Product shape is neither necessary nor sufficient. Where shape correlates with survival, the underlying causes are segment knowledge, workflow entrenchment, and a data loop. Shape is the visible surface of the moat, not the moat.

I recently read an essay where the author’s argument goes like this. AI application companies have spent four years rotating through defensibility stories to raise money. First, fine-tuning was the moat. Then evals. Then model routing. Each story expired, and a new one slid in. He calls this a shell game, and he is right.

Then he offers his own answer: the real moat is product shape. The fit-for-purpose arrangement of screens, workflows, and views that a focused team builds for one vertical. A great kettle can’t be a great toaster. Model providers can’t maintain a thousand shapes. So shape yourself tightly to one job, and you’re safe.

It’s a clean argument. It reads well. It feels true the way a rhyme feels true. It’s also the fifth story in the shell game he just described. I want to take the argument apart properly, because it’s not stupid. Parts of it are correct, and the parts that are wrong are wrong in an instructive way. The founder is a smart operator describing something real. I believe he’s just labelled the wrong thing as the cause.

The Claim, Stripped Down

Strip the essay to its irreducible parts, and you get one testable claim: a fit-for-purpose product design is a durable competitive defence against model providers.

For that to be a moat, in the plain meaning of the word, two things need to hold. First, companies with strong shape should survive. Second, companies without it should die, or at least bleed. That’s what “necessary and sufficient” means when you take the Latin off. If bad-shape companies thrive and good-shape companies die, shape isn’t the cause of anything. It’s a decoration on top of whatever the real cause is.

So the test is simple. Find a company with a famously bad shape and a fortress business. Find companies with beautiful, fit-for-purpose shapes that got run over anyway. If both exist, the claim is dead, and the interesting question becomes: what were the survivors actually protected by?

Both exist. In quantity.

Not Necessary: The Ugliest Product in Finance

The Bloomberg Terminal is a black keyboard and an interface that has looked broadly the same since the 1980s. Amber text. Cryptic function codes. A learning curve that firms pay to train people through. By every standard in the founder’s essay, this is a shape disaster. If fit-for-purpose design were the moat, Bloomberg should have been dismantled decades ago by any of the dozens of competitors that shipped cleaner, friendlier, cheaper screens. Many tried. Money was spent.

Instead, a single seat costs $31,980 a year as of January 2025; the price has only ever moved up, and there were about 325,000 subscribers as of 2022, generating more than 85% of Bloomberg’s revenue.

Why? Not shape. The moat is the other 350,000 people. Traders message each other on the terminal. They trade through it. Their entire daily workflow, their contacts, their muscle memory, lives inside it. Bloomberg’s own product page sells the network, not the screens. Leaving Bloomberg means leaving the room where your counterparties are, not switching software. That’s workflow entrenchment plus a network, and it’s strong enough to carry the worst shape in professional software for forty years.

One case doesn’t prove a rule. But one case is all you need to kill a “necessary” claim. A company with terrible shape built one of the deepest moats in software. Shape is not necessary.

Not Sufficient: Three Bodies

Now the other direction. Companies that had the shape and lost anyway.

Jasper, the AI writing tool. The first big AI application company had exactly what the essay prescribes. Not a bare chat box. Templates for fifty marketing jobs. Brand voice controls. Team workflows. Campaign structures. A real, opinionated, fit-for-purpose layer for marketers, built years before anyone else. The differentiation was the marketing-vertical UX layer, templates, prompt scaffolding, a brand voice the model would not naturally produce. Then the model provider shipped a plain chat box at $20 a month, and revenue fell from roughly $120 million to $88 million. The shape stayed. The customers left. What the company never built was the thing underneath: proprietary data the model provider didn’t have, workflow lock-in that made leaving painful, switching costs beyond a saved template. The shape was the whole moat, which is to say there was no moat.

Stories. Snapchat invented a genuinely new product shape. Ephemeral, vertical, tap-through stories. Novel enough that analysts wrote about it the way people now write about AI interfaces. Instagram copied the shape feature-for-feature in 2016, and within eight months Instagram Stories passed Snapchat’s entire user base, hitting 250 million daily users against Snapchat’s 166 million by mid-2017. Same shape, both companies. The winner was the one with distribution: 700 million people who didn’t have to download anything.

Slack. The best-shaped workplace product of its decade. Loved, polished, fit for purpose in every detail. Microsoft copied the shape, bundled Teams into Office 365 at no extra cost, and passed Slack in daily users within two years of launching. Teams sits above 320 million users today. Slack sold to Salesforce. Slack’s shape was better the entire time. It didn’t matter, because the fight was never about shape. It was about who already owned the buyer relationship and could make the copy free.

Three different decades of software, one pattern. Shape is visible, and anything visible is copyable. What actually decided each fight was invisible from the outside: distribution, bundling economics, data, network, switching cost. Shape is not sufficient.

The Test Is Running Right Now, In His Market

Here’s where the essay stops being an abstract argument and becomes a live experiment: the founder’s structural assumption is being falsified in his own vertical this quarter.

The assumption: model providers can’t maintain a thousand product shapes, so they won’t come after yours. The first half is true. The second half doesn’t follow. They don’t need a thousand shapes. They need the ten biggest, and legal is on every version of that list.

The record. Anthropic has shipped 13 vertical products since February 2025, including products for coding, design, financial services, small business, and law. Claude for Legal launched in May 2026 with 80-plus legal agents, contract review, Word integration, and Freshfields, Quinn Emanuel, and Holland & Knight already in production, priced at roughly $20 per user per month. Three weeks later, OpenAI hired the founder of a $3.2 billion contract management platform to build its own legal vertical. Read that carefully. A model provider recruited the person who built the defining contract-management shape of the last decade, specifically to build shapes.

My only question in all this is, rhetorically, whether you’d buy a computer from a car company. The market already answered. Global law firms are running legal AI from a model lab into production at consumer prices. And the price point matters as much as the entry. When the lab’s version costs $20 per seat, and the vertical tool costs 10 times that, the vertical tool’s pricing survives only if it’s defending something the lab can’t replicate. A screen layout is not that thing.

Two more forces make it worse. First, the labs control the layer everyone builds on, and at least one has stopped consistently warning partners before launching competing products. Your supplier is your competitor, and it can see the demand data. Second, the shape layer itself is being unbundled from below. The document platform at the core of most large law firms shipped an open protocol server that lets any AI agent access governed legal content, permissions and audit trails intact. When the data layer opens to all agents, the proprietary container ceases to be a container. The shape becomes a skin any agent can wear.

None of this means every legal AI application dies. It means the ones that live won’t be saved by their screens.

Why the Story Sells Anyway

If the evidence is this lopsided, why does the shape story land so well? Three mechanisms, all well documented, none flattering.

Easy ideas feel true. The brain grades claims partly on how smoothly they process. Psychologists have shown that statements that are easier to read, easier to say, or delivered as rhymes are judged more accurate than identical claims stated awkwardly. “A great kettle can’t be a great toaster” is a perfect fluency machine. You can picture it. You can repeat it at a partner meeting. The essay’s own opening actually names this trick: each expired defensibility story worked because it could be explained in a few sentences and made to feel legible. Then the essay does the same thing with kitchenware. The fluency of the kettle line is doing the work the evidence should be doing.

You only autopsy the survivors. “Look around: every mature product you own is fit for purpose.” True, and empty. The products you own are the survivors. The graveyard is full of beautifully fit-for-purpose products that died anyway, and you don’t own them, so they never enter the sample. Judging what causes survival by examining only survivors is the oldest error in business writing. Every dead product had a shape too.

Builders overvalue what they built. Behavioural researchers call it the IKEA effect: people assign inflated value to things they assembled themselves, even wobbly ones. A product founder’s craft is the artifact. The screens, the flows, the redlining surfaces. When a founder reasons about defensibility, the artifact is what he controls and loves, so it becomes the moat. That’s attachment wearing analysis as a costume, not analysis. Notice the essay never mentions distribution, retention, or switching costs until a late paragraph, where it quietly concedes that, compared with other app companies, you need “network effects, data, brand and raw speed.” That paragraph is the honest essay. It’s also a confession: if shape doesn’t stop a copycat startup, it never will stop a lab with unlimited money.

And underneath all three sits the incentive. The essay was published by a founder, days after a product launch, into a fundraising environment that rewards exactly one thing: a defensibility story that fits in a sentence. The genre demands the shell game. Fine-tuning, evals, routing, shape. Same shell, new pea.

What the Moat Actually Is

So what protected the companies that actually survived contact with a bigger player? Strip away the stories and three mechanisms show up again and again. They’re not glamorous, which is partly why they don’t trend.

Segment knowledge. Not “we know lawyers.” Knowledge that is expensive to acquire and invisible from the outside. Which clause disputes actually block deals. What a claims adjuster checks before escalating. How a pharma rep’s call gets audited. The clearest case is the company that built a CRM for drug companies on top of Salesforce’s own platform. It rented its shape from the platform it could have feared. What it owned was the regulatory workflow: the compliance rules, the validation requirements, the audit trails that life-sciences sales legally require. The platform owner could see the shape perfectly and still couldn’t cross the knowledge gap economically. Even the analysts covering the AI vertical boom keep landing on the same answer when asked what defends these companies: distribution and data, not interface. Segment knowledge is why the shape was right, not the other way around.

Workflow entrenchment. The product becomes the place where the work lives, not a tool the work passes through. The system of record. Bloomberg again: contacts, chat history, trade flow. Ripping it out means reconstructing your working life. In enterprise software, this shows up as a brutal, boring truth: the buyer isn’t optimizing for the best product; the buyer is minimizing blame. Enterprise purchases are made by people who get fired for a bad call and merely nodded at for a good one. Loss aversion runs procurement. “Nobody got fired for buying IBM” was never a joke about IBM; it was a statement about how organizations price risk. Entrenchment converts your product from a choice into a default, and defaults don’t get re-litigated every budget cycle. Shape can’t do that. Only accumulated dependence can.

A data loop. The product generates data that improves it, and that data can’t be bought elsewhere. The legal AI company currently winning the enterprise market reached $190 million in annual recurring revenue and an $11 billion valuation without training its own foundation model. Its defence, as the market reads it, is the depth of firm relationships and the proprietary usage inside them. A real data loop compounds: more usage, better output, more usage. It’s the only one of the three mechanisms that gets stronger while you sleep. Note that the shape essay’s author, to his credit, mentions his own market-data tracking in passing. That one line is worth more than the rest of the essay. He filed his actual moat under miscellaneous.

Here is the relationship between these three and shape, stated plainly. A team with deep segment knowledge will almost automatically ship a fit-for-purpose shape because they know the job cold. So when you look across the market, good shape and survival move together. That’s the correlation the essay noticed. But the shape is the exhaust of the knowledge, not the engine. Copy the shape without the knowledge, and you get the copy of the painting, brushstrokes and all, worth nothing at auction. The three companies in the graveyard above had shape without the underlying mechanisms. Bloomberg has the mechanisms without the shape. The causality only runs one way.

The Positioning Read

There’s a positioning lesson buried here, and it’s the same lesson the essay’s author half-knows and keeps stepping past.

A position isn’t what you say about yourself. It’s what the market can verify about you when your words are removed. Run the test on “our moat is our product shape.” Remove the words. What’s left is a set of screens that any funded competitor can study in a free trial and rebuild in two quarters, and that a model lab can absorb into a $20 subscription. Nothing verifiable survives. Now run the test on Bloomberg. Remove every word Bloomberg has ever published about itself. What’s left is 350,000 professionals who cannot do their jobs without the box, and a price that has risen for forty years without churn. The position survives with the words deleted because it was never in the words.

This is the difference between a story about defensibility and defensibility. A moat is a claim about future margins, and there’s exactly one place it shows up: the numbers. Retention that doesn’t budge when a cheaper copy launches. Pricing power that holds through a platform shift. Sales cycles that shorten because the buyer’s risk question is already answered. The essay contains analogies, one support-ticket anecdote, and a launch announcement. No retention curve. No churn number. No pricing evidence. For a claim whose entire content is “we will keep our margins when attacked,” that’s not an oversight. That’s the tell.

The deeper positioning error is treating positioning as an output. Shape, in the essay, is something you present to the market. But a durable position is an operating filter, not a presentation layer. It’s the list of things a company refuses to build because doing so would dilute its knowledge advantage. It’s the decision to go deeper into one workflow when the board wants breadth. The companies with real moats can usually state, in one sentence, the thing they know that nobody else has paid to learn. The shape follows from that sentence. It never precedes it.

What Would Change My Mind

Intellectual honesty requires naming the falsifiers, so here they are.

If, three years from now, the labs’ vertical products have stalled in legal, finance, and design, and independently shaped applications have held pricing and retention against them without deep data or entrenchment advantages, the shape thesis gains real support. Early enterprise adoption of a lab’s legal product could be tourism; production logos are not the same as renewed contracts.

If it turns out that shape is where the data loop starts, meaning you can’t collect proprietary workflow data without first shipping the opinionated surfaces that capture it, then shape is upgraded from decoration to precondition. This is the strongest version of the founder’s argument, and he should have made it. Shape as the intake valve for the moat, rather than the moat. I’d sign that essay.

And if agent-driven interfaces genuinely fail in daily enterprise use, if humans insist on stable, hand-built screens indefinitely, then the shape layer retains more value than I’ve granted. The support-ticket anecdote points somewhere real. People do hate change. Whether that inertia survives a generation of workers who grew up delegating to agents is an open question, and I hold it as one.

Finally

If you run an AI application company, here’s what this changes this week.

Stop presenting your interface as your defence, internally or to investors. Anyone who’s watched this market can name the graveyard, and the pitch reads as either naive or laundered.

Run the removal test on your own company. Delete every word of your deck. What can the market still verify? If the answer is a screen layout, you have a head start, not a moat, and head starts have expiry dates.

Then audit yourself against the three mechanisms. What do you know about your segment that cost real money to learn and doesn’t appear in your UI? Where does your product hold data or history that would be painful to export? What loop makes your product better with every use in a way your model supplier can’t see? Whatever you find, that’s the thing to feed. The roadmap question is not “what surface do we add.” It’s “which surface deepens the entrenchment and the loop.”

And run the one-question stress test the whole argument reduces to: the model lab ships your shape tomorrow at $20 a seat. What keeps your customers on Thursday?

Whatever your answer is, that’s your moat. If your answer is the shape, you don’t have one yet. Build the engine. The paint you already know how to do.