AI and the Physical Economy
The next decade of AI won't be won by the companies with the best models. It'll be won by the ones who wire AI into the physical, regulated, slow-moving parts of the economy.
There's a comfortable story about the future of AI: big labs ship bigger models, developers build apps on top, the world gets smarter. That story is mostly right for the digital economy — code, content, customer support.
It's mostly wrong for the physical one.
Where the real labour lives
Most GDP in developed economies isn't produced by knowledge workers staring at screens. It's produced by people doing things in buildings: processing claims, handling inventory, dispatching technicians, reconciling invoices, translating legal documents into operational tasks.
These are the jobs that generative AI is supposed to transform. But the transformation is stuck — not because the models can't do the work, but because the work sits inside workflows that were designed for paper.
A Swiss property manager doesn't need a smarter chatbot. They need an assistant that can read the email from the tenant, cross-reference the lease database, check the maintenance ticket queue, draft a response in the house tone, schedule a site visit in the calendar, and log everything for the quarterly report. Each individual step is trivial for a modern model. The connection between the steps is the hard part — and that's where builders, not labs, create value.
The wrong metric
Tech discourse measures AI progress in benchmarks: MMLU, HumanEval, ARC. Those benchmarks are useful for research. They are nearly useless for predicting economic impact.
The metric that matters is integration density: how many real-world workflows a given AI system actually touches, how deeply, with what error tolerance. A model that scores 90% on every benchmark but doesn't integrate with SAP, Outlook, Rimo R5, or ImmoTop2 generates zero economic value in a Swiss SME. A model that scores 75% on benchmarks but plugs into three ERPs and writes native Hochdeutsch generates millions.
Who wins
Three groups win the next decade:
1. Model labs — the obvious ones. OpenAI, Anthropic, Google DeepMind. They set the ceiling.
2. Platform providers — Anthropic's Managed Agents, OpenAI's Assistants, AWS Bedrock. They make it possible to ship agent systems in days instead of quarters.
3. Local implementation studios — small teams, usually solo founders, who understand one specific vertical and build custom systems for real clients. This is the category that will quietly create most of the economic value, and the category that receives the least press.
The Swiss angle
Switzerland is unusually well-positioned for category 3. High labour costs mean the ROI on automation is obvious — saving 10 hours per week on email triage is worth roughly CHF 2'000–3'000 per month per employee. The SME density is enormous — hundreds of thousands of companies with 5–50 employees, most of them operating with processes designed in the 1990s. The regulatory environment is strict but navigable — DSG and revDSG force local solutions, which shuts out some foreign competitors.
The winners in this category will not be the biggest or the loudest. They'll be the ones who show up, ship in days, and stick around.
What I'm building
I'm building AI assistants for Swiss SMEs — starting with property management and expanding from there. The first live product (Swiss Immo Assistant) is documented in the rest of this journal. The thesis behind it is everything I wrote above, compressed into one sentence:
*The model is not the product. The integration is.*