AI for Glass and Metalwork Shops: Standardize Measurements and Quotes
How glass and metalwork shops can standardize measurements and quote data so an AI agent builds proposals without inventing specifications.
Want to apply this in your company?
Book a call with Juan →An glass or metalwork shop can only use an AI agent to build proposals safely when the input arrives standardized. A measurement scribbled on a notepad, a photo with no reference point, or an order without a defined glass type forces the agent to fill the gaps with assumptions, and those assumptions turn into wrong prices and wrong lead times. The fix is a measurement form with required fields, validation rules, and a clear point where a person takes over.
Key takeaways
- The agent needs measured data, not guesses. Every specification should come from a known source, such as an on-site measurement or a choice the customer confirmed.
- Required fields and fixed units eliminate most quoting errors before any calculation starts.
- Measured data, chosen options, and calculated values should be stored separately so each one can be audited.
- Anything outside the expected range or left blank should go to human review, without exceptions.
- Standardization starts with a pilot on one type of job, with a record of every proposal, and only then expands.
Why the agent invents specifications when measurements are loose
A language model is good at completing text in a plausible way, and that is exactly the risk in quoting. If a customer writes "window 1200 by 1000 with glass," the message is missing the glass type, the thickness, the frame colour, the opening direction, and whether the measurement is the rough opening or the finished frame. An agent without clear rules may choose standard clear glass, assume a white profile, and keep going, producing a proposal that looks complete and is not.
The problem is not the model's capability. It is the input. That is why standardization is a business task: define which information must exist, who collects it, in which unit, and with what level of confirmation. After that, the agent only organizes and calculates with what has been confirmed.
Fields the measurement form needs to have
The form should be the same for every enquiry, whether it comes through WhatsApp, the counter, or a site visit. The minimum fields usually include:
- Customer identification and the job address, with the date of measurement.
- Installation location, such as bathroom, living room, facade, or balcony, and the floor.
- Width and height in millimetres, with a note on whether they are rough opening or finished dimensions, and who took them.
- Product type: window, door, shower enclosure, balustrade, gate, canopy, or another item from the shop's catalogue.
- Glass type and thickness, selected from a closed list rather than free text.
- Profile type and finish, also from a closed list, with the colour confirmed by the customer.
- Opening direction, fixing type, and accessories such as locks, handles, or tracks.
- Photos of the location with the measurement visible, including at least one angle that shows the whole opening.
- Site access notes, such as scaffolding, stairs, or building access restrictions on working hours.
Each required field should have a defined format. A measurement is accepted only as a number with a fixed unit. Glass is accepted only if it appears in the catalogue. If the customer does not know a value, the field is marked as pending, not filled in by assumption.
Keep measured, chosen, and calculated data apart
A reliable proposal has three layers. The first is what was measured: width, height, photos, and site conditions. The second is what was chosen: glass type, colour, and accessories, each with a recorded confirmation. The third is what was calculated: area, profile lengths, hardware quantities, cost, and lead time. The agent can calculate the third layer from the first two, but it should never mix the layers.
Without this separation, when a customer asks why a quote came in high, nobody can answer precisely. With it, the team can see whether the difference came from a corrected measurement, a glass choice, or a calculation rule, and adjust the right source.
Rules the agent follows and when a person takes over
Validation rules should be deterministic, written in code or in a rules spreadsheet, not only as instructions to the model. Some are simple: a measurement outside the range the shop fabricates, a product that does not exist in the catalogue, a glass type that the manufacturer does not allow in a given location, or a required field left empty. In those cases the agent does not produce a proposal. It records the reason and passes the case to a responsible person.
AI fits where judgment is needed, such as explaining options to the customer, interpreting an unclear description, and suggesting clarifying questions. The decision to accept an unusual measurement or substitute a material stays with a person. This design also supports compliance with Brazil's LGPD, because the agent collects only the data the proposal needs and does not keep personal information beyond what the enquiry requires. It is worth defining who can see job photos and how long they are stored. For a broader view of this kind of data handling, see International Data Transfer Under LGPD When Using AI Providers.
On security, customer messages are untrusted text. A customer may, on purpose or by accident, write instructions that try to change the agent's behaviour, such as asking for a discount or telling it to ignore the measurement rules. The agent should treat those messages as data, never as orders, and any price action should require approval. The OWASP guidance on prompt injection prevention covers this risk in detail.
Illustrative scenario: quote for a shower enclosure and a sliding window
Illustrative scenario: a metalwork shop receives, through WhatsApp, a request for a bathroom shower enclosure and a sliding window for the same apartment. The customer sends two photos and types the measurements from memory. The agent replies with the form questions: whether the measurement is rough opening or finished frame, which glass type they want from the catalogue options, and which profile colour. The customer chooses the glass and the colour but does not know whether the measurement is the rough opening. The agent marks that field as pending, does not calculate the proposal, and books a site visit. On the visit, the measurement is confirmed, the calculation is redone from the measured data, and the proposal goes out with the three layers kept separate. If the window had been outside the shop's fabrication range, the case would have gone straight to the technical team without an automatic proposal.
How to apply the method in a glass shop
The method has four steps: map, limit, pilot, and hand over. Mapping means describing how measurement and quoting happen today, with the timeline of a real proposal and the points where information gets lost. Limiting means defining which data the agent may use, which actions it may take, and what always stays with a person. Piloting means testing one type of job, such as shower enclosures, against a success criterion agreed in advance, for example the number of proposals sent back for correction, with a record of every decision. Handing over means giving the workflow to the team with documentation and deciding whether the next step is to expand, adjust, or stop. This is a way of working, not a promise of results. The pilot shows whether the form works in the shop's daily routine.
To test the rules before serving real customers, use simulated measurement and order cases, as described in Mock Tools for Testing AI Agents: Safe Sandbox Testing Guide. To understand how data, rules, and models can be separated into layers, the article on Company Superintelligence Layer Architecture helps clarify responsibilities.
Plan to standardize measurement capture
- List the product types the shop fabricates, and reduce the catalogue of glass, profiles, and finishes to the options that are actually sold.
- Write the measurement form with required fields, fixed units, and closed lists for glass, profile, and finish.
- Define who measures, how photos are recorded, and when a site visit is mandatory.
- Create validation rules for measurements, products, and empty fields, and define how cases are handed to the team.
- In the quoting system, separate measured data, chosen options, and calculated values.
- Test with past proposals: compare what the agent would have built with what the shop actually sold, and fix the differences.
- Run a pilot on a single product type, record every proposal, and review the rejected ones with the team before expanding.
Frequently asked questions
Do we need a new system to standardize measurements? Not necessarily. Many shops start with a digital form and a rules spreadsheet, and only later assess whether it is worth integrating with the quoting system they already use. What matters most is that the form exists, has required fields, and is the same across every sales channel.
Can the agent build the proposal on its own if the customer agrees? It can draft a proposal from confirmed data, but the final proposal should go through a human check whenever a measurement is pending, a product is outside the catalogue, or a discount is requested. This review stops an input error from turning into a sales commitment.
Which customer data needs care under the LGPD? Name, job address, phone number, and photos of the property are the most sensitive in daily work. Collect only what the proposal requires, state the purpose of using that data, and define how long photos are kept and who has access to them.
How do we know whether the pilot worked? Define the criterion before starting, using a simple measure such as how many proposals come back for correction because of a measurement or specification error. If that measure improves and the team can run the form without improvising, it makes sense to expand. If it does not improve, the next step may be to adjust the form or stop and review the process.