The September signal in every architecture office is the same: the models sat quiet through August, and now the inbox is full again. Our KI Back 2 Work Camp on 23 September 2026 is built for exactly that morning — the one where the tender is due, the team is back, and the AI you meant to adopt in July is still sitting in a browser tab.

Start where the research does. At ETH Zurich, the Design++ centre has spent years asking a narrower question than the hype: not can machine learning draw a building, but where in a real design workflow it measurably removes hours. Dr. Mikhael Johanes, the Design++ postdoctoral fellow who closed our KI Sommer Camp with a keynote on machine learning in architectural design, framed it plainly — the useful AI is the one that reads your project data, not the one that poses for a render. That is the line we draw straight into your desk.

This is a follow-on, not a first date. As we documented in our KI Back 2 Work Camp notes, the day continues the thread from the three-day KI Sommer Camp we ran 2–4 July 2025 at PAZ Central, Niederdorfstrasse 77 in Zurich — GenAI, LLMs, Python and Grasshopper against a live competition brief, priced at CHF 1,790–2,190 per seat and running 09:00–16:30 daily. The Back 2 Work format is shorter and blunter: one session, 09:00 onward, aimed at the tasks that actually clog the first week back.

←TODAY: In 2026 the bottleneck is not model access — it is knowing which desk task to hand it first. →3012: The offices that survive automate the boring 60% and keep the judgement. Fulcrum: AI only pays back when you already know, in hours, what it is replacing.

What we work on is deliberately unglamorous: cleaning and tagging project data so it is searchable, triaging the repetitive Archicad automation that eats a junior’s Monday, and wiring a small custom assistant that answers questions about your files rather than the internet’s. PAZ-BOX — our Archicad add-on that lets you talk to the model by chat and voice — is on the table where it genuinely fits a task, named as a tool, not a pitch.

Atelier: Every Büro coming back from summer has the same argument this week — whether to adopt AI broadly or ban it quietly while the partners decide. Do neither. The Monday move: pick one recurring task your team did more than five times last month — say, renaming and classifying imported IFC elements — and time it honestly before you automate anything. That single number is what turns an AI experiment into a business case your practice can defend.

Hack: List every element the model already knows, before you automate a single one. The fastest way to scope an Archicad automation is to count what is there, grouped by type, and let the tallest bar tell you where the hours hide. Run this against a live project through the Archicad Python connection — no new plugin, just the API that ships with the software. Read first, script second; the count is your brief.

from archicad import ACConnection
conn = ACConnection.connect()
acc = conn.commands
elements = acc.GetAllElements()          # everything the model holds
print(len(elements), "elements \u2014 group these before you script them")

Who it is for: architects, BIM and project coordinators, and engineers who left July with a plan to bring AI to the office and came back to the same manual routine. You do not need to code to attend; you need one workflow you are tired of. We keep the group small on purpose — the same reason PAZ Academy graduated 39 certified Building System Specialists through a five-month, six-hours-a-week program rather than a faceless cohort: feedback loops only close when the room is small enough to hear each screen.

Sign-up and the exact timing sit on our event page. Bring a laptop and one real task from your own desk — the session is worth most to the person who arrives with a specific Monday problem, not a general curiosity about AI.