Stop Bolting AI On. Start Building AI-Native.

01 /The Concept

User adoption is where most products quietly die. Sammy exists to close that gap – an AI agent that learns how a product actually works, then uses that knowledge to drive onboarding, documentation, and retention.

The key insight is how Sammy learns: it explores the product the same way a first-time user would. It navigates to the site, clicks every link and button, sees what happens. Every path through the product gets recorded into Sammy's database, which it then uses to do the following.

Documentation generation is the starting point. Sammy produces written descriptions of how users complete tasks – as text documents or screenshot-illustrated walkthroughs. Alternatively, a developer can sketch a workflow as a bullet-point outline and Sammy expands it into a fully fleshed document with relevant screenshots and connected prose.

Those same descriptions translate directly into live in-product tooltips, guiding users in real time as they navigate. Different user types get different guided paths, tailored to what they're typically trying to accomplish.

Sammy can also convert any workflow into a demo video – useful for sales decks or SEO landing pages where each page shows how the product solves a specific problem.

There's a quietly powerful fourth capability: while exploring the product, Sammy surfaces bugs and broken flows and flags them for the developer. The bugs get found before users report them. And because the discovery happens in the context of real user workflows, it's far more meaningful than abstract test coverage.

Sammy re-crawls the product on a recurring schedule, which means documentation and videos automatically stay in sync with new releases. It also catches regressions introduced by updates.

Pricing is usage-based rather than a flat subscription – charged per generated document, video, or discovered bug, starting at $0.30 per action. Sammy's own calculator estimates that a mid-sized product with four-plus screens, multiple user flows, and 10,000 users translates to roughly $3,500/month.

Sammy Labs, the company behind the product, is currently in Y Combinator's accelerator and has raised $500K. The platform launched on the YC site two days ago.

02 /Why It Matters

A few days ago, a story circulated among developers about someone who wanted to build an AI support agent for their product. They quickly realized the standard approach – loading documentation and FAQs into a knowledge base – had a fundamental problem: you have to create that documentation first. And every product update makes part of it stale, requiring manual regeneration.

So the developer built an alternative: a support agent that answered questions by referencing the actual current source code and each user's live data. Roughly speaking, it "ran" the code in its head and answered based on the output. Surprisingly, this worked – though keeping it from leaking source code and personal data back into the chat took some effort.

Conceptually, this is the same principle Sammy operates on: both use the most authoritative and current source of truth – the actual product. Sammy reads the front-end experience through a browser; the support agent reads the back-end logic through code. Different windows onto the same reality.

The two approaches could plausibly be combined in a future version – an agent that both answers user questions with screenshots and video context, and traces the root cause of front-end bugs back to the source code.

Aristotle laid out "first principles" thinking roughly 2,400 years ago: when reasoning through a problem, strip back to the most fundamental underlying truths and build up from there, rather than inheriting whatever intermediate assumptions others have accumulated.

In software, most AI tools are still layered on top of legacy product architectures. Knowledge bases for support are an example – they exist because humans can't read source code on the fly to answer every ticket. But AI can, which means the intermediate workaround is obsolete.

The same logic applies to CRM. Today, CRMs capture what salespeople choose to write down, filtered through their individual judgment and biases. An AI-native CRM would instead store raw transcripts of calls, emails, and video meetings – plus product interaction logs – and generate reports directly from those primary sources, not from what any one person decided to summarize.

That's exactly what the startup Day.ai built ([related review](/review/chtoby-pobedit-nuzhno-peredelat)), raising $4M in a Sequoia-led seed round last summer. Day.ai explicitly calls its product "AI-native" – meaning the architecture was redesigned from scratch around AI capabilities, not retrofitted with AI features on top of an old structure. The distinction matters more than it sounds.

03 /Opportunities

The broad direction: build platforms and tools designed from first principles rather than grafted onto legacy architectures.

What makes this moment distinctive is that AI can now process large volumes of raw primary data and return competent, actionable outputs. That capability eliminates the need for the intermediate layers that previously existed because humans couldn't handle raw data at scale – knowledge bases, manual CRM entries, documentation written by hand. Those intermediaries were workarounds. They can now be replaced with direct-from-source reasoning.

The question worth asking: in your domain, what are the "knowledge bases" – the intermediate summaries that exist only because raw data used to be too hard to process? Those are the places where AI-native rearchitecting will create the biggest step changes.