The new, new AI-native playbook, or: everything you thought was the product isn't
Like most, I feel like I'm constantly struggling to keep up with the pace of AI development these days. Not just in the technological sense, but mainly trying to make sense of the paradigm shifts it induces. As a founder, it's ever more important to understand how this impacts how we build, ship and sell products. The playbook feels like it's constantly shifting, so I wanted to summarize some of the trends I've seen.
I hope these serve as a useful reference as I myself build towards my next adventure.
AI start-ups have changed a lot since I started working in the industry. Each wave would naturally disrupt the previous. Traditional Machine Learning companies, experts at feature design for decision trees got swept away by the Deep Learning wave, where the model would learn the discriminatory boundaries in some latent high dimension purely based on data. Then, the LLM-powered product blew those products out the window: why bother training your own model at all? Pre-training will do the job for you. Finally, the agentic iterations arrived, with amazing post-training routines and systematic approaches to integrated vertical ontologies, custom workflows and specialized tool integrations. And now, the latest autonomous agent step function is promising to disrupt things once again: why bother even designing anything, just describe your very niche flow and let the agent figure out a skill to accomplish your task without you even prompting it.
This bitter-pilled revolutionary cycle has become increasingly shorter and shorter. It has killed many incumbents at each interface. It also makes the position of a founder ideating very tenuous: every phase shift needs to be anticipated, timing is more important than ever.
These days I've been spending a lot of time having conversations with many actors of the AI space: investors, founders (experienced and prospective), users, potential customers. I've been spending at least as much time building, prototyping, testing and confronting various ideas. And I keep having the same experience.
A potential lead describes one, two, five, ten time sinks they have. The instinct fires: smells like friction, great, I can build a tool for that. Until they walk you through the SaaS graveyard: someone has been here already, and has charged them $49/month to solve their issue. Ten tabs, ten dashboards, ten products that each try to solve one problem well enough to justify their cost. CRM, invoicing, contract management, scheduling, project tracking, a tool that turns meeting notes into action items, another tool that turns action items into meeting notes. The circle of enterprise life.
Problem solved, right? Actually not. People don't use the product in the way it was "intended". No one has time to learn a product. No one is watching that onboarding video. No one is adapting their workflow to someone else's opinion about how work should flow.
I have seen thousands of agentic versions of the same SaaS app. Promise to solve some sliver of the issues, slap some chat or agent on it. The result is exactly the same: users don't care, and you get studies showing how 90% of AI initiatives fail. The problem was never the automation. It was the assumption that the product gets to dictate the shape of the work.
In the latest verbiage, what people are really asking every time is: can you just teach the agent the skill to solve my problem?
This keeps happening, and I think it's the thread that connects a bunch of things I've been seeing about what actually works when you build AI-native products right now. Not a theory of moats, not a strategy deck, more like a set of observations from the field that I've been trying to crystallize into something coherent.
Here's my attempt.
The product is the substrate, not the solution.
The ten-dashboard meeting keeps replaying because SaaS was built on a specific deal: we find a friction-driven wedge, we design the workflow, you learn our interface. That deal made sense when building software was expensive and exclusive. It makes no sense when a user can (or expects to be able to) describe a capability in a sentence and have it running in minutes.
What this means, concretely, is that the product you're building isn't a tool; it's the surface that tools get built on. Each skill a user creates is one more reason they never leave, not because you've locked them in with contracts or data migration headaches, but because they've accumulated thirty micro-products that are shaped exactly to how they work. Not how you thought they should work. How they actually do.
The moat isn't the model. The moat is the accumulated institutional sediment: the skills, the context, the muscle memory, the context graph that builds up on your platform over time.
It also means building traditional apps is not worth it anymore: the interface to your product has become abstracted into a one-of-one interaction surface, tailored to each user's specific workflow and preferences.
Your product shouldn't have an address.
A contractor on a roof isn't opening your web app. A procurement officer in Lagos is on WhatsApp. A lawyer at midnight is in their email. But this isn't just about role-specific interfaces. It's about how many portals we have to interact with products at all. Computers brought companies into our homes, the cloud brought them to our browsers, mobile into our hands. Each came with their boundaries: a program, a website, an app. We're now going one layer further: access everything from everywhere all at once.
Claude Code just shipped session handoffs from terminal to phone. That's the direction: the product isn't a place you go, it's a layer that follows you around, an abstract floating entity ready to receive your every wish: a literal cloud. WhatsApp, email, voice, a Word document, a terminal, whatever. Every surface is an interaction surface, and every interaction should be universally powerful. Remember when some functionality was available only on desktop and not on the mobile website? Yeah...
Of course this isn't a moat, anyone can build a WhatsApp bot. But it's a prerequisite. If you can't reach people in the surfaces they already inhabit, you never collect the context. You're just another tab in the graveyard, a rusty gate no one wants to push.
Stop organizing. Start retrieving.
Every enterprise AI project I've seen hits the same phase: "we need to structure our data first." Then six months of taxonomy design, ontology workshops, tagging sprints. And then the ontology is incomplete, because ontologies are always incomplete (that's basically their defining feature) and the project stalls while someone argues about whether a "client" is a "customer" or an "account."
The counterintuitive move that actually works: keep your knowledge flat. Good atomic concepts, a solid corpus, and let the model do the linking at query time. Don't precompute every possible relationship. Don't build a cathedral of categories. Build a pile of well-labeled bricks and trust retrieval to assemble them just in time.
You still need anchors: canonical entities, verified facts, the stuff that has to be right. And you absolutely need verification after the fact. But the connective tissue, the "which of these twelve documents answers this question", that's what the model is for. A flat knowledge layer with strong retrieval beats a perfect taxonomy that ships a year late and is stale on arrival.
Close the loop, you stochastic parrot.
There's a clean line between AI products that work in production and AI products that work in a pitch deck, and it's this: does the agent see what happened after it acted?
A coding agent that generates code but never watches it run is a suggestion engine. One that opens a browser, sees the result, and iterates: that's a tool. The same principle scales to everything: a contract drafting agent that never sees redlines, a scheduling agent that never checks if the room was actually booked, an email agent that never reads the reply.
Successful agents are loopy. They act, observe side effects, adjust. Most AI products today are still open-loop: they produce an output and throw it over a wall. The ones that close the loop are the ones that survive contact with reality.
Validation is the new friction.
AI didn't eliminate the bottleneck in most workflows. It displaced it.
The bottleneck in software development is not writing code anymore. It's understanding the problem, making design decisions, reviewing changes. Now that generation is nearly instant, the review gate is the entire constraint. Same pattern in legal: the bottleneck isn't drafting contracts, it's legal review. In government, it's not processing applications, it's the approval chain. In healthcare, it's not writing up notes, it's clinical sign-off.
The work was never the bottleneck. The oversight was. AI just made this painfully visible.
Which means the real opportunities are in industries built almost entirely out of approval gates: government, legal, healthcare, regulated finance. Not because they're "ripe for disruption" (ugh), but because the shape of the work there is mostly gate and very little generation. Collapse the generation step and the product becomes the gate itself: the review workflow, the compliance check, the approval chain with AI sitting inside it, not replacing the human judgment but compressing everything around it.
Build for the gate. That's where the value is.
The journey and the destination.
Terry Tao has this way of describing mathematical discovery: when you walk the path yourself, trying things, failing, backtracking, you pick up insights along the way that have nothing to do with the destination. Side trails. Intuitions about which problems are even worth attempting. The journey has externalities that matter more than the answer.
AI helicopters you to the solution. That's incredibly useful. But you skip the walk, and the walk is where most of the genuinely novel thinking happens: the unexpected connections, the "wait, this is related to that" moments.
I don't have a neat product insight here. It's more of an honest limit. The best AI tools right now make you fast. Making you generative, actually expanding what you'd think to explore. That's a different, harder problem, and I don't think anyone's cracked it yet. The companies that figure it out will be the ones that matter in five years. The rest of us are building very sophisticated helicopters.
Which isn't nothing. But it's worth knowing which game you're playing.