What 100 AI leaders say will win in the next era of AI products

Executive overview

AI foundation models are becoming commodities. Every company will soon have access to comparable models, so the race shifts to experience. The builders who win will be those who design AI that feels alive, earns trust, and anticipates needs before users articulate them.

Start with what the technology can now do, then pull toward user need — it's a back-and-forth, not a linear path.

The winners of the AI race will be determined by great user experience.

The five new interaction primitives

1. Chat with anything — persistent, ambient conversation at any scale, applying Metcalfe's law across all products.
2. Semantic resize — any content can be made longer, shorter, more formal, or casual to match the user's current state.
3. Remix — effortless style transfer and cross-pollination across all content types.
4. Format translation — move between text, audio, video, and structured formats with minimal fidelity loss.
5. Attention orchestration — managing multiple agents running at different time scales; the "lobby problem" of what users do while agents work is unsolved.

Building capability-first products

• Conventional wisdom says start with customer need or design vision; in AI it is more effective to start with what the technology can now do.
• The "why now" question matters — being first to identify a new LLM capability creates durable advantage.
• Treat each model as a distinct material with its own properties; play with it directly to discover what it can do.
• Good product is a push-pull between capability and need, not a one-way translation.
• "Ship to learn" — but shipping speed does not equal learning speed; iteration without progression is common.

Designing for real problem-solving

• What users ask for is the tip of the iceberg; surface the underlying goal, not just the stated request.
• Satisfy the explicit ask first — earn the right to help with the broader need.
• Real problem-solving is non-linear: diverge, explore branches, prune, converge. Today's chatbots are too linear.
• Design for long-horizon, multi-session problems, not single-turn exchanges.
• Blank-page friction and mid-task dead ends are the two key moments to solve.

Matching AI capability to real user problems

• Map the top customer problems first, then identify where AI capability addresses them — not the reverse.
• The reverse (adding AI because of FOMO, then finding use cases) wastes iteration cycles.
• Established brands like Grammarly face an additional constraint: any AI addition must preserve existing trust, not just add capability.
• Moving from post-writing correction to pre-writing composition assistance is an example of a capability-mapped expansion.
• Studying user prompts as a research method is now essential — prompts are the new button clicks.

UX as the ergonomics of AI

• Poor UX with powerful AI is like a heavy-duty drill with a terrible handle.
• Decades of physical ergonomics research apply: safety, comfort, efficiency. AI UX is at day one of that journey.
• Outcome satisfaction — not just conversation completion — must be measured.
• Personalisation that feels like care (not data extraction) is the goal; the interaction with a "magic dog" AI character showed hyper-personalisation can emerge from a two-line exchange.
• The right question is not "do users trust AI?" but "does this specific product make them feel seen and heard?"

Responsibility and the next generation

• AI amplifies everything — including mistakes. Getting the foundations right matters more than in previous platform shifts.
• Children interact with AI differently from adults; leaving room for that exploration may surface things adults cannot anticipate.
• Access is equalising: geography and background matter less when AI tools are universally available.
• Founders need a beginner's mind — trying to control every variable closes off larger opportunities.