AI-First Product Development: What Founders Get Wrong

A two-person startup now has access to the same AI models as a global software company. Writing code has never been cheaper or faster. That raises an uncomfortable question: if everyone can build faster, what actually gives a startup an edge?

This question ran through a recent Swiss Startup Association webinar, AI-First Product Development, in partnership with Miro and Holycode. Tony Beltramelli, Head of AI Research at Miro, and Laurent Decrue, Co-Founder and Chairman of Holycode, compared how their teams build with AI today. Beltramelli set the tone early: “We all have access to the same AI these days. There’s no longer any secret sauce.”

Their answer was not a tool. It was everything around the tool: the workflows, the review process, the people and, increasingly, the parts of a business that AI cannot copy. Four lessons stood out.

Speed Comes From the Process, Not the Prompt

Most developers use AI the same way: prompt, copy the code, check it by hand. The gain sits only where code gets typed. Holycode’s own estimate puts this at around 15 to 25 percent faster coding, on work that makes up roughly a third of a project.

Holycode has instead redesigned the whole chain. Engineers trigger fixed commands that start structured, agent-driven workflows, and the work then passes through an automated review platform in which several AI agents, running on different models, check architecture, security and best practice before anything returns to a human. Decrue compares it to “a very senior developer that is not tired” and says it has cut review time by around 95 per cent. In his company, a delivery team that once needed seven people now runs with a project manager and a developer. These are the company’s own figures, and it took nine months of trial and error. But the principle applies to any startup. “Everything happens in a chain, and it’s not you chatting with the interface anymore,” Decrue said. 

A human still approves every client release. Miro, which serves customers in sensitive industries, keeps more human reviews but leans on AI to filter weak pull requests and on automated testing. The lesson for founders is simple: an AI licence for every engineer is a starting point, not a strategy. The real gains come from standardising how AI is used and automating the checks, so that review does not become the new bottleneck.

When Building Gets Cheap, Discovery Becomes the Bottleneck

Remove one bottleneck and another appears. Once Holycode had sped up delivery and review, Decrue saw the constraint shift “to the front of the process”: understanding what the client actually needs. 

Now an AI workflow researches the client and its competitors before the first meeting. Two workshops test the assumptions, and the system then produces user stories, architecture documents and a clickable prototype. “Six months ago it was nowhere,” Decrue said of prototype quality.

Beltramelli called prototyping the biggest unlock AI has brought to Miro: putting an idea in front of customers and getting feedback within a day. But he warned against confusing output with progress. “It’s easy to build, but it’s still hard to know what to build and how to build it.” If an MVP can be built in days, the expensive mistake is no longer building slowly. It is building the wrong thing quickly. The time AI saves belongs in customer conversations, sharp problem definition and fast validation.

Your Engineers Are Becoming Managers

Both speakers were candid about the human side. When Holycode made AI mandatory, some developers resigned. Many who stayed miss the deep, focused work they grew up loving, as the job moves towards architecture, requirements and communication. It is also more tiring. Agents work around the clock and ask for decisions late in the evening, and some employees struggle to switch off because an answer tonight means the system can work overnight. “We are all becoming managers without being trained as managers,” Beltramelli said, pointing to constant multitasking as a major cause of burnout.

At the same time, expertise matters more, not less. Beltramelli cited research showing that AI output is only as good as the knowledge of the person directing it. An engineer who understands scaling, security and the customer’s domain gets far more from AI than an average one, becoming “like a senior manager” guiding a team of virtual developers.

For founders, that means hiring for judgement and domain depth, setting clear norms about availability, and not assuming that faster tools automatically make a healthier team.

When Anyone Can Build, Distribution and Trust Become the Moat

If software is easy to copy, what protects it? Beltramelli was direct: “In an age where everybody can build something with AI, the biggest edge you have as an organisation is not your technology, it’s your distribution.” He pointed to Figma and Slack, products people stay loyal to even when alternatives exist. Decrue added that while new apps have exploded, usage concentrates on a few winners customers already trust.

In B2B, trust takes a different form. Larger companies rarely replace a vendor with something built over a weekend, Decrue argued, because they need someone accountable for security and compliance.

Both also saw opportunity in less obvious places. An audience member noted that many companies, particularly in biotech, still run on spreadsheets and paper. For them, a simple product, perhaps with a voice interface, competes with paper rather than with Silicon Valley.

Key Takeaways for Founders

• Standardise how your team uses AI. Individual prompting habits do not scale.
• Automate review and testing early so quality keeps pace with output, and keep a human on final approval.
• Spend the time AI saves on discovery and validation. Building the wrong thing is now the bigger risk.
• Hire for judgement and domain expertise, and protect your team from always-on agent work.
• Treat brand, distribution and trust as your defensible assets, not your code.

Final Thoughts

For decades, the hard part of a software startup was building the product. Engineering was scarce and expensive, and that scarcity protected those who had it. That protection is fading fast.

What stays scarce is harder to automate: knowing which problem is worth solving, the judgement to steer machines in the right direction, and the trust of customers who have too many options. None of this is new. These are the fundamentals of building a company, and AI has simply removed the excuse to neglect them.

Beltramelli expects people to stay essential in deep tech and regulated fields, where general AI models still struggle. Switzerland has plenty of both. The founders who do best may be those who use AI to build faster, then spend the time they save on what AI cannot do for them.

About the Webinar

This article is based on the webinar AI-First Product Development, hosted by the Swiss Startup Association in partnership with Miro and Holycode on 22 September 2026. Tony Beltramelli, Head of AI Research at Miro, the visual collaboration platform, and Laurent Decrue, Co-Founder and Chairman of software development company Holycode, discussed how their teams build products with AI, how code review and product discovery are changing, and what startups can still defend when anyone can build.

The full webinar replay is available in the Swiss Startup Association Education Library, free for members. Not a member yet? Join the community and get access to expert sessions like this one. 

Subscribe to our newsletter to hear about upcoming webinars and events.