What is an AI-first product?

A friend and I were deep in editing one of my essays. At the end, he did a really great job distilling his editorial comments into a clear set of action items. This distillation was nowhere in the AI recorder’s generated meeting notes. It was pure slop. Moreover, it got a bunch of names wrong (for example, “Aaron” instead of “Erin”). It didn’t really understand what the meeting was about either, so most of the notes landed as either wrong or unhelpful.

A trained human admin wouldn’t have made these mistakes. They would’ve intuited what was important, had enough context about the named entities to transcribe them correctly, and would’ve recognized how to make the notes most legible to me.

I’d have felt seen by a trained human in ways that existing note-takers can’t emulate. More broadly, most AI products I use in my day-to-day workflow can’t see nor learn about me over time. They see a thin archetype of me, not me. They either churn out slop or shove me straight into an uncanny valley.

It’s tempting to conclude that the latest models are incapable of such collaboration. But today’s frontier models are more than capable of powering incredible experiences. Product builders lack the right conceptual frameworks for thinking about AI-first products.

Errors in deterministic software usually show up as bugs you can fix. In contrast, errors in stochastic software (i.e. LLMs) show up as misunderstandings you can only resolve iteratively. An AI system often fails at its task because it lacks relevant context and makes a bunch of implicit assumptions that turn out to be wrong. There’s no deterministic method of a-priori resolving such misunderstandings. It’s a necessarily iterative process. It’s impossible for both users and product builders to write complete specifications ahead of time that anticipate all such misunderstandings. For example, even the user might not be able to enumerate every situation ahead of time in which the voice recorder should highlight or deprioritize a sentence. Taking this seriously implies a profound shift in how AI-first products should be designed. Rather than treating the model as a static component that either works or doesn’t, the entire product is framed as an evolving conversation of testing value between the system and its users.

This insight can be formalized into three criteria for whether or to what extent a product is AI-first.

1. The product’s core user journey relies on LLM inference calls.

2. The product is in an ongoing process of surfacing and resolving misunderstandings between itself and the user. It’s mindful of the tolerance window of this specific user for this specific interaction, given the type of workflow that the user is engaged in.

3. The product maintains and uses a rich, long-term memory of the user and their social milieu in service of value creation.

The voice recorder’s frictions are easy to place when re-examined through the lens of these criteria. The recorder clears the first criteria because it already uses AI in its core loop. But it totally fumbles the second and third. It doesn’t have a good way to surface or resolve misunderstandings about what the meeting was really about, and almost no memory or orientation toward my actual world, so I end up with slop instead of something that feels like notes from a Chief of Staff.

These criteria point to a very different worldview and corresponding ontology for products that were simply impossible pre-LLMs. In this worldview, the system stops acting like a static tool that executes a formal specification (i.e. computer program). Instead, it behaves like a dynamic process that formulates, tests and falsifies hypotheses about what’s most valuable to its users based on evidence.

Imagine a voice recorder that, after three months, knows your standups need ticket numbers but your 1:1s need emotional subtext. That notices when a conversation shifts from routine to existential and adjusts its summary accordingly. That learns your threshold for clarifying questions and stays just under it. That tracks not just what you said in meetings but how your priorities have shifted across them, and uses that to anticipate what you’ll find valuable before you ask. In that world, the product is increasingly made of feedback loops rather than static features. It uses the flexibility and rapid improvement of frontier models to keep probing for opportunities in value creation.

Zooming out shows that the same pattern applies to other AI products. A sales tool that is AI-first keeps updating its sense of which relationships and deals are truly consequential. A coding assistant that is AI-first keeps learning your codebase, your style and your team. As more people use these systems, they can generalize in two directions simultaneously. That is, they build personal models of individual users and population-level models of value, then test which patterns actually transfer. They increasingly become scientists in the discipline of economic value creation with integrity.

Satisfying all three of these product criteria is profound. It yields a system that feels “alive” and “responsive” because it’s actively testing for what a user finds valuable. The most valuable AI-first product for an individual is one that keeps trying to fit itself to that user’s evolving conception of value over the longest possible time horizon. The same holds at the population level. The most valuable product is one that quests for value for an entire population over the largest possible horizon while still seeing each individual in front of it.

We’ve barely begun to scratch the surface of what all of this means. AI-first products sit on a spectrum based on the degree to which they satisfy the three criteria. All three criteria are interlocking and interpenetrating with each other. Rapid model improvements will yield possibilities that are difficult to imagine right now.

Will you build products capable of absorbing these improvements? Or will you keep bolting LLMs onto product surfaces that feel alienating?

Please email varun [at] doubleascent.com if you’re actively building AI-first products. I’d love to hear and learn from you!