On 24 June 2026 PAZ Academy runs an Archicad Powersession on KI in Zürich — a hands-on day on pulling artificial intelligence directly into the Archicad workflow: AI-assisted design variants, text-to-geometry, image-to-model, GDL and parameter automation, and the part most demos skip — where the limits, the data-protection questions, and the quality controls actually sit. The promise is deliberately unglamorous: tools that work on Monday morning in the office, not on a conference stage.

Let me draw the system the way I draw any production pipeline — as a dependency graph, not an architecture diagram. An AI-in-Archicad stack has four layers stacked on each other: the model (your .pln, the system of record), an API layer that exposes geometry and parameters to code, a codegen layer that turns intent into commands, and an inference layer — the model that proposes variants or reads an image. Each arrow between layers is a place where queues build and trust leaks. The headline feature is always the inference layer. The failure mode is almost always the boring arrow underneath it.

←TODAY: In 2026 Archicad 29 ships a documented Python/JSON API — the model is finally scriptable without IFC round-trips. →3012: By the Zurich-3012 horizon the BIM model is a queryable substrate, spoken to in language, not clicked through in palettes. Fulcrum: The office that wins is not the one with the best AI — it is the one that mapped which layer it actually depends on before the vendor’s roadmap changed underneath it.

Where this is possible now and was not five years ago: the API layer matured. Graphisoft-native automation — the Python/JSON connection, the Hotlink Manager — means an AI suggestion can land as native Archicad elements with real properties, not orphan meshes pasted on top. PAZ’s own work sits exactly here: PAZGPT already lets an architect talk to the model (“add a column there”, “what does this floor cost?”) and have it execute in-model, and the PAZ Grasshopper↔Archicad Library fills the missing parts of the current Grasshopper-to-Archicad connection so a parametric definition survives the trip into BIM. The inference is the easy part to buy; the wiring is the part PAZ had to build.

The quality and privacy questions are not a footnote — they are the second single-point-of-failure most offices don’t draw. As t3n reported on DSGVO compliance, AI can take roughly 75% of the documentation load off a team, but it structures while the human decides — “KI strukturiert, der Experte entscheidet.” And per a NTT DATA study covered by Computerworld, rising AI use is hitting the structural limits of existing IT architectures, not just model accuracy. Translate that to a studio: your client’s site survey, your competition entry, your cost model are now leaving the building unless you’ve drawn where the data goes. PAZ has covered this thread before — the KI & Datenschutz material and the Ethik von KI course exist precisely because the variant generator is useless if the data path can’t be defended.

Atelier: Run the AI variant on the massing study, then push it through PAZ Volumenstudie so the result lands as native walls, slabs and zone stamps — and let Kostenrechner nach SIA416 price it in CHF/m² before anyone falls in love with the render. AI proposes; the SIA norm decides what’s defensible.

Hack: This Hack teaches you to attach code to a running Archicad and see your automation surface — the first step before any AI sits on top. The medium is runnable Python; the domain is Workflow. Install the official connector with pip install archicad, open your project, and run:

from archicad import ACConnection
conn = ACConnection.connect()            # attaches to the open Archicad
acc = conn.commands
walls = acc.GetElementsByType("Wall")    # every wall, as element IDs
print(f"{len(walls)} walls — this is what an AI agent can touch")

Those element IDs are exactly what a codegen layer hands to an inference call. Once you can list them, you can rename, classify, or feed them to a model — but you now also know precisely what leaves the machine.

Bring one question to the Powersession on 24 June: draw your real dependency graph this week — model, API, codegen, inference — and find the third single point you didn’t know you had. Then come learn to automate the arrows, not just admire the layer on top. Register at pazacademy.ch.