AI-native software

Tab. Tab. Tab. Done. Cursor moves me through my code terminal and rewrites code. No prompt, no mouse click.

I text "Hey Poke" in WhatsApp to get started. No signup form, no app to install. The AI assistant lives where I do.

The moment I end my meeting, Granola expands my notes with the transcript. Not a separate summary, only what matters to me.

AI-native interactions can take many shapes, but they all feel the same: frictionless.

We're used to friction

The first wave of software reduced friction by changing the medium. Sharing a digital document is easier than handing out paper copies. The potential of using software was huge but came at a cost.

We had to learn to think like computers. There are forms to fill, buttons to click, rules to follow. Even simple tasks - like noting why a deal was lost - get abstracted into fields and drop-downs.

As software matured, it chased one-size-fits-all. Users got clunky interfaces. Legacy tools turn a simple change into countless clicks through endless options. Lots of friction.

A new task hierarchy

AI-native software promises something different: less friction in tools we already take for granted. Many mistake this for AI features. That's not taking it far enough.

Doug Engelbart - who invented the computer mouse - used the term task hierarchy to describe how large goals break into smaller subgoals. To write a report we have to get sources, outline a draft and many micro-tasks like move a cursor, insert text, etc.

The real opportunity for new tools is to reshuffle the task hierarchy - to look at every interaction and ask: "Could an intern with the right context do this?" In legacy software, you tell the computer how to do something. In AI-native software, you say what you want done.

Take reporting. Understanding your business is important, but reality is complex. Traditional reporting tools solve this with plenty of filters and buttons. In modern systems, users just say: "What changed in sales this quarter, and why?" The tool translate the question into the queries and runs it by itself. From 14 clicks to one prompt.

An AI-native Applicant Tracking System (ATS) doesn't only ask about a job title when you open a new job. It asks why you're hiring in the first place and what an ideal candidate looks like.

This seems like more work at first, but similar to onboarding an actual intern, setting up tools with internal context is an investment for the future. By understanding your intent, the ATS can source talent, write job descriptions and screen candidates for you.

While traditional software digitized our workflows, AI-native tools reimagine them. This means that success is defined differently. Traditional software tracks company-wide adoption rates. The more active users inside a company, the better. Agentic products look for the opposite: the more tasks are done without human oversight, the better.

Best practices embedded

In traditional software, best practices are encoded in interfaces - templates, workflows, checklists. In AI-native software, best practices live in the model's weights.

Today I can prompt "draft this sales email like Chris Voss would" without explaining who he is. The negotiation strategies of his book "Never Split the Difference" are embedded in the model.

Software shifts from supporting decisions to guiding them. I can't stop thinking about how different today's software looks like when it's not just a tool we operate but a partner we consult.

We already see this shift in our personal lives. Therapy and companionship are among the most popular AI use cases today. Recently, one of our investors told me he and his wife use ChatGPT voice mode to settle arguments.

When AI helps us navigate intimate personal decisions, it's only a matter of time before it does the same at work. As LLMs improve their memory and continuous learning capabilities, they climb the task hierarchy - from doing what we say to helping us decide what to do.

Our expectations increase

The more we use AI, the higher our expectations for everyday tools become. When I get a deep research report in minutes, I won't spend half an hour setting up templates.

So how else can a new tool imitate my writing style?

An AI-native CRM looks through my past mails and sets up a system prompt to match my communication style. There's no need for templates. I get a tailored draft for every message.

Building this isn't easy. Much of our personal preferences are implicit. We know why the AI-written draft to our parents looks weird - but it's tricky to say why. Telling the model to "not use corporate language" doesn't work as well as "always reply with: "Yes Mom, I'll find a real job soon"."

We underestimate how much knowledge we've built subconsciously over the years. But getting it right makes the difference between AI-native and AI-slop.

With expectations changing, the gap from new to old software only increases over time. Marc Andreessen predicts that every incumbent will eventually be replaced - not only because AI can do more, but because users will expect more.

When we train our attention spans on Instagram reels, multi-step forms will be left behind - like broccoli on a child's plate. And that's the real promise of AI-native software: a world where we use computers with intent, not clicks.

A world with less friction.