I've stopped being afraid of LLMs taking my job.
Not because the threat isn't real. Because I finally understand what it actually targets.
We are hearing thousands of variations of "SaaS is dead" ad nauseam. It started to click when Nicolas Bustamante put it plainly in early February 2026: LLMs absorb the interface layer[1]. Interfaces get commoditized, it's API versus API, only proprietary data survives. Ben Thompson pushed it from the hardware side two weeks later: thin clients are back, agents need no UI.[2] Two different angles, same verdict.
The interface moat is dead. That's settled. Or is it? Settling it didn't close the question — it opened a better one. Three very different games are now visible that weren't before. Most software companies don't know which one they're in.
The interface moat is dead — and everyone agrees
For a whole tier of software, interface complexity was the moat — or at least, it looked like one: a financial terminal charging $24K/year for UI over commodity data, a legal research platform charging premium for search over public case law. Switching costs dressed as value.
Nicolas put it plainly:
"Knowledge workers spent years learning specialized interfaces. The muscle memory is real. They're not paying for data. They're paying to not relearn a workflow they've spent a decade mastering."
— Nicolas Bustamante, The Crumbling Workflow Moat (2026)
LLMs absorb the interface layer. Interfaces get commoditized. It's API versus API. Only proprietary data survives.
Thompson came at it from the hardware side. The history of computing oscillates between thick and thin clients. Mainframes meant thin clients. PCs were thick clients. Mobile was hybrid. AI pulls everything back to thin:
"The point of an agent is not to use the computer for you; it's to accomplish a specific task. Everything between the request and the result, at least in theory, should be invisible to the user."
— Ben Thompson, Thin Is In (2026)
Agents are the extreme case: no UI at all. One looked at software, the other at hardware. Same conclusion.
Jason Lemkin (SaaStr) added the time dimension: the 7–10 year stability runway B2B used to have is gone. Competitive edges are now measured in months, not years — a startup he invested in had four clones within two weeks; PagerDuty, at $500M ARR, sat flat at 15k customers for four years before DataDog shipped a clone.[3] "System of record" is not a growth strategy: if you're the database underneath someone else's AI layer, you get low churn but also flat growth and slowly declining relevance. Lemkin, drawing on Rory O'Driscoll (Scale VP), framed stickiness the same way: distribution advantage lasts as long as product stickiness — accounting gets decades, task management gets months.[4]
John O'Nolan, founder of Ghost, sees the same compression at the code level. In February 2026, Cloudflare took a competitor's open-source codebase and rebuilt it from scratch in one week: one engineer, Vite, $1,100 in AI tokens.[5] O'Nolan's diagnosis: the premise that code is intrinsically valuable — "difficult to maintain, expensive to write" — is gone. If that's true for open source, it's coming for closed products too.
Software that hid behind switching costs — that's what's being repriced. But repriced to what?
Not dead — sorted. Scale or Craft.
Thompson hedges, though. His closing paragraph:
"UI isn't just about how to use a computer, it also, as Benedict Evans noted [...], embeds critical aspects of how a business works. Open-ended text prompts in particular are a terrible replacement for a well-considered UI button that both prompts the right action and ensures the right thing happens."
— Ben Thompson, Thin Is In (2026)
He frames this as a transition problem. It's permanent — and more interesting than he makes it sound. The death of the interface moat doesn't kill software. It reveals two independent axes that were always there but masked by the switching costs: Scale and Craft.
Scale is about what you accumulate: proprietary data, distribution lock-in, brand equity, regulatory capture. Craft is about what you stand behind: taste — the judgment LLMs optimize away from — and accountability — a named person responsible for the social agreements the software encodes.
| Low Craft | High Craft | |
|---|---|---|
| High Scale | Data Fortress (Bloomberg by data, Salesforce by lock-in) | Golden quadrant (Apple) |
| Low Scale | Commodity — 5% margins, restaurant economics. Segment lives; premium dies. | Tech atelier (37signals, Ghost, Buttondown) |
The interface moat used to hide which quadrant you were in. Now you can see. Three quadrants survive with different physics. One got repriced to commodity. That repricing is what Bustamante described — and it's real. It's just not the whole story.
Start from a shared premise: everyone uses the same LLM.
Ted Chiang called LLMs "a blurry JPEG of all the text on the Web."[6] Taste is deviation from average. LLMs optimize for average. If every company uses the same model, all products converge: same features, same UX, same pricing. But markets reward differentiation — which is exactly why there's competitive pressure to NOT use the consensus-generating machine.
| Company | Proprietary data? | Interface complexity? | Why they win |
|---|---|---|---|
| Linear | No | No | Knowing what to leave out |
| Superhuman | No | No | Speed as design decision |
| 37signals | No | No | Opinionated defaults |
What we see them competing on is judgment: what to include, what to exclude, how it should feel. These are Tech Ateliers. DHH: "What if there was no next level? What if you just did great work at the level you're at?"[7] Saarinen: "Building quality requires three elements: the belief that quality matters, the skill and taste to recognize it, and the discipline to maintain it."[8]
An LLM trained on pre-iPhone data would have recommended a physical keyboard. The training data said so. Taste requires understanding what you build. Opacity prevents craft. And agent orchestration is a harder interface problem than anything we had before — the ceiling just got higher, not lower. Ralv.ai is a literal StarCraft for AI agents: 3D spatial interface, drag-select, deploy-to-task.[9] Who decides what the map shows? That's craft. LLMs don't design StarCraft.
Accountability is the second component — same premise, different branch. Everyone using the same LLM is not just a convergence problem — it's a liability. When the model changes, what you can deliver changes, and you have no say in it. That's not a vendor relationship. That's dependency.
Software carries something deeper than taste: an identity. This is the identity-technique loop at work.[10] Tools get powerful → people learn → identity crystallizes around the technique → they build who they are around what they use → more tools get built for that identity → repeat. A generation of Bloomberg professionals didn't just learn a terminal — the loop ran long enough that the technique crystallized a professional identity. They communicate through Bloomberg, evaluate work through Bloomberg norms, have built an entire economy of shared judgments around Bloomberg. That's not a switching cost. It's a shared language built through the loop.
George Sivulka, the CEO of Hebbia, formalized what the loop produces: "Software is a stored process. It's not a neutral tool: it's an opinion for how a group of people should collaborate, encoded in a durable system."[11] Sivulka's conclusion: "You can't engineer load-bearing software from scratch and expect people to just agree on it, because load-bearing software contains social agreements for how a team will cooperate."[11:1]
The cost of switching isn't relearning a UI. It's dissolving the identity the loop produced — renegotiating how your team works together, unlearning what counts as good work, rebuilding the shared language. That's why vertical software is stickier than the interface moat thesis predicts: the moat was never the interface. It was the identity encoded in it. And that identity demands accountability — someone has to be named as responsible for the social agreements the software carries. "The AI told me to" isn't a legal defense. When you buy from 37signals, you're trusting DHH and Jason Fried: people with names, track records, skin in the game. When you buy from "an LLM," who are you trusting?
Illich's test for a convivial tool: it "can be easily used, by anybody, as often or as seldom as desired, for the accomplishment of a purpose chosen by the user."[12] A platform where the vendor can change your capabilities without your consent fails that test.
The other axis. Scale is what you accumulate — and what can't be prompted into existence. Three assets define it.
Proprietary data: Bloomberg's moat isn't the terminal — it's the data feed. Salesforce's moat isn't the UI — it's the customer graph. When agents mediate everything through APIs, the data is the product. The interface was always incidental.
Brand equity: Bustamante's own platform, Fintool, has a brand. Ghost has a brand. A brand is a trust shortcut. Agents don't have brands; people and companies do. When the interface disappears, the brand becomes the primary surface. Creator/founder brands compound here — DHH is both a craft and a scale asset.
Institutional lock-in: Regulatory capture, compliance requirements, contractual embedding. Accounting software gets decades of stickiness because switching means reauditing. Healthcare platforms persist because of certification. These aren't interface moats — they're structural positions that AI can't prompt around.
Three games remain
The commodity zone is the floor. Software without scale or craft doesn't disappear — it gets repriced to restaurant economics. The segment lives; the premium dies. Most SaaS that complains about "AI disruption" is discovering it was in the commodity quadrant all along — the Software Devolution Cycle running its course. The interface moat was the only thing justifying premium pricing on commodity value. Now it's 5% margins, high failure rates, and grinding competition. Bustamante described this repricing accurately. He just didn't ask what was above it.
Three quadrants survive. Each has different physics.
The Data Fortress
High scale, low craft. Proprietary data, distribution lock-in, regulatory capture. Bloomberg by data, Salesforce by platform. These companies survive because what they accumulate can't be prompted into existence. The data is the product; the interface was always incidental.
The risk: Bustamante's fully agentic future. When agents mediate all software through APIs, the Data Fortress becomes invisible infrastructure — high retention, flat growth, slowly declining relevance. You're the database underneath someone else's AI layer. Low churn, but also no pricing power. The moat holds; the premium erodes.
The tech atelier
Low scale, high craft. The craft moat in action. 37signals, Ghost, Buttondown — profitable, taste-driven, direct customer relationships, no VC pressure. These companies survive because what they decide can't be averaged by a model: what to include, what to leave out, what to refuse to automate.
If humans outsource taste decisions to models, the signal that training data learns from disappears. Bernard Stiegler called this proletarianization: "the loss of knowledge — savoir-faire, savoir-vivre, and theoretical knowledge — through its exteriorization in machines."[13] The pipeline eats itself. Taste is deviation from average. If the average stops moving because humans stopped deviating, there's nothing left to deviate from.
And the accountability moat compounds it: the identity-technique loop breaks in a headless world. In Bustamante's fully agentic future — where no human logs into SaaS, only agents do — no skill formation happens around a product. No identity crystallizes. Which means the only identity that can survive is the founder's. DHH's identity is 37signals in a way that no API endpoint ever can be.
The tech atelier is the Craft Game made explicit: build something excellent for people who care. At least I hope so.
If you're building here:
- Your differentiation is judgment — what you include, what you leave out, what you refuse to automate. Document those decisions. They're your moat.
- Resist the pull toward chat interfaces for routine workflows. Design deliberately: every element of your UI is an encoded decision about what matters.
- Build on a stack you can read and repair. Opacity prevents craft; transparency enables it.
If you're VC-backed and feel the repricing:
- Audit your moat honestly: proprietary data, distribution, or founder brand. If none of those, you're repriced to commodity.
- Don't confuse "interface complexity" with "product value" — LLMs will strip the former; only the latter survives.
- Consider whether you're playing the right game. Becoming profitable and small is better than staying big and commoditized.
The golden quadrant
High scale, high craft. Apple is here. Not a lot of people can join — not even with $100 billion. Sam Altman and Jony Ive are trying to build a competitor from scratch. They have the money, the best designer in the world, access to every talent pool. They will probably fail. No ecosystem. No developer network. No decades of cumulative innovation. No foundry relationships. No millions of apps.
It's a paradox: the same forces that create scale + craft also create lock-in so total that no alternative is viable. The quadrant exists. It's occupied. The door closed behind the occupant.
That's worth a full article of its own. For now, the observation: the golden quadrant proves that scale and craft can coexist — but only if you built both before the moat became visible.
The interface moat is dead. It wasn't protecting software — it was hiding which quadrant you were in. Now you can see. Three games remain. The commodity zone is everyone's default unless you have scale, craft, or both.
The question isn't whether SaaS is dead. It's which quadrant you're actually in.
VC-backed platforms and hyperscalers benefit from commodity lock-in and distribution at scale; tech ateliers are the countervailing force, building on craft and accountability. The structural barrier is funding: growth capital rewards scale over craft, so the default path is the one that burns. Leverage points exist — procurement that favors transparent, repairable tools; community and regulation that name conviviality as a criterion — but they require naming the alternative, not just critiquing the incumbent.
I've stopped being afraid of LLMs taking my job — not because the threat isn't real, but because I now know which quadrant I'm going for.
The Craft Game is also a political position: building and naming the alternative explicitly — tech atelier, Digital-Mittelstand, transparent tech — so it's a visible path, not a scattered exception.
Nicolas Bustamante, The Crumbling Workflow Moat (2026). ↩︎
Ben Thompson, Thin Is In (2026): "What workflows will transition from UI to AI, and thus from a thick client architecture to a thin one? Current workflows are TBD; future workflows seem inevitable." ↩︎
Jason Lemkin, Is SaaS Dead? No. But One Thing Is Clear: It's Unstable. (SaaStr, 2026). See distill. ↩︎
Jason Lemkin, Is SaaS Dead? No. But One Thing Is Clear: It's Unstable. (SaaStr, 2026), citing Rory O'Driscoll (Scale Venture Partners) on stickiness: distribution advantage lasts as long as product stickiness; accounting gets decades, task management gets months. ↩︎
John O'Nolan, Open Source in the Age of AI (2026): "The premise that code is intrinsically valuable because it's a scarce resource — difficult to maintain, expensive to write — is being rapidly overturned." ↩︎
Ted Chiang, "ChatGPT Is a Blurry JPEG of the Web," The New Yorker (2023): "Think of ChatGPT as a blurry JPEG of all the text on the Web... if you're looking for an exact sequence of bits, you won't find it; all you will ever get is an approximation." ↩︎
DHH, It Doesn't Have to Be Crazy at Work (2018): "What if there was no next level? What if you just did great work at the level you're at?" ↩︎
Karri Saarinen, "Why is quality so rare?" (2024): "Building quality requires three elements working together: The belief that quality matters fundamentally, the skill and taste to recognize it, and the discipline to maintain it." ↩︎
Ralv.ai: "Too many tabs. Too many terminals. Zero clarity. [...] Games already solved this. RTS games have been optimized for decades to give humans intuitive control over hundreds of units." ↩︎
Ian Vanagas, "The engineeringification of everything" (PostHog, 2026): "Tools change skills, skills reshape identity, and identity demands new tools." See distill. ↩︎
George Sivulka, In Defense of Vertical Software (2026): "Software is a stored process. It's not a neutral tool: it's an opinion for how a group of people should collaborate, encoded in a durable system." ↩︎ ↩︎
Ivan Illich, Tools for Conviviality (1973): "Tools foster conviviality to the extent to which they can be easily used, by anybody, as often or as seldom as desired, for the accomplishment of a purpose chosen by the user." ↩︎
Bernard Stiegler, Automatic Society (2015): "Proletarianization is the loss of knowledge — savoir-faire, savoir-vivre, and theoretical knowledge — through its exteriorization in machines." ↩︎