There is a particular quiet in a Swiss atelier the week the holidays end. The out-of-office replies come down, the competition deadline that felt distant in July is suddenly six weeks out, and the good intention to “finally sort out the AI thing” resurfaces — usually to be buried again by Thursday. We built PAZ KI Back 2 Work, running 23 September 2026, for exactly that week.

Start where the real leverage is, not where the marketing points. The clearest picture of what “AI at the desk” should mean is not a chatbot — it is a model that stays honest with itself. From Archicad 24 onward, the software derives a Structural Analytical Model in the background from the physical model and keeps it current as the design moves, collapsing beams and columns to 1D members and slabs and walls to 2D members, then exporting to solvers like SCIA Engineer or ETABS through the Structural Analytical Format (SAF) rather than surviving a lossy IFC round-trip. That is the standard the session holds AI to: coupled, auditable, correcting itself when you change your mind — not a second source of truth drifting away from the first.

←TODAY: On 23 September 2026 we sit a room of returning architects and engineers down and wire one AI move into their actual Monday workflow. →3012: The offices that treated AI as a coupled model, not a novelty, are the ones whose archives their heirs can still read. Fulcrum: A tool only pays off when it stays honest with the model it feeds — that is as true of an LLM as it was of the analytical model in Archicad 24.

Who it is for is narrow on purpose: architects, BIM and project managers, and structural or civil engineers who want AI running through their work, not sitting beside it in a browser tab. We have run this shape before. Our KI Sommer Camp (2–4 July 2025, 09:00–16:30 daily at PAZ Central, Niederdorfstrasse 77 in Zürich, at CHF 1,790–2,190) worked on real competition briefs and closed with a keynote from Dr. Mikhael Johanes, a Design++ postdoctoral fellow, on machine learning in architectural design. The KI Bootcamp we ran 24–27 September 2025 — three 30-minute intensive sessions plus a one-hour bonus masterclass on day four, at 49 CHF — proved the same thing: the participants who leave with one repeatable move outlast the ones who leave with a list of forty tools they will never open again.

So Back 2 Work teaches one move you keep. We name our own kit factually where it fits: PAZGPT, our chat-and-voice interface to Archicad models, is the kind of Archicad automation the session grounds itself in — not a product to admire, a workflow to adopt. The session page carries the date, the outline, and the sign-up.

Atelier: The Büro shift this September is not “adopt AI” — it is deciding which single task the whole team agrees to route through a model first, so you learn the failure modes on something low-stakes. Your Monday move: pick one recurring, boring, high-volume task — renaming exported views, triaging the shared inbox, tagging drawing revisions — and give exactly that task to AI for two weeks before you touch anything a client sees.

Hack: Rank your first-week-back task list by urgency without reading all of it. This calls a local model to score each line 1–5, so you open the pile already sorted. Swap the endpoint for your provider of choice.

import ollama
for t in open('inbox.txt').read().splitlines():
    r = ollama.chat(model='llama3', messages=[{'role':'user','content':f'Rate urgency 1-5, number only: {t}'}])
    print(r['message']['content'].strip(), t)

Run it, pipe the output through sort -rn, and the top three lines are your morning. That is the whole doctrine of the session in four lines: one intention, one model, one honest result you can check. Block the morning of 23 September now — the desk-return only comes once a year, and it is the cheapest week to change a habit.