Working model · last revised 22 August 2026 · Swiss market only · figures unaudited
Open questions & risks
Self-serve conversion — materially de-risked by going hybrid
The model previously assumed zero sales reps across every segment, including CHF 32,000 mid-market
contracts. That was its weakest assumption. Now hybrid: self-serve carries the 3,840
small and micro customers (CHF 8.7k ACV), while five reps carry the 960 medium customers holding 48.5%
of ARR. Cost CHF 0.95M/yr against CHF 1.4M of NPV — cheap insurance on half the revenue.
What remains untested is self-serve conversion at the small end: CHF 8,664/year is
still well above the CHF 10–50/seat range where self-serve normally operates. A real trial-to-paid rate
is the number that settles it — and the first signed customers are the first chance to measure one.
Engineering coverage — resolved
Raised 3 → 5 engineers (1-in-5 on-call, cover for illness and departure). Cost: CHF 0.4M of NPV,
78.1 → 77.7. Cheap insurance against the single-point-of-failure risk on a business-critical platform.
79% EBITDA margin will get the model dismissed
Rule of 40 ≈ 139. The most profitable software companies at scale run 35–45%. Recommend presenting
a deliberately conservative margin externally and treating the lean structure as unstated upside.
CHF 10,000 ticket size may trigger a FIDLEG prospectus
Public bond offerings are generally exempt only above CHF 100,000 denomination. Needs legal review
before any marketing — Domenig & Partner.
Seats per customer — 5.5, on a two-tier basis
The first pass scoped this to business-administration roles only (25% of SME headcount → 3.0
seats). That was too narrow: AtlasOS handles mail, calendar, documents and projects, which is not
confined to admin staff. Rebuilt as two tiers — a core of admin/management roles (25%
of headcount, 35–50% adoption) plus an extended tier of other regular PC users (~30%
of headcount, 15–25% adoption, lower because the marginal user is a lighter-touch case). Blended across
the customer mix: 5.5 seats, band 4.0–7.3.
Still an assumption, and the most valuable one to kill. The adoption rates are
pure judgement — no public benchmark exists for them. Capability ("anyone who uses a PC could use it")
is not the same as willingness to pay CHF 200/month per head, and that gap is where this number
actually lives. The EasySystems CRM plus one pilot customer would settle it.
Developer segment is real but small — NPV CHF 2.2M
Now modelled (see scenario G). It is not the large upside previously assumed here.
TAM is 36.5M professional developers, but the seat only prices at ~CHF 32/month — AtlasOS carries no
inference cost, so it sells as a supervision layer on top of a Copilot or Claude Code seat,
not as a replacement. Base NPV CHF 2.2M (range −0.7M to +13.2M), and every explicit year is
cash-negative. Binding constraint is distribution, not product: 330,000 free users by 2031 from an
unknown Swiss vendor with no sales team. Real risk of commoditisation if GitHub or Anthropic ship
orchestration into seats developers already own.
International expansion is worth ~CHF 107M — engineering constraint disputed
Now modelled (scenario H). Austria first, not Germany: Austria has 601,001 SMEs
— 0.96× Switzerland, not a rounding error — is German-speaking, and BMD alone covers ~80% of Austrian
tax advisors, so 2–3 integrations reach most of the market. Germany is 5.2× larger but fragmented, and
its real gate is not code: DATEV serves ~80% of German tax advisors and requires
25 live customers plus 3 references before Marktplatz listing — calendar time no
engineering can compress. Start that onboarding in 2027 regardless.
The research called headcount the binding constraint — that estimate does not survive
contact with the evidence. It put Swiss maintenance at ~51 person-months/year (4.25 FTE just to
stand still) and concluded expansion needs 11 engineers in 2027 rising to 28 by 2031. But the existing
~35 skills and ~10 live integrations were built by one person, using the agent tooling
this company sells. A benchmark derived from conventional teams cannot be applied to a company whose
product exists to invalidate it — the same error that earlier pushed the staffing plan to 160 people.
Working estimate: 5–7 engineers, not 28.
Two things genuinely do not flex, and neither is about productivity. On-call is rota
arithmetic — five gives a 1-in-5 rotation with cover for illness and departure, which is why
5 is the floor. And DACH integration work is real: a specialist European
accounting-API vendor lists Germany as DATEV/sevdesk/Lexware and Switzerland as bexio/Abacus with
zero overlap, so nothing carries across at the connector layer however fast it is written.
Reuse from Switzerland is near zero at the connector layer — Abacus/Bexio/Topal share nothing
with DATEV/BMD. Compliance is a tailwind though: Germany's B2B e-invoicing mandate lands 1 Jan 2027
(>€800k turnover) and 1 Jan 2028 (all), forcing a software refresh exactly in the entry window.
Scenario H is now broadly comparable
The international model assumes 5.0 seats per customer, against the base case's 5.5 — close
enough that its CHF 107M can be read alongside scenario D without rebasing. It was briefly incomparable
while the base sat at the narrower 3.0-seat figure. Its real constraint is headcount, not seats.
Key-person risk — substantially mitigated, not eliminated
The skill library and integration layer were built by one person, which is both the efficiency
proof and the obvious concentration risk on a 10-year bond. Mitigation already in place:
an existing 7-person software team within the group has full access to the codebase and the system is
documented — so continuity does not depend on one individual, and the same team is available as surge
capacity for the DACH integration push.
Two things still need closing before a bondholder treats it as settled. The team is
shared, not dedicated — it delivers customer projects, so it de-risks continuity rather than
adding throughput, and the plan should say which it is counting on. And access plus
documentation is not the same as operating experience: the test that converts this from a
claim into evidence is whether someone other than Michael has shipped a production change to AtlasOS
and carried an on-call rotation. If the team sits in a different group entity from the bond issuer,
an intercompany service agreement makes the dependency contractual rather than informal.
EU AI Act Art. 50 is already in force — since 2 August 2026
Not a future deadline. Any AI system interacting with people must disclose that it is AI, and
AI-generated content must be machine-readably marked by 2 Dec 2026. The Digital Omnibus deferred the
high-risk rules to Dec 2027 but did not touch Art. 50. Applies to a Swiss vendor whose output
is used in the EU. Also gate the agent away from HR and creditworthiness use cases, or it becomes an
Annex III high-risk system by Dec 2027.
EU data-transfer position is strong — and is a sales argument
Swiss adequacy holds in both directions (EU→CH via Decision 2000/518/EC, confirmed by the
Commission's 15 Jan 2024 review; CH→EU via FADP Annex 1), so no SCCs are needed either way. As a
processor for German controllers, AtlasOS is not directly caught by GDPR at all — only
contractually via Art. 28. The one thing that undoes this is a US model sub-processor
touching customer books, so an EU/CH model option should be architected in early.
Reverse-charge VAT treatment unconfirmed
Cross-border B2B SaaS is assumed to be reverse-charge, so no German VAT registration. Needs a tax
advisor before it is relied on.
Signed customers are not in the model — and their terms are unrecorded
First external customers have signed (Michael, 22 Aug 2026). Nothing about them has reached this
model: not the contract values, not the seat counts, not the start dates. Two consequences. The forecast
is unchanged and still begins from zero in 2027, so early billing is unmodelled upside. And the single
most valuable calibration available — real seats per customer against the modelled 5.5, which moves NPV
more than any other input — is sitting in those contracts and is not being read.
TODO: verify — who signed, contract value, seat count, go-live date.
Placement is now assumed at 2–4 months, and the bridge converts rather than repays
Raise timing: CHF 15M over the next 2–4 months, with a CHF 250–750k convertible
loan in the meantime at the bond's own 6.0%, converting into the bond at placement, and issued in the
bond's own CHF 10,000 denomination so conversion leaves no fractional ticket to settle. That settles what
was open before. Note the coupling: if the bond denomination changes — the FIDLEG
question below is the likeliest reason it would — the bridge denomination has to move with it, or the
1:1 conversion stops working. It converts into debt, not equity, so the cap table is untouched; and because it
converts rather than being repaid, a placement does not have to fund a cash repayment out of its own
proceeds. The bridge is sized inside the CHF 0.9M peak funding requirement, so it is not new money in the
plan either.
What the conversion structure does not fix is a failed placement. If the bond is
never placed there is nothing to convert into, and the loan stands against a business whose own cash flow
does not turn positive until 2028. The 2–4 month window is also an assumption, not a commitment — it is
short for a CHF 15M placement, and the FIDLEG question below is the most likely thing to extend it.
Sizing at the low end until the placement is credible remains the conservative read.
Marketplace product line has no model
Every figure here is AtlasOS only, so none can be compared against the full CHF 15M raise.