On 22 July we ran an Archicad Powersession on PAZGPT, and the moment the room changed was not the moment something spoke. It was the moment someone asked which walls in this storey have no finish assigned — and got a number back instead of an afternoon.

We had named that pain ourselves, six weeks earlier. Our June Powersession on Wandfinishes & Materialisierung (15 June 2026, 11:30–12:00 CET, online, in German, CHF 0.00) was thirty minutes about finishes managed by hand: lists and legends drifting apart, every change resetting the documentation, the model quietly knowing the answer the whole time. A model is made of parameters. Almost nobody talks to them.

Why now, and not five years ago

Graphisoft shipped the first BIM software in 1984. The Grasshopper connection passed its tenth anniversary, as our own Archicad Freebies page notes. None of that is new — what changed is the cost of the first sentence. For a decade, asking your model a question meant knowing how to ask, and that barrier was never really a technical one. It was a translation problem, and translation layers have always done more for adoption than features do.

The test we actually care about

SIA 416 does not describe a drawing. It partitions volume — Geschossfläche, Konstruktionsfläche, Nettogeschossfläche — into a statement a model can be asked to prove. So the honest question about any AI layer on Archicad is not whether it can hold a conversation. It is whether it can produce a number a quantity surveyor will sign.

The fear in the room is real, and it is not about this tool. It is about the class of tool that will not show you its hands: the layer that edits your model and cannot tell you which elements it touched; the round trip that sends your geometry somewhere you have no retention agreement with; the capability that exists only as a service and disappears when the service does, taking the workflow your office built around it. That is the risk we teach against, and it is why we built our own.

PAZGPT is ours. It runs inside your own Archicad session, and every action it takes is one you could have typed yourself and one you can read back afterwards. That is not a promise about our intentions — it is a property of where the thing sits. Ask the same question of every tool in your office, ours included.

So we start with the boundary, not the magic. Write your read-only boundary list before you run your first prompt. Then count the walls with no surface assigned, before and after any AI-assisted edit. A number that moves the way you expected is trust. A number that moves otherwise is a bug you caught for free.

This is for the Archicad power user drowning in repetitive selection work, the BIM manager enforcing standards across a team, and the office owner who suspects the quantities are right but cannot prove it before Friday — the same audience we have served since the Zurich workshop space opened three minutes from the Hauptbahnhof, next to the Polybahn. PAZGPT and the rest of the in-house tooling live on the PAZ Archicad add-ons page.

The next one

ARCHICAD Automation Bootcamp — 22–25 October. Three 30-minute sessions and a one-hour bonus masterclass, 49 CHF. Not a demonstration: you leave with your own read-only boundary list, your own before-and-after check, and a model that answers you. Places at pazacademy.ch/workshops/.

←TODAY: the answer has been in the model for a decade; the question was the expensive part.
→3012: the building answers in language, and every answer still shows its working.
Fulcrum: an AI you cannot audit is not a colleague, it is a confident stranger — the difference is a boundary you wrote first.

Source: original article