a company of one.

The agentic-AI tooling is real but uneven — and in Germany the legal form decides as much as the tech. A top-down survey of what's shipping, what's hype, and what's lawful.

the build is the easy half. the company is the hard half.

One reading for the technical founder shipping product; one for the same person deciding how to incorporate, get taxed, and stay compliant in the EU — then a staged plan that puts them in order.

four figures that frame the bet

Two on what the tooling can and can't do, two on what it costs to be wrong.

36%

Success of a 20-step workflow at 95% per-step reliability. Error compounds.

41–87%

Failure rates across 7 multi-agent frameworks (MAST, NeurIPS 2025).

~48%

Combined tax on distributed GmbH profit; ~30% if retained.

40%

Of agentic-AI projects Gartner expects cancelled by end of 2027.

übung 1 — die linie

Why "95% reliable" is not reliable: per-step success compounds. The line falls from 100% to about 36% by step twenty. This — not model quality — is the core unsolved problem.

Compounding error curve A line chart showing the success rate of a workflow at 95% per-step reliability, falling from 100% at step 0 to roughly 36% at step 20. 100% 80% 60% 40% step 0 10 20 36%

übung 2 — die fläche

Maturity by company function. Solo leverage is concentrated in engineering, content and support; finance, legal and accountability remain thin — the parts a human still has to own.

Strategy / CEOleast automated forming
Product / engineeringmost mature today shipping
Operations / coordinationorchestration maturing maturing
Customer supportstrong, with caveats maturing
Sales / marketingcontent well-served maturing
Business-process automationbroad but shallow partial
Finance / CFOaugmentation, human sign-off thin
HR / legal / compliancethinnest, high-risk thin

the tooling is real but uneven.

Part 1 — the available-today win is a human firmly in the loop on anything irreversible.