AI in an architecture firm: 6 questions for any tool

Almost every office now has someone trying out AI. It usually starts with ChatGPT, a draft for the building description or a summary of a long permit. That often works surprisingly well. It gets hard with the question that costs the most time in daily work: what applies to this project, at this location, in this building class, and where does it say so?

This post is for office principals and project leads deciding whether an AI tool comes into the office, and which one. It names six questions to put to every vendor, including us.

1. Which law does the tool work with?

Austria has nine building codes, and each state declares the OIB guidelines binding in its own edition and on its own date. A tool that knows “building law” only from its training cannot tell you which version applies in your state. Many AI tools for planning offices are also built for Germany: state building codes, DIN, VOB and HOAI rather than RIS and OIB.

Ask: Where do the legal texts come from? Which states are covered? How current are they, and does the tool say which edition it uses?

2. Can you check every answer against the original?

In planning, an answer without a citation is an opinion. You need the section, the clause of the guideline or the page of the drawing, with a click that opens the passage. Check whether the source is merely attached or also verified: language models occasionally invent citations that look plausible.

Ask: Is every citation checked against the source text before it appears? Does it open at the cited passage?

3. Does it work with your own documents?

Most questions depend not on the law alone but on the law and your project: the zoning plan, the permit, the section. And often on your office: how did we solve this detail on the last housing scheme? A tool that searches only legal texts answers half the question.

Ask: Can the tool read drawings, including as images? Does it search the office archive? Are one office’s documents kept apart from every other office’s?

4. Does it say what is missing?

The underrated mark of a good tool is that it admits what it does not know. If the zoning plan is missing, the answer should say so rather than sound as if it had read it. If the building class is missing, it should ask, or state its assumption.

Ask: What happens when a key document is missing? Does the answer show its assumptions?

5. What happens to your data?

Drawings, tenders and unpublished projects are confidential. Ask specifically: who processes the requests, in which country, and is your data used for training? A vendor who answers openly, even where the answer is not ideal on every point, is more trustworthy than one who just says “secure”.

Here is ours: Piloti does not train models on your data, drawings remain the property of your office, sign-in runs through WorkOS (USA) and AI requests through OpenRouter (USA) to model providers that may be based outside the EU. The details are in the privacy policy.

Ask: Which providers are involved? Is my data used for training? Who owns the documents?

6. Does the answer stay stuck in a chat?

A good answer that disappears into a chat history does little for the project. Ask whether an answer can become a file note, a checklist or a review report in the project, whether the team can reach it, and whether open points are followed up.

Checklist for the first conversation

• Which legal sources, which states, how current?
• Citation down to section, clause or page, verified and clickable?
• Your own drawings and office archive as sources, kept apart from other offices?
• Handling of missing documents and open assumptions?
• Providers involved, training, ownership of the data?
• Does the answer become a document in the project?

The fairest test

No checklist replaces a real question. Take one from a current project whose answer you already know, and put it to every tool you are considering. Pay less attention to how fluent the answer sounds than to whether you can find every statement in the original.

If you want Piloti in that test, send us the question and we will show you the answer. Why we built Piloti this way is on Why Piloti; how it differs from Reiner AI, in the comparison.