AI Window Configurator: When AI is the Interface, Not the Brain

In short. AI interprets the user's language and nothing else. The rules of the trade—how many sashes are needed for a certain width, when a window or door becomes a French window, which glass is required by the safety standard—remain deterministic, verifiable code, always the same. The technical document comes out correct by construction, not because "the model usually gets it right." It's a three-stage architecture—interprets → decides → executes— which is valid for windows and doors but applies identically to any complex configurator: nautical, custom furniture, systems, industrial products with combined parameters.

Where we come from: the form configurator

The window-and-door configurator we discussed previously does one thing and does it well: it translates the user's choices—family, material, opening, measurements, sash distribution, mullioned windows, glass—into a dimensioned technical drawing and a PDF list, with a single SVG engine shared between screen and print. It's solid, proprietary to the window and door manufacturer, with no lock-in. It remains an excellent tool, and for many contexts, it's still the right choice.

However, it has An entry threshold: you have to know what to look for. Menus, fields, drop-down menus. A customer accustomed to windows and doors fills them out without thinking; an occasional customer—a private individual renovating a home—sometimes stops at the "type of opening" because they don't know which one they have. The form is precise but asks the user to speak the machine's language.

The question we asked ourselves is simple: what if the machine spoke the user's language?

The evolution: the configurator that understands words

The AI Configurator (try it here) reverses the interaction. Instead of filling in fields, the user describes what they want. who wants, in his own words, as he would say over the phone: "a double-leaf wooden window 120 x 150 cm," or "a large sliding window with a surface area of ​​about 2.40 m high by 2.45 m wide, double-glazed." The tool interprets, draws in real time on the same technical table as before, explains its choices, and builds the bill of materials—always without prices, always leaving the economic evaluation to the window and door manufacturer.

Real requests are messy, and this is the testing ground. They arrive in meters instead of millimeters, with the Italian decimal point, with the height before the width, with local terms, with annotations that the software doesn't model—"white color," "triple seal," "insulated casing." A configurator that only accepts clean inputs is a demo toy. A tool that supports real requests is a working tool. We spent a good part of the development precisely here: ensuring that "ML 2.40 (Height) ML 2.45 (WIDTH)" reliably becomes a 2450 × 2400 mm window, and that "white color" is stored as a note for the estimate provider instead of confusing the parsing process.

The heart of the article: where AI goes, and where it doesn't

Here we come to the point that really matters, and which is worth understanding even if you don't write software. Because it's the difference between a configurator you can trust and one that occasionally puts the wrong window on the list with the air of having done its job.

A linguistic model — What we call "AI" — it's extraordinary at understanding language and terrible at consistently following the rules. If you ask it the same thing twice, it can answer you in two different ways: it's its nature, not a defect to be corrected. It's great at interpreting "sliding window" and understanding that you mean a large window. It's terrible at deciding that that window, being 6 meters wide, can't be a normal window but must belong to another family with other dimensional limits. The first is interpretation. The second is a rule of the trade, and a rule can't change depending on the model's mood.

So We have clearly separated three stages.

AI interprets. It reads the user's text, however messy, and extracts the intention: what type of window, what dimensions, what material, what annotations. It is the only stage where the linguistic model lives, and it is the only one where its flexibility is an asset.

The rules decide. Once the intention is understood, deterministic rules—normal code, always the same—establish the technical facts. A window more than a meter wide takes two sashes, below it takes one: always, not "usually." A full-height window triggers the safety glass recommendation according to UNI 7697. A sliding door too wide to be a window is promoted to a sliding glass door; a glass door too narrow is reverted to a French window. These decisions are not made by the model. The code does, and it always makes them in the same way.

The tools execute. They design the window, calculate the list, generate the PDF. It's the same single-source engine as the previous configurator: what you see on the screen is what you get. that comes out in print, by construction.

The practical result is that the final document never depends on the unpredictable part of the system. The model can also describe in words "I configured three windows and a French window" while underneath only one is being prepared: it doesn't matter, because what ends up in the list is decided by the rules, not by the model's narrative. We trust verified data, not words. It's an architectural choice, and like all architectural choices, it's only visible when something could have gone wrong and didn't.

A concrete example: the 90×230 French window

Let's take a real case, one of those that seem trivial but aren't. The user requests a "90×230 sliding French window." The word "door" pulls the model toward the door family—a blind door. But a French window isn't a door: it's a tall, glazed window, through which you can pass. 90 cm wide, moreover, it's below the minimum size for a sliding glass door.

The model, left alone, oscillated: sometimes it was a door, sometimes it was a glass window inflated to a size the user hadn't requested. Each time, a different result. The solution wasn't "convincing the model better"—it was installing a deterministic network that, after interpretation, corrects: if the text says "French door" but the chosen family is door, it's changed back to window; if it's If it's too narrow for a glass window, it remains a French window with its actual size. Now the result is always the same, because the variable part of the system no longer decides.

It's a small example, but it's the paradigm in miniature: AI understands, the rules guarantee. Repeated for dozens of cases—the measurements in meters, the number of sashes, the safety glass, the grouping of identical windows in a single line with quantities—it builds a tool that behaves like a competent technician, not like a lucky text generator.

Comparing the two configurators

One doesn't replace the other: they solve the same problem for different users.

The form configurator is Precise and direct for those who already know what they want. Zero ambiguity, zero processing time, every field under control. It's ideal for the expert user, for the retailer, for those who compile dozens of configurations a day. The limit is the entry threshold: it requires knowledge of the terminology.

The AI configurator lowers that threshold to zero: anyone who can describe in words what they have in mind gets a result. It's ideal for the occasional end customer, for the request that comes in from the website, for those who don't know the name of the opening type but know that "the kitchen window opens to the side." The cost is a small processing time and the need, for us who build it, to carefully shield everything the model could misinterpret—which is exactly the architectural work mentioned above.

The beauty is that they share the same heart: identical design engine, identical bill of materials, identical technical drawing board aesthetics, identical white-label philosophy and no lock-in. The AI ​​didn't rewrite the configurator. It added a natural language front door to an already robust engine.

Why this scheme is valid well beyond windows and doors

And here we are at the reason why this article is about architecture and not just windows and doors.

The window and door frame is a product with combined parameters: few variables — family, material , opening, measurements, doors, glass—which combine in thousands of valid configurations, each with precise technical constraints. But windows and doors are by no means the only product made this way. A boat to be fitted out, a custom piece of furniture, a system to be sized, an industrial machine with its options, a tank, a pergola, a modular kitchen: they all share the same logical structure. Parameters that combine, trade constraints that must always be respected, a technical document that must be correct.

For all of these, the scheme interprets → decides → executes functions identically. The domain changes—the rules of a boat are not those of a window—but architecture is not. AI puts the front door in natural language; the deterministic rules of the trade remain the reliable core; the tools produce the drawing and the document. What we've built for windows and doors is, ultimately, a reusable framework for a whole class of problems.

It's the same principle that guides everything we do, and which lies behind our Designed. Not assembled: not a product to be resold in the same way to everyone, but a solid architecture to be adapted, each time, to the specific job of the user.

Frequently Asked Questions

What's the difference between the form-based configurator and the AI ​​one? The former is used by filling in menus and fields: precise and immediate, ideal for those who already know the terminology. The latter is used by describing in words what you want: it lowers the entry threshold to zero, ideal for the occasional end customer. They share the same design engine, the same bill of materials, and the same white-label philosophy.

Does AI decide the technical characteristics of the window or door frame? No, and it is the core of the project. AI only interprets the language of the request. The technical characteristics—number of sashes, family, safety glass, dimensional adjustments—are established by deterministic rules in the code, which are always the same. The final document never depends on the unpredictable part of the system.

Can you trust an AI-based configurator? You can trust it because AI is not left free to decide. It does only one thing—understand what the user requests—and everything that matters for the correctness of the document is entrusted to verifiable code. We've built safety nets for every case where the model might misinterpret, and we've validated them with reproducible tests before putting them into production.

Does it handle "real," poorly written requests? Yes: that's exactly what it's designed to do. Measurements in meters or millimeters, commas or decimal points, height before width, local terms, and technical annotations that the software doesn't model (colors, casings, gaskets) are recognized—and when they can't be modeled, they're kept as notes for the estimate provider instead of being discarded.

Does it show prices? No, like the form configurator. It produces the correct technical specification and the bill of materials, leaving the economic valuation to the window and door manufacturer, who calculates it based on their own production and installation variables.

This approach can be used Can it be applied to other products? Yes, and that's why we talk about it as architecture and not just as a product. The "interpret & decide & execute" model applies to any product with parameters combined with technical constraints: nautical, custom furniture, systems, machinery, industrial products. The domain changes, not the architecture.

Is the code proprietary, like the other configurator? Yes. Same white-label philosophy and no lock-in: each company owns its own installation, with its own brand and code.

Want to see it in action?

The AI ​​Configurator is online and interactive. You can try it now by writing a request in your own words and watching the window take shape, or compare it with the configuratore a form from which it was born. And if you produce windows — or any other product with combined parameters — and you want a quote tool with your brand, your code and the right architecture underneath, tell us your needs: we'll build it custom.