Moatability, What Every AI Founder and Builder Must Understand to Move Beyond Product-Market Fit

Why understanding moats is a must for any startup wanting to survive the age of AI

Listen to to this article here.

Everyone is building an AI vertical startup, even Y Combinator wants you to build one. And in my opinion the vast majority of these startups will fail. Every pitch I heard recently was about a team that wants to build an AI startup for accounting, for legal, for education... etc etc but have we all considered that 

1) there is at least a thousand other teams thinking about building exactly that idea, and 

2) as an example, the daughter of that accountant you want to compete with can literally vibe-code a software today that will be safer and more customised to use than every fancy OpenAI, ElevenLabs or Anthropic wrapper solution you and I can think of. So even if you bring me 10 pre-seed customers, that is not enough anymore. Let me elaborate more.

Five years ago, the path from pre-seed to seed to Series A was all about evidence. We had an assumption, we turn it into a hypothesis, ran a test, and brought back the numbers to discuss with your team and backers. Each round was permission to de-risk the next set of assumptions. And your traction was the holy grail of evidence of where you will be in 18 months from now. We as founders have one job: create traction as evidence for each hypothesis as we move forward. 

That game is mostly over. Not because numbers stopped mattering, but because they stopped being scarce. AI has made it absurdly cheap to build a product, generate early usage, and show a chart that points up and to the right. Literally everyone can manufacture traction these days, which made traction stop being a proof of a validated business model or team. We need a new way beyond traction to validate the trajectory of the venture and whether it will die or fly. 

From my own experience every serious investor now is looking, even when they ask for your metrics, they are looking in behind to one question:

If a well-funded team cloned your product next month, what would still be impossible for them to deliver? 
aka. What is your moat?

The answer is now the single most valuable thing an AI builder can understand.

Where the word Moat comes from

“Moat” is not a startup term. It comes from Warren Buffett, who liked to describe a good business as an economic castle, and the durable advantage protecting it as the moat around that castle: the thing that keeps the people trying to take your castle from getting across.

Buffett talked about moats constantly but never really sat down and categorized them. The person who did that was Pat Dorsey, then head of equity research at Morningstar, in his 2008 book The Little Book That Builds Wealth. He sorted durable competitive advantage into four sources: intangible assets, switching costs, network effects, and cost advantages. That framework, written for stock pickers almost twenty years ago, is the cleanest lens I have found for building AI companies today, as long as you translate it from “what should I buy?” into “what should I build?”

And nope “loyalty” is not a moat (and never really was)

The thing most founders often mistake for a moat: customer love. And it is a big mistake. Go ask OpenAI in front of Anthropic if loyalty matter or not and you will quickly understand that loyalty doesn't matter in this current startup environment.

Loyalty is rented, not owned. No matter how delightful your product is, at some point a new product will appear that is cheaper, or faster, or both, and in the world of AI, that point arrives roughly every quarter. “Customers love us” and “we are better and cheaper” are not moats. They are nice conditions that last exactly until someone shows up nicer and cheaper. And that can be this afternoon.

A real moat does not depend on you being the best option this week. It works by making leaving costly, irrational, or simply not worth the bother, structurally, independent of how everyone feels. That is the bar. If your advantage evaporates the moment a competitor matches your price and speed, you do not have a moat. You have a head start.

The four moats, translated for AI builders

Here is Dorsey’s framework reframed for AI founders and builders. For each one I have put the honest gut-check, plus the concrete way I validate it as a venture builder: what to listen for in customer interviews, and what to put on a landing page to see if the signal is real. Treat these as the evidence you want before you raise the next round.


1. Intangible assets: proprietary data, regulation, and trust

For an AI company this rarely means patents. It means assets a competitor cannot simply buy: a proprietary or exclusive data set, a hard-won regulatory approval or certification, or genuine institutional trust in a domain where being wrong is expensive (health, legal, finance, defense).

  • The gut-check: If a competitor had your exact model weights, what would they still be missing? If the answer is “nothing,” this moat is not there.
  • In customer interviews: Ask what data they would never hand to a new vendor, and why. Ask whether certification or compliance is a gate to even running a pilot (“Would you trial a tool that was not SOC 2 / HIPAA / cleared?”). Listen for the phrase “we can only work with vendors who...” That sentence is the moat talking.
  • On the landing page: Make trust the hook and measure it. Put security badges, named design partners, and compliance language above the fold, and gate signup behind a qualification step. If trust signals lift conversion, you have found an asset worth deepening; if no one cares, this is not your moat.
  • To deepen it: Build a data flywheel only your usage can produce, or go earn the certification that takes competitors years and lawyers to copy.

2. Switching costs: make leaving painful

The most underrated moat for AI startups, and in my experience the most achievable early. Switching costs are everything the customer would have to redo, relearn, re-integrate, or risk if they left you.

  • The gut-check: Write down, literally, the steps a customer would take to rip you out and move to a competitor. If that list is short and painless, you have a problem.
  • In customer interviews: Ask them to walk you through the last time they migrated off a tool: how long it took, what broke, what they lost. Ask what lives inside their current tool that they would hate to rebuild. The depth of the groan tells you how high a wall you could build.
  • On the landing page: Lead the call to action with embedding, not signup: “Connect your data,” “Sync your workflow,” “Import from [incumbent].” If people will hand you an integration or upload before they will even pay, you are watching switching costs form in real time.
  • To deepen it: Embed into the workflow, not next to it. Hold the customer’s accumulated context (history, fine-tuned behavior, configurations, integrations) until your product becomes the place where their work lives.

3. Network effects: the product gets better as more people use it

Value that grows with each new user. Classic for marketplaces and platforms, but for AI there is a sharper version: a data network effect, where every additional user makes the underlying model measurably better for everyone else.

  • The gut-check: Does user number 10,000 get a genuinely better product than user number 100 did, because of the other 9,900? If the product is identical regardless of how many people use it, this moat does not apply, and you should not pretend it does.
  • In customer interviews: Probe whether the product gets more valuable when their peers, team, or supply-chain partners are on it too. Listen for “I would want my whole team on this” or “it is only useful if my clients use it too.” That is a network effect declaring itself.
  • On the landing page: Build invite mechanics into the test (“Invite your team,” referral access, a public leaderboard or directory) and measure whether people actually send invites unprompted. Spontaneous invites are the cleanest early proof of pull.
  • To deepen it: Make usage feed a shared asset (a model, a graph, a marketplace) rather than sitting in siloed accounts, and defend the cold-start phase hard. It is your only window of real vulnerability.

4. Cost advantages: structurally cheaper to operate

The ability to deliver at a lower cost than anyone else can, not just lower than you currently charge. For AI: proprietary infrastructure, a smarter architecture, an efficient inference setup, or a distribution channel that gets you customers for almost nothing.

  • The gut-check: Could a competitor match your unit economics by spending money? If yes, it is a price, not a moat. The advantage has to be structural.
  • In customer interviews: You mostly cannot validate this one through interviews: willingness-to-pay tells you about demand, not about your cost structure. The honest test happens inside your own build and ops, not in front of the customer.
  • On the landing page: Same caveat. A landing page can prove people want it cheap; it cannot prove you can serve it cheap. Do not fool yourself by reading demand as a cost moat.
  • To deepen it: Be ruthlessly honest about whether you have this at all. Most AI startups do not: they rent compute at roughly the same rate as everyone else. If your edge is a frontier-lab subsidy or a temporary pricing trick, treat it as a head start and go build one of the other three.

Why desirability, feasibility and viability are no longer enough

If you build ventures for a living, the work is really a sequence of tests. IDEO framed innovation as three lenses, and a good studio treats each one as something you actively validate rather than assume. You test desirability with customer interviews and landing pages, asking whether people actually want this. You test feasibility with prototypes and technical spikes, asking whether you can really build it. You test viability with pricing experiments and unit-economics models, asking whether it can make money. For two decades, passing all three was the working definition of a venture worth backing.

It is no longer enough. That framework was built for a world where building was the hard part, and in 2026 it is not. AI has collapsed the cost of passing all three tests. A venture can be provably desirable, feasible, and viable, and still be copied into irrelevance within a quarter. Desirability, feasibility, and viability now tell you that something is worth building. They tell you nothing about whether you can keep it. The three lenses describe the castle. They are silent on the moat.

This is why I do not treat moatability as a fourth lens. A lens is just another angle on the same question, namely whether we should build this. Moatability is a different question altogether, so it is a fourth way of thinking: not what makes this worth building, but what makes it hard to take once it is built. You cannot get there by looking harder through the first three lenses. It sits around them, the way a moat sits around a castle.

And here is the part most people miss. A different way of thinking still demands the same discipline. Just as you would never skip testing desirability, feasibility, or viability, you cannot leave moatability to faith. It has to be tested with real experiments, on the same calendar and at the same table as the other three, and the moat gut-checks and the interview and landing-page tests earlier in this piece are exactly those experiments.


This is where most corporate venture studios need to adapt. Studios are usually rigorous about testing the first three lenses, but they treat defensibility as a matter of opinion or a single slide in the deck. My argument is simple: give moatability its own tests, its own evidence, and its own gate, scored at the same table as the other three. A venture that passed desirability, feasibility, and viability but has run no moatability tests has not been fully validated. It has just been built quickly, which in 2026 is the easy part.

How to actually integerate this article's know-how into your team's work?

You do not need all four moats. Most enduring companies have one deep moat and maybe the beginnings of a second. The mistake is having none and dressing up a head start as a fortress.

So as you move from pre-seed to Series A, run this alongside your metrics deck. For each moat, ask the gut-check honestly, then go get the evidence, in the interviews you are already doing, on the landing pages you are already running. Pick the one or two moats you can realistically deepen given who you are and what you have got, and make deepening them an explicit, funded objective, not a thing you hope emerges on its own. Then integrate tracking and improving that moat-traction in your OKRs and teams' targets in addition to your growth targets.

Numbers will get you the meeting. The moat is what makes the numbers worth believing, and, in 2026, what gets you across the chasm.


This is an opinion-and-experience piece, not a research paper. The moat categories are borrowed from Pat Dorsey’s The Little Book That Builds Wealth (2008), and the three-lens model from IDEO; the translation to AI, the “moatability” way of thinking, the venture-builder validation steps, and any strong opinions are mine.
Sources


Email Post

Comments

Post a Comment