'Certification is too expensive for startups' — the math says otherwise

One of the most common objections in Physical AI startup circles is that safety certification is too expensive and too slow. The argument typically goes: we're a startup, we need to move fast, and we can't afford to spend six to twelve months and significant budget on a certification process before we have revenue.

The math doesn't support this argument. It actually argues the opposite.

Certification bodies don't publish fixed price lists — each project is scoped individually. What is public are the benchmarks for plain machinery CE marking: roughly €3,500–€16,500 including risk assessment, with notified body fees of about €3,000–€8,000 per submission. A full functional safety programme for an AI-enabled safety function sits well above that. Risk assessment, safety architecture design, FMEA, fault tree analysis, hardware metrics calculation, software lifecycle compliance, and third-party assessment by TÜV, UL or exida together land in the tens to low hundreds of thousands of euros, depending on system complexity, target SIL, and scope. For a focused product with a well-defined safety function, the investment is a fraction of a single product recall.

Now compare that to the cost of not certifying.

The third annual Pulse of Quality in Manufacturing survey — 2,263 managers and directors at mid-to-large manufacturers across the US, UK and Germany, fielded in the first quarter of 2026 — found that 75% had experienced a product recall in the previous five years, and that 59% said each recall cost between $10 million and $49.9 million. That share was 39% in the 2024 survey and 48% in 2025. The cost of getting it wrong is not stable. It is climbing roughly ten points a year.

These are general manufacturing figures. For safety-critical robotic systems the exposure is concentrated. Robert Williams was killed on 25 January 1979 at Ford's Michigan Casting Center in Flat Rock, when a one-ton transfer vehicle in a Litton Industries parts-retrieval rack struck him from behind. In August 1983 a jury awarded his estate $10 million — the largest personal injury verdict in Michigan history at the time — finding that basic safety measures such as alarms and automatic shutdowns were absent. The award was raised to $15 million in January 1984 before Litton settled for an undisclosed sum. The system's only stop function depended on a human remembering to lock it out. That is one incident, one lawsuit, one plant. Today, with fleet-deployed robotic systems, the exposure scales with every unit in the field.

From 20 January 2027, when Regulation (EU) 2023/1230 replaces the Machinery Directive, machines and safety components whose safety functions rely on fully or partially self-evolving machine learning fall under Annex I Part A. Notified body assessment — EU type-examination or full quality assurance — becomes mandatory, and self-declaration is no longer available. Since the Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026, machinery sits in Section B rather than Section A of the AI Act's Annex I, which makes the Machinery Regulation the single technical route for AI safety functions in machines rather than a second assessment layered on top of the AI Act. A company that ships into Europe without proper conformity assessment doesn't just risk a fine — it risks having its product pulled from the market entirely.

The framing matters. Certification isn't a development cost. It's insurance — but priced as a line item rather than a premium. And it isn't a one-off: certificates carry surveillance, and under 2023/1230 a software change that alters a safety function can make you a new manufacturer with a new declaration of conformity. The startup that budgets for that still spends a rounding error against a recall in the $10M–$50M band. The startup that skips certification faces open-ended liability exposure from the first unit deployed.

The expensive choice is skipping it.