Measuring Everything Means Nothing If You Can't Act on It

01 /The Concept

Making good product decisions requires understanding how users actually behave inside the product – where they move confidently, and where they stall or give up.

Aggregate statistics come from tools like Mixpanel or Amplitude, but those platforms don't let you investigate what individual users are doing. Session replay tools fill that gap – but insights from watching individual sessions have to be manually synthesized in your head.

The result: with conventional analytics tools, it's genuinely hard to answer meaningful questions like "What share of users abandon the product while waiting for this slow operation to complete?" or "How much more frequently do paying users engage with this feature compared to free users?"

Human Behavior built a tool that analyzes every user session to deliver substantive answers to exactly those kinds of questions.

Setup is simple: either connect the Human Behavior SDK to your service to start capturing sessions, or point it at an existing session data store from PostHog, LogRocket, or HotJar.

From there, the AI processes the sessions, automatically extracting key moments and assembling them into user narratives – for example: "user added item to cart, navigated to checkout, encountered an address validation error, clicked submit several more times anyway, returned to the product page, visited the support page without taking any action, and then left the site with the item still in the cart."

That narrative database can be queried in plain language. Ask it to surface sessions where "users abandoned cart at checkout," or "spent more than two minutes on a product description page," or "clicked the upgrade button but never completed the upgrade." The AI returns a matched list of sessions, with human-readable summaries and access to the underlying recordings.

Pricing starts at $500/month for up to 15,000 sessions, with 30-day retention. Higher volumes or longer retention require a direct conversation with the team.

Human Behavior is currently in the Y Combinator accelerator and posted its platform launch to the YC site yesterday.

02 /Why It Matters

Interestingly, the previous Y Combinator batch included a similar session analysis tool from Decipher AI ([related review](/review/po-nastojashhemu-proryvnoj-produkt-ne-dolzhen-uluchshat)). Since that earlier review, Decipher has pushed further – it now features a dashboard where the AI proactively surfaces problems it detects in user sessions, from software bugs to unexpected drop-offs.

"Further along" is the right framing because Human Behavior is heading in the same direction. The platform's stated goal is to surface all "aha moments" in user behavior that can drive product optimization – friction in onboarding and activation, adoption patterns for new features, upgrade flow drop-offs, and user reactions to product experiments.

The platform just doesn't yet have a polished way to surface those insights automatically. But it's getting there.

And this is a broader trend reshaping analytics as AI enters the space: the shift from data-collection systems that require humans to draw conclusions, to insight-generation systems where the AI does the synthesis and delivers the findings directly.

Retellio ([related review](/review/podkast-sozdannyj-lichno-dlja-tebja)) is a good example from an adjacent domain. The startup raised $1.3 million in its first round last December. Its AI analyzes customer support calls, identifies the most important recurring themes, and produces short podcasts summarizing the insights – with supporting quotes from actual calls. Retellio pushed one step further still: not just surfacing insights, but thinking about how those insights get delivered to the people who need them.

As Retellio puts it, "insights mean nothing until someone hears them." And that's a fair point. How many people reliably open dashboards on a regular basis, even when they suspect something useful might be there? Listening to a 30-minute summary podcast during a commute or a workout, though – that's a much easier habit to sustain.

Writ ([related review](/review/jeto-ja-vizhu-no-chto-s-jetim-delat)) raised $3.8 million last spring with nothing but a teaser page promising to rethink what business analytics should look like. It has since launched, and its platform does three things:

- Automatically refreshes source data from multiple inputs – what conventional analytics systems do.

- Automatically detects anomalies in that data – surfacing insights, as the platforms above do.

- And immediately generates recommendations for what to do about them – which is the next step beyond.

The step after that would presumably be to actually implement those recommendations and observe the results: detect anomalies, generate insights, produce recommendations, apply them, measure the outcome, repeat.

Incident.io ([related review](/review/kak-ni-stranno-no-sdelannoe-ii-budet-chashhe-lomatsja)) has raised $96.2 million and is already moving in that direction – a platform that helps engineering teams detect software errors and trace their root causes. The founders' thesis: since AI is writing an increasing share of production code, it would be logical to put AI in charge of diagnosing and resolving the problems that AI-written code introduces.

03 /Opportunities

The old saying holds: "You can't manage what you can't measure." But that's only part of the story. There's no point measuring what you can't draw conclusions from. And conclusions that don't lead to action don't accomplish anything.

In other words, pure data-collection analytics platforms are only the tip of the iceberg. Underneath, there should be at least two more layers: a system that draws conclusions from the data, and a system that generates actionable recommendations from those conclusions.

Historically, humans were responsible for those two layers. Now, AI can take them on. Building new analytics platforms that include both layers – that's the relevant and timely direction to pursue.

The clearest starting point is any domain where first-generation analytics platforms are already widely adopted. The user base exists; the data habits exist; the need for a better synthesis layer is obvious. Which domain would you go after first?