AI for Building Supply Stores: Quotes, Stock and Delivery

AI for Building Supply Stores: Quotes, Stock and Delivery

For a building supply store, AI works best as an assistant that reads the customer's materials list, checks stock and delivery status in the store's own systems, and builds a draft quote. Quantities, prices and lead times reach the customer only after calculation rules have run and someone at the counter or in sales has checked them. The gain is answering fast without letting a mistake in bags, meters or rebar gauge turn into a loss. That is the practical core of AI for building supply stores.

Key takeaways

• AI can make sense of a messy customer list (text, photo, voice note, PDF), but prices and stock must come from the store's system, never from the model's memory.
• Quoting from a list has three main risks: an ambiguous product, a wrong unit of measure and a quantity calculation error. Each one needs a rule and a checkpoint.
• Stock and delivery lookups are read-only. Start there, before allowing any action such as reserving, changing an order or promising a date.
• Human review is not a brake on the project. It is what lets you go live safely and measure what the AI gets right.
• A pilot with one flow, a record of every decision and one agreed metric is worth more than a generic automatic attendant.

Where AI really helps in a building supply store

Customer service in a building supply store has a particular trait: customers rarely know the exact name of what they need. They send a list typed on a phone, a photo of a notebook page, a voice note from the site foreman or a project PDF. Many items look alike, and the difference between them is technical: rebar gauge, type of mortar, grout color, pipe diameter.

That is where AI helps. It is good at interpreting free text, separating item from quantity, recognizing that "bag of cement" and "50 kg cement" probably mean the same product, and asking clarifying questions. It is poor as a source of truth. Prices, balances and lead times change every day and belong to the store's management system.

So the division of roles is simple:

• The AI reads, organizes, asks questions and drafts.
• The store's system (ERP, price table, stock control) supplies the data.
• Deterministic rules do the math and apply limits.
• A person checks anything with financial impact before it is sent.

If the store already uses Bling, Omie, Tiny or an in-house system, the technical path usually goes through an API. It is worth reading first about integrating AI with Bling, Omie or Tiny, including the risks of giving the AI write access.

How to turn a materials list into a quote

The flow below is a starting point. Each step has a clear owner, which makes it easier to record and audit later.

• Receive the list. The customer sends text, a photo, a voice note or a PDF on WhatsApp. When a file arrives, field extraction should follow rules and review, as described in extracting data from PDFs with AI.
• Structure the items. The AI splits each line into requested product, quantity, unit and note, without yet choosing the final product.
• Match to the catalog. Each item is searched in the store's product records. If more than one option is plausible, the AI does not pick on its own: it asks the customer or flags the item as ambiguous.
• Convert units. Bags, kilos, meters, square meters, thousands, bars and boxes are handled by a conversion table in the product records, not by the model.
• Calculate. Totals, permitted discounts, freight and payment terms come from store rules, run by code or by the ERP itself.
• Generate the draft. The AI writes the message with items, quantities and values, highlighting anything that was assumed.
• Review. A person checks the risk points and approves, corrects or sends the draft back for a question to the customer.
• Send and record. The quote goes out, and the store keeps the original list, the draft, the corrections and who approved it.

The critical step is the calculation. Language models can get arithmetic wrong, and material quantities involve yield, waste and rounding up to whole packages. If the store offers estimates (for example, how much mortar for a given area), the formula should come from the store or the manufacturer, be applied by code, and be presented to the customer as an estimate they can adjust.

What the AI can say about stock

A stock check looks simple, but it is where customer trust is most fragile. If the AI says a product is available and it is not on the shelf, the cost is a rescheduled delivery or a stalled job site.

A few rules keep the answer honest:

• The stock answer must come from a system query made at the time of the conversation, not from fixed text or an earlier reply.
• Distinguish physical balance, reserved balance and available balance, if the system separates those numbers. If the store does not trust the balance for certain items, those items are flagged for human confirmation.
• For items sold loose or by volume, such as sand, gravel and cut wire, the AI should say that availability depends on confirmation.
• If the query fails or times out, the correct answer is to admit it could not verify and call a person. Inventing a plausible answer is the worst outcome.
• Avoid promise language. "Shown as available in the system right now" is different from "guaranteed for you."

The same logic applies to substitutions. If the requested product is out of stock, the AI may suggest registered equivalents, but only those the store has marked as acceptable substitutes, and always pointing out the difference. Swapping one brand for another may be acceptable. Swapping a strength class or a technical standard is not a decision for an automatic assistant.

Delivery status: answers that avoid phone calls and empty promises

Questions like "has my order left yet?" and "what time will it arrive?" are repetitive and depend on data the store already has: invoice issued, picking, loading, route, driver, proof of receipt. This is a good first use case because it is read-only and carries low financial risk.

A good design separates what the AI answers alone from what it passes on:

• Answers alone: current order stage, the delivery window already set by logistics, the address on file (partially masked), and fixed guidance such as who may receive the goods and what to check when unloading.
• Passes to a person: requests to change the address or date, complaints about damage or missing items, invoice discrepancies, any request for an exception.
• Never promises: an exact time when logistics has not set one, or any cancellation or refund.

If the store wants to notify customers proactively, for example when an order leaves for delivery, that falls under another set of rules: contact consent, approved templates and frequency. The article on WhatsApp outbound messages, templates and opt-in covers that care in detail.

There is also the data protection side. Under Brazil's LGPD, address, phone number and individual or company tax IDs are personal data or tied to people. The AI should query only what it needs, confirm the customer's identity before showing order details (for example, by asking for the order number or another detail only they would know), and not keep data longer than the service requires. The ANPD, Brazil's data protection authority, is the official reference for following guidance.

Human review: what to check before the quote goes out

Review does not need to cover everything. It should concentrate where an error costs the most. A practical rule is to require review whenever any of these conditions appears:

• An item flagged as ambiguous or with more than one candidate in the catalog.
• A converted unit (square meters to boxes, kilos to bags, linear meters to bars).
• A quantity estimated by the AI or by a formula rather than given by the customer.
• A product with low stock, doubtful stock or made to order.
• A discount, special condition or freight outside the standard table.
• A high-value order, based on a limit the store defines.
• A new customer, an address outside the delivery area or a job site that is hard to reach.

The reviewer should see the customer's original list, the draft and the doubt flags on the same screen, so they can decide in a few seconds. If the reviewer has to open three systems, review becomes a bottleneck and people start approving without looking.

The AI drafts the quote, the store signs it.

That sentence sums up the accountability rule. The customer receives a price in the store's name, and the store answers for it. The technical controls, such as output validation and action limits, are covered in guardrails for AI agents.

Illustrative scenario

Imagine, hypothetically, that a customer sends a WhatsApp voice note that is transcribed as: "I need cement, sand, some bricks and that thin rebar for a small slab." The AI structures the items but notices three gaps: there is no quantity for the bricks, "thin rebar" could be more than one gauge, and "small slab" gives no area. Instead of filling in with guesses, it asks for the gauge, the brick type and the dimensions. With the answers, the system converts units and calculates. The draft reaches the counter with two items flagged: the estimated rebar quantity and freight outside the standard table. The attendant adjusts the waste allowance, approves it, and the quote goes out. The record keeps what was requested, what the AI proposed and what the attendant changed.

How to apply the method: map, limit, pilot, hand over

The method we use in AI projects for companies fits this case well, without requiring a major overhaul.

Map: survey how a quote is produced today. Who receives the list, where they look up price and stock, which errors have already happened, which delivery questions keep coming back. This map becomes the starting point, and the current average response time (measured by the store itself) becomes the baseline.

Limit: define what the AI reads, what it never changes and when it calls a person. At the start, read-only access to catalog, price, balance and delivery status. Least-privilege authorization principles, such as those in the OWASP authorization guidance, help decide what each tool may do.

Pilot: choose one real flow, for example quoting a simple list over WhatsApp or checking delivery status, with a record of every decision and a metric agreed in advance, such as time to draft and number of corrections made by the reviewer. To structure that measurement, see how to measure an AI automation pilot.

Hand over: at the end of the pilot, give the team the documentation, the rules and the tracking dashboard, and decide from the records whether the flow should expand, be adjusted or stop. Stopping is also a valid result when the data shows the flow does not pay off.

On architecture, what makes sense for catalog queries is to retrieve information from the store's system at the moment of the question and hand it to the model, a concept known as retrieval-augmented generation. AWS has a plain-language explanation of it. Anthropic, for its part, recommends starting with simple, predictable solutions before more autonomous agents, which fits the principle of limiting before expanding.

Frequently asked questions

Can AI calculate the material quantity for a job on its own? It can help organize the data and explain the estimate, but the calculation should follow formulas defined by the store or the manufacturer, run by code. The final quantity needs to be confirmed by the customer and, in higher-risk cases, by someone on the technical or sales team.

Is it safe to let AI quote prices on WhatsApp? It is safe when the price comes straight from the price table or the ERP, not from text generated by the model. For off-table prices, discounts and special terms, the AI should only forward the case to a person who approves before the amount is communicated.

Do I need to connect AI to my management system to start? To check stock and delivery status reliably, yes, at least with read access. If integration is not yet feasible, a smaller pilot can use the AI to structure the list and prepare the draft, leaving price and balance for the attendant to fill in.

How do I stop the AI from promising a lead time the store cannot meet? Limit its answers to logistics data that is already confirmed, such as order stage and the set window, and forbid promises of an exact time or of exceptions. When the customer asks for something outside that, the AI should log the request and pass it to the responsible person.

Sources

• Anthropic: Building effective agents
• AWS: What is retrieval-augmented generation
• OWASP: Authorization Cheat Sheet
• ANPD: Autoridade Nacional de Proteção de Dados