AI Transparency

1. Where AI appears in our products

Each product has a conversational assistant in a side panel: the AI Designer in CabDesign, and the assistant dock in StackDesign. You open it, you type, it answers, and it can drive the tools on the page for you.

That is the whole of it. The rest of both products is ordinary software: the drawing canvas, the 3D view, the material and hardware catalogues, the pricing, the cut list optimiser and the nesting are all deterministic code. We do not use AI to score you, to price you, to decide what you are shown, or to make any decision about you or your account. There is no AI in the sign-up flow, the billing, or the support forms.

The assistant is optional. Every tool in both products works without opening it.

2. What the assistant can do

Two useful things, and it is worth separating them.

It answers. Ask how to approach a run of cabinets, what a term means, which material suits a job, how a hinge or a slide works, or what to do next. This is ordinary conversation and nothing changes in your project.

It acts. Ask it to do something and it can carry out a defined set of actions in your project: add a cabinet to a wall, change a dimension, assign a material, set up a room, send work through to the cut list, and so on. The set of actions it can take is a fixed list written by us. It cannot invent a new one, it cannot reach outside your own workspace, and it cannot touch your account, your billing or anyone else's data.

The changes that matter ask first. Before an action with real consequences, the assistant puts a confirmation card in front of you and waits. That covers deleting a cabinet, sending work to the cut list, applying a whole room preset, and any change that would sweep across more than one cabinet at a time. The card tells you how many things are in scope and what will be left alone, so you are agreeing to something specific rather than to the word "yes".

Everything it does is undoable. An action it takes lands as a single undo step, so one undo puts it back the way it was.

3. What it cannot do

We keep a written list of the things people reasonably ask for that the products cannot do yet, each with an honest one-line answer, and that list is handed to the assistant as part of its instructions. When you ask for one of them it is supposed to tell you plainly that it cannot, and offer the nearest thing it can. Countertop design and placing an order for materials are two examples on that list today.

Beyond the product's own limits, the assistant is not:

• A professional adviser. It is not an architect, an engineer, a code official or a lawyer, and nothing it says is professional advice.
• A code or permit check. It does not know your jurisdiction's building code, your inspector, or the permit conditions on your job.
• A site survey. It cannot see your room, your walls being out of square, the real sheet in your rack, or the tolerance of your machine.
• A price quote. Anything it says about cost is an estimate built from what you entered, not a quotation from a supplier.
• Reliable about products. It can name a part that does not exist, or quote a specification that has changed. Check the manufacturer's own documentation.

And, honestly: it can be wrong about ordinary things too. It is a language model. The same question can produce different answers, and a wrong answer usually arrives sounding exactly as confident as a right one. That is the nature of the technology rather than a defect we are about to fix.

4. Where the numbers really come from

This is the most important thing on the page, so it gets its own section.

Cut lists, part sizes, bills of material, sheet nesting layouts and every other measured output are produced by our geometry engine. The engine is deterministic: it computes from the parameters in your project by fixed rules, so the same project produces the same numbers every time. No language model invents any of it, and no language model is asked to do arithmetic on your behalf.

What the assistant does instead is change the inputs. When you ask it to make a cabinet 900 mm wide, it sets that parameter and the engine recomputes. The number you end up cutting to came out of the engine, not out of the conversation. That is a deliberate architectural decision, and it is the reason a chat assistant can be in a tool like this at all.

Deterministic is not the same as correct for your job. The engine cannot know that your measurement was off, that the sheet in your rack is 18.2 mm rather than 18, or that the wall bows. It computes faithfully from what it was told.

5. You are the one who checks it

The same rule is in the Terms of Service, in the sections on the assistant and on engine output, and it is the one thing on this page we would ask you to read twice. A wrong unit is a wrong cut, and neither the assistant nor the engine is standing in your shop to catch it.

If you pass what comes out of our tools to a client, a colleague or an inspector, you are the one presenting it, and you are responsible for having checked it. The Use Policy says the same thing from the conduct side.

6. What it remembers, and how to make it forget

So you do not have to say the same things every session, the assistant keeps a small memory in your own workspace. It holds three kinds of thing:

• Facts about how you work, from a short fixed list of keys we defined in advance: your preferred units, the shape and dimensions of the room you are working on, the style you build in, and a handful of others of that kind. It is a closed list, so the assistant cannot invent a new sort of thing to remember about you.
• A rolling summary of the conversation so far, regenerated at the end of a session and kept short.
• The last few turns of conversation, capped at a dozen, so it does not lose the thread mid-sentence.

You can see it and delete it. The assistant panel has a "What I remember" list showing every fact it holds, each with its own Forget control. Deleting a fact removes it. The whole memory lives in your workspace, so it goes when your project or your account goes, and nobody else can read it: access is enforced in the database itself, per workspace.

The memory is not a transcript archive we mine. It exists so the assistant is useful on the second day, and that is all it is for.

7. What we send, and who receives it

When you send a message, our server sends that message and the project context needed to answer it to our AI provider, which generates the reply. The call is made server side, so your browser never talks to the provider directly and no provider key is ever exposed to you.

Who the provider is. Anthropic. Every customer conversation in either product runs on an Anthropic model. Anthropic is listed as a sub-processor in the Privacy Policy, alongside everyone else who processes data for us.

There is a second AI provider, and it is connected but not yours to use. Our server also carries a route to Google's Gemini API, and a third labelled ChatGPT that is not connected to anything. The switch can only be turned on by one of our own platform administrator accounts. There is no address you can type, no setting in your account and no supported way for you to reach it, so no customer conversation is sent to Google. We use it for our own testing. It is disclosed here, and Google is listed as a sub-processor in the Privacy Policy, because the connection is live on our server and we would rather name a thing that exists than describe Anthropic as the only provider wired up. If we ever make a second provider available to you, we will say so here and in the Privacy Policy before any of your data reaches it.

What is not sent. Your billing details, your password, and other people's workspaces are never part of an AI call. Our product analytics never receives your designs, your dimensions, your customers' records or your assistant conversations.

Please do not paste anything into the assistant that you would not want processed by a third-party AI provider, including other people's personal information you have no right to share.

8. Training

What we do. Nothing. We do not train, fine-tune or evaluate any model on your designs, your dimensions, your customer records or your conversations. We build no models of our own.

What our provider commits to. We use Anthropic through its standard commercial API, and under those terms the content we send and the replies we receive are not used to train models, and the provider's operational logs are deleted on a short rolling window. That is a contractual commitment made by the provider rather than something we can enforce from inside our own code, and we would rather describe it that way than imply we have a technical guarantee we do not have. We re-check those terms when they change, because vendor terms drift.

What we ask of you. The Use Policy asks the same thing in reverse: do not use the assistant, or its output, to build, train or evaluate a competing model or dataset.

9. Limits on how much you can use it

Every AI conversation costs us money, so there is a daily usage ceiling per account. When you reach it the assistant tells you in plain words that you have hit today's limit, that your design is saved, and that you can keep working with the tools on the page. It does not go quiet, and it does not pretend to answer with something worse.

We would rather have a limit that is honest about itself than an unlimited promise that degrades silently. If a ceiling is getting in the way of real work, write to us.

Working around the limit, for example by rotating accounts, is a breach of the Use Policy.

10. When it gets something wrong

It will get things wrong. When it does, the useful thing you can do is tell us what you asked, what it said, and what you expected instead. There is a bug report button inside the product once you are signed in, and during the open beta a report that leads to a fix earns free service under the Open Beta Terms.

If the assistant does something in your project you did not want, undo puts it back. If it says something harmful, offensive or plainly unsafe, report that too and tell us it was the assistant, because those reports get looked at first.

We do not promise the assistant will be accurate, and the Terms of Service disclaims warranties on its output. That is the legal position. The practical position is that we want to know when it is wrong, and we fix what we can.

11. Changes, and how to reach us

We will update this page when what the assistant does, what it can reach, or who processes it changes. The effective date at the top always shows the current version. A change of AI provider, or a new place AI appears in the products, gets named here and in the Privacy Policy before it reaches you.

Bespoke Woodcraft Studio LLC, a California limited liability company
688 N Rimsdale Ave
Covina, CA 91722, United States

Email: support@cabdesign.app for CabDesign, or support@stackdesign.app for StackDesign, answered within one business day.