Leaderboards don’t ship products. Customers do.

Google's Gemini 4 Argon now sits at the top of the Vals Index, and almost nobody can use it yet. So nobody can say whether it's actually useful. That's the trouble with leaderboards: they measure a model, not whether it helps anyone.

The AI that's winning right now is built by people who use it obsessively and fix what breaks. That's the benchmark that matters: the customer. And it's the same test for any AI your own company builds.

Who is each model for?

Look across the major labs and most have a clear customer in mind.

• Anthropic focuses on businesses and helping people do professional work well.
• OpenAI has a very large consumer product and a growing enterprise business.
• Meta is all in on consumers, with Muse as its latest push.
• Google earns a lot from compute, serves billions through search, and has Gemini. What's less clear is where a frontier model like Argon fits: in search, in Gemini, and why it would be more useful there than what's already shipping.

More strong labs in the game is good for everyone; more players tends to mean better and safer products. But a model that tops a benchmark without a clear customer hasn't yet passed the test that counts.

Build for change, not a roadmap

Traditional software shipped on a roadmap, with updates every month or two. AI products don't work like that anymore. The intelligence underneath keeps improving, sometimes every week, and the product has to grow with it while helping people understand what's changing for them, in a hopeful way rather than a fearful one.

That's a hard product problem, and a lot of money is being spent on it. None of it works without starting from a clear picture of who the customer is and what they're actually trying to do.

Use what you sell

The best AI products come from teams that use them constantly, notice what goes wrong, and make those failures matter. That's as much a business discipline as an engineering one.

Using AI for real errands shows quickly where each tool shines. One agent may be slower but very good at pulling data from many connected accounts and holding a lot of it together; another may be faster at working through websites and taking action. In practice you end up using both, each where it's strongest. That kind of judgement only comes from use, and companies need it about their own products: which jobs they truly handle well, and which they don't.

Beyond the chat box

Most AI products since ChatGPT have borrowed its design: a chat box, then a toolbar, then a panel for documents. That's only the beginning of what AI products can look like. Some newer ideas point in other directions:

• A walk in the woods. An app that uses image recognition to help children name the trees and plants around them, using AI to get them outside rather than adding screen time.
• Plans with friends. An assistant that joins a group chat and helps coordinate the next dinner or film.

Not all of these will work, but there's far more room than the chatbot pattern suggests, and there are billions of people with different needs.

Small moments of care

Customer obsession often shows up in small moments. Picture a parent asking an assistant for perspective after a child's minor accident. A good answer is careful not to play doctor, but it also notices how the child might be feeling, and suggests that this probably isn't the moment for a lecture. That bit of empathy in a stressful moment is what customer obsession looks like in practice.

The bottleneck has moved

In 2024 and 2025, most of what went wrong with AI came from the models. Increasingly, the rough edges come from the product: how the AI reaches people, fits into their day and hands work back. People already use agents for taxes, travel plans, school schedules and making sense of medical history. The models are good enough for much of that; the experience around them often isn't yet.

So the next year in AI is likely to be about products, not benchmarks. When the next model tops a leaderboard, the useful questions are simpler: who is it for, where do they meet it, and what does it do for them that nothing else does?