AI for Glass and Metalwork Shops: Custom Quotes With Technical Review

AI for glass and metalwork shops works best for custom quotes when the agent builds the pre-quote from standardized data (measurements, photos and materials) and a technical lead validates it before any number becomes a proposal. The agent organizes the information, calculates with the company's own rules and flags what is missing. Whether a measurement is usable, whether the glass is suitable and whether the job is feasible stays with a qualified person.

Key takeaways

• The gain comes from standardizing the input: a measurement sheet, a photo checklist and a materials catalog with your own item codes.
• The agent produces a pre-quote with a clear status (complete, incomplete or needs site visit), never a final quote.
• The technical lead validates critical measurements, thicknesses, glass type, structure and fixing before anything is sent.
• Prices and calculation rules live in a database the company maintains, not in the model's memory.
• Start with one product type (for example, a shower enclosure or a simple gate), log every decision, and expand only after that.

Why do quotes in glass and metalwork shops get stuck?

A made-to-measure quote depends on information customers rarely know how to give. They write that they want a shower enclosure, send a crooked photo of the bathroom and ask how much it costs. Whoever answers has to find out the width, the height, whether the wall is plumb, whether the floor slopes, what color the profile should be, whether anything blocks the door from opening and how the pieces will be carried in.

Metalwork is similar. A gate needs the opening size, the type of movement, material, finish, automation, fixing points and the condition of the masonry. A misread becomes rework on cutting, welding or glass that cannot be reused on another job.

So the bottleneck is not just typing up the calculation. It is guiding the data collection, checking that it is coherent and translating it all into a list of materials and services. That repetitive part is exactly what an agent can take over, as long as the input has the same format every time.

What the agent does and what stays with the technical lead

Splitting responsibilities clearly avoids the biggest risk: a polished but wrong number reaching the customer. A practical split looks like this.

• The agent runs the conversation on WhatsApp, asks for missing data and refuses to proceed with inconsistent information.
• The agent organizes measurements and photos into the standard sheet and identifies the product within the company's catalog.
• The agent calculates quantities and builds the materials list using rules and tables supplied by the company.
• The agent raises alerts, such as an unusual measurement, a very large opening, a risky location or a photo with no size reference.
• The technical lead validates critical measurements, glass thickness and type, gauge and structure type, fixing and installation conditions.
• The technical lead decides when a measuring visit is needed and when the job requires specific technical documentation.
• The sales person approves the final price, terms and lead time before it goes out.

The pattern behind this split is well known: predictable workflows for what is a rule, human judgment for what is a risk. Anthropic describes the distinction between defined workflows and more autonomous agents in its Building effective agents guide, which is a useful reference for deciding how much freedom to give the agent. For a technical quote, the less freedom on the numeric part, the better.

How to standardize measurements: the sheet the agent must fill in

Without a standard sheet, every conversation is different and the agent ends up guessing. With a sheet, it knows exactly what is missing. A minimal sheet, adapted to your product type, usually has these fields:

• Product type: shower enclosure, window, door, mirror, guardrail, gate, grille, staircase, canopy or any other item in your catalog.
• Installation location: room, floor, indoor or outdoor, and whether there is moisture, strong sun or salt air.
• Opening measurements: width and height at no fewer than three points (start, middle and end), recorded in millimeters or centimeters, always in the same unit.
• Opening condition: plumb, level, squareness, wall type, and whether there is tile cladding or an old frame.
• Who measured and how: the customer, an architect, a mason or the company's team, and with what instrument.
• Desired finish: profile or paint color, glass type, hardware type.
• Access: stairs, elevator, distance to the installation point, and whether hoisting or extra crew is needed.
• Desired timeline and customer notes.

Two rules help a lot. First, every measurement that comes from the customer is tagged as reported, and only becomes confirmed after technical validation or a site visit. Second, the agent never converts an opening into a production measurement. Clearances, setbacks and tolerances are your shop's decision, and they enter as registered rules reviewed by someone who knows the trade.

How to ask for photos the agent can actually use

Bad photos lead to bad quotes. The customer should get simple instructions, with the same checklist every time. A lean version:

• A wide photo of the room, so the whole opening and its surroundings are visible.
• A front-on photo of the opening, taken from straight ahead, with the phone at the height of the middle of the opening.
• Detail photos of the corners, floor and top, where unevenness and leaks tend to show up.
• A photo with a size reference, such as an extended tape measure or a sheet of paper of known size held against the opening.
• A photo of the existing fixing point, when it is a replacement or renovation.

The agent can check the basics: whether each type of photo is there, whether it is legible, whether the whole opening is visible. When something is missing, it asks again and explains why. What it should not do is take a precise measurement from a photo. An image is for context, for identifying the type of opening and for supporting validation. It does not replace the tape measure.

There is also a privacy point. Photos of homes and apartments show the inside of the customer's residence, sometimes with people and documents in view. Ask only for what is needed, tell customers what the photo will be used for and define how long it is kept. Brazil's data protection authority (ANPD) is the official source of guidance on personal data processing, and the choice of legal basis and retention period should go through your legal counsel.

How to organize the materials catalog so the agent does not invent things

The agent only calculates well what has been registered. That makes the materials catalog the central piece. Instead of letting the model infer what tempered glass of a given thickness is or which profile suits a sliding window, register each item with a code, description, sales unit, expected waste and usage rules.

For a glass shop, this includes glass types (clear, tempered, laminated, mirror), available thicknesses, profile lines and colors, hardware, seals and accessories. For a metalwork shop, it includes tube and sheet types, gauges, finishes, paints, locks, hinges, motors and accessories. For each assembled product, record the recipe: which items go in, in what quantity and as a function of which measurements.

The agent queries this database and builds the list. If something is not in the catalog, the correct result is a not-registered-item alert, not an invented suggestion. This is the logic of retrieving information from a reliable source before answering, which architecture literature calls retrieval-augmented generation (RAG) and which AWS explains in its What is retrieval-augmented generation page. For price tables and numeric rules, though, a direct, deterministic lookup is better, with the model only calling the function.

One practical consequence: prices and margins stay in a spreadsheet or system owned by the company, under the control of whoever answers for them. The model does not learn prices and does not repeat them from memory.

Illustrative scenario: a bathroom shower enclosure on WhatsApp

The scenario below is illustrative and does not represent any real customer.

A person writes to a glass shop asking for a sliding shower enclosure. The agent introduces itself, explains that it will prepare a pre-quote that a technical lead will review, and asks for what the sheet requires. The person gives three width measurements and two heights, sends the requested photos and chooses clear tempered glass with a black profile.

The agent checks the sheet and notices that the three widths differ by more than the tolerance the company has registered. It does not pick one of them. It marks the pre-quote as needs measuring visit and records the reason. Then it builds the materials list from the product recipe, flags the items that depend on the final measurement and leaves the price as a reference range calculated from the company's table, with a notice that it is subject to confirmation.

The technical lead opens the panel, sees the sheet, the photos and the alert, and decides to schedule the visit. After measuring, they adjust the fields, confirm the glass type and release the quote. The sales person reviews terms and lead time and sends the proposal. Every step is logged: who supplied what, what the agent calculated, what was changed and who approved it.

How to roll it out with control: map, limit, pilot and hand over

The rollout follows the same method in any operation. First, map: how a quote is created today, which questions get asked, where rework happens, which mistakes have already occurred and how long it takes from first contact to proposal. That picture is the baseline for knowing later whether anything improved.

Second, limit: define which products the agent handles, what data it can see, which actions it can take and what requires human approval. A quote sent to the customer, a discount and any promise of lead time stay with people. This fits the principle that significant decisions made automatically should be open to review, a topic we cover in automated decisions under LGPD.

Third, pilot one small, real flow, with logs and an agreed success criterion, such as the number of complete sheets on the first round or the time the technical lead takes to validate. Before using real data, it is worth testing with simulations, as we explain in mock tools for testing AI agents. Finally, hand over: document the flow, train the people who operate it, define who maintains the catalog and decide whether the pilot will be expanded, adjusted or ended.

If your operation also builds custom furniture or other made-to-measure items, the logic is close to what we describe in AI for custom furniture makers. For the channel connection, the guide on how to integrate AI into your company's WhatsApp covers the technical choices.

Which risks should you watch once the agent is live?

The first risk is false confidence: a well-formatted pre-quote looks right even when it starts from a wrong measurement. The mitigation is to always show the source of each data point and keep the validation status visible to whoever approves.

The second is content coming from the customer. Messages, files and photos can contain instructions that try to steer the agent off course, which is called prompt injection. Keep the agent on minimal permissions and treat any external text as data, never as an order. OWASP maintains useful references on this in its Top 10 for LLM Applications.

The third is catalog drift: prices and products change, and an outdated database produces wrong proposals. Assign an owner and a review routine. The fourth is overall risk management, for which the NIST AI Risk Management Framework offers a structure to identify, measure and address risks in AI systems.

Also track the corrections the technical lead makes. They show where the sheet is ambiguous, where the catalog is incomplete and where the agent should have asked for more information.

Frequently asked questions

Can AI replace the measuring visit?

It should not. It reduces unnecessary trips by identifying complete, coherent sheets, but a measurement that becomes a glass or metal cut must be confirmed by whoever is technically responsible. When in doubt, the right move is to book the visit.

Can the agent calculate the quote on its own?

It can build the materials list and apply the rules and table registered by the company. The final price, terms and exceptions should go through human validation. The model should not estimate prices from memory.

Do I need a special system to start?

Not necessarily. A pilot can use a structured sheet, a catalog spreadsheet and WhatsApp, with a simple review panel. What matters is having a log of each decision and a clear owner for validation and for maintaining the catalog.

How should customer data and photos be handled?

Ask only for what is needed, state the purpose, restrict who has access and set a retention period. Consult legal counsel on the legal basis and use ANPD guidance as a reference. If an AI provider processes the data outside Brazil, also review the requirements covered in international data transfer under LGPD when using AI providers.

Sources

• Anthropic: Building effective agents
• AWS: What is retrieval-augmented generation
• OWASP Top 10 for LLM Applications
• NIST AI Risk Management Framework