Founding a startup with AI in 2026: what changes in practice

I see an important opportunity to make ideas easier to try out. But a tool capable of producing code or preparing a campaign still needs to meet a problem worth addressing. My view is that greater ability to execute increases the responsibility to choose well.

The founder's work shifts

When some execution becomes more accessible, a founder can test paths that previously required more resources. That does not make every task simple or remove the need for specialists. Understanding the customer, setting priorities and recognizing the solution's limits still require attention.

My starting point is to specify before executing. Who will use it? Which task needs to improve? What information is available? What will remain outside the first version? AI can help organize these answers, but it should not supply assumptions where we still need to learn from people.

Productivity needs context

A working paper published by the NBER in 2023 studied the introduction of an AI assistant in customer support and found productivity gains that varied with workers' experience. The result belongs to that work context. Generative AI at Work, NBER.

A METR study published in July 2025, meanwhile, observed experienced developers working in open-source repositories they knew well. With the early-2025 AI tools used in the test, they took longer to complete the tasks. METR study of experienced developers.

To me, these studies reinforce a management question: what happens in our task, with our tools and our quality standard? A sense of speed and a result accepted by the customer are different things. We need to measure the entire process.

A prototype starts a conversation; it does not finish validation

Imagine a founder who wants to help small businesses organize customer requests. A first version can help observe how someone records a request, finds information and decides on the next step. That contact produces better questions. It will still be necessary to discover whether the solution is useful and whether anyone wants to adopt it.

A working demonstration also does not prove the operation is ready. Incorrect data, exceptions, maintenance and review can consume the time saved in building. Before expanding use, we need to understand who monitors failures, what requires intervention and how much it costs to sustain delivery.

What I would put into a new founder's routine

• Choose a real task and record how it happens today, including time, difficulties and rework.
• Define a small first deliverable, with someone able to assess whether the result is correct and useful.
• Compare the entire effort with and without AI. Include preparation, review, correction and tool costs.
• Talk with customers and specialists. A well-framed question can keep you from automating the wrong problem.

I remain optimistic about the possibility of more people being able to build. Access to tools can expand the starting point, but sector knowledge, trust and distribution need to be developed. The initial team size should follow the work and responsibilities, without a universal formula.

That is the combination of learning, building and exchange I want to bring together in the NewHack ecosystem. For someone starting now, I would choose a concrete problem, run a carefully assessed test and seek people who help reveal what the first version still does not solve.