status and notices without promises
A contractor wants to know one simple thing: when the material arrives and who will receive it. The agent can only answer well if that information exists somewhere reliable, either an integrated dispatch system or a spreadsheet kept current by the logistics team. If the information does not exist, the right answer is to say the check is in progress and pass the case to a human agent, without inventing a time window.
Two additional points need care. The first is data protection under Brazil's LGPD (Lei Geral de Proteção de Dados): addresses, phone numbers and the names of recipients are personal data, and the agent should use only what the delivery purpose requires. The second is the channel. Messages sent inside the 24-hour customer service window follow different rules from messages the company initiates, as explained in WhatsApp 24 Hour Customer Service Window: How an AI Agent Should Act. Proactive delivery notices, for example, depend on approved templates and the customer's permission.
Material questions: when to answer and when to hand off
Most counter questions repeat: what the unit of sale is, whether the product comes with a fiscal note, what the packaging is, whether it is suitable for outdoor use. These answers can be automated if they exist in a registered technical data sheet. The risk appears when the agent fills gaps with generic knowledge. A simple decision guide helps prevent that:
- If the answer is in the registered technical data sheet, the agent answers and cites the source.
- If the question involves sizing, mixing, loads or a technical standard, the agent hands it to the counter specialist or the responsible engineer.
- If the customer reports a problem with material already delivered, the agent opens a case with the data and does not suggest a cause.
- If no data sheet is registered for the product, the agent says it does not have the information and does not fill the gap with assumptions.
Before any answer goes live, test these cases with difficult questions and simulated data. The practice is described in Mock Tools for Testing AI Agents: Safe Sandbox Testing Guide.
Illustrative scenario: a site ordering mortar and steel
The following is an illustrative scenario, built only to explain the flow and not based on any real customer. A site foreman writes on WhatsApp asking for tile adhesive mortar and a few steel bars, delivered on Thursday, and refers to a brand only as the one they always use. The agent identifies the registered job site and asks which mortar brand and which steel gauge are meant, because "the one we always use" is not a field in the system. It then checks stock. The mortar is available. Part of the steel is reserved for another order. The agent reports the real situation, drafts the order with the mortar, and leaves the steel as an open item for the buyer to decide between waiting for restock or accepting another gauge. The model calculates no amounts, and the order is recorded only after the foreman confirms the summary. Every step stays in the log: the message received, the data consulted, the rule applied and the person who approved.
Records, permissions and LGPD
A store agent needs to leave a trail. For each conversation, the record should show what was asked, which data was consulted, which rule was applied and what changed in the system. That history is what allows you to review errors, adjust instructions and answer a customer who disputes a piece of information.
Permissions should follow the principle of least privilege. The agent checks stock and creates drafts, but it does not change prices, cancel orders or access employee records. For general guidance on access control, the OWASP Authorization Cheat Sheet is a good starting point for your technical team. Another risk is prompt injection: a customer may paste text into the chat trying to change the agent's behavior. OWASP also publishes a list of risks for applications built on language models, the OWASP Top 10 for LLM Applications, which is worth sharing with the people who build the integration. In practice, the pricing rule must live in the system, not in an instruction written in text.
Under the LGPD, the company needs to define the purpose of each piece of data it processes, limit collection to what is necessary and clearly tell customers when they are interacting with an automated system. The national data protection authority (ANPD) publishes guidance on its government website. The final legal decision depends on the specific case and should go through the company's responsible person or specialized legal advice.
How to apply the method in a pilot
The approach we recommend for a store follows four steps: map, limit, pilot and hand over. First, map the current process: the channels through which orders arrive, who answers each type of question, and where the most common errors happen, such as a wrong item or an incomplete address. Then limit data and actions, defining what the agent can consult and what only a person can do.
For the pilot, choose a single channel and a small set of products. Agree beforehand on an observable success metric, for example the share of orders that reach the buyer without retyping. Record everything and review the conversations with the team. At the end, the decision is objective: expand, adjust or stop. No result is guaranteed, and a pilot that misses the agreed metric is still useful information.
If the company also serves contractors after delivery, it is worth seeing how the same principle of limits applies to after-sales support in AI for Construction and Real Estate Developers: Buyer Support.
Frequently asked questions
Do I need to replace my store system to use an AI agent? Not necessarily. In many cases, the agent can read a well-maintained stock spreadsheet or a daily export from your current system. What matters is a reliable source for stock, reservations and delivery status. Without that source, the agent can only answer with uncertainty, and the best practice is to hand those cases to a person.
Can the agent close an order on its own if the customer is already registered? We do not recommend it. A registered customer may still change the quantity, the job site or the specification, and an order recorded by mistake creates rework in stock and billing. The safer model creates a draft, sends the summary and records the order only after explicit confirmation. Recurring orders can be simpler, but they still need a record.
How do I stop the agent from inventing delivery dates? The rule needs two layers. In the instructions, the agent should know that it only gives dates that come from the system. In the integration, the date field should come from a real query, not be generated by the model. When the query fails or returns nothing, the standard response is to say the information is being checked and to route the case to a person.
How long does a pilot like this take to show useful signals? It depends on message volume, data quality and how much time the team can spend reviewing conversations. What can be said with confidence is that the pilot needs enough time to cover the most common order types and the exception cases. Before deciding, the team should read the conversation records, not rely on a single isolated indicator.