There’s a question that breaks most SaaS business models in under ten seconds.
Pick any product on your company’s software invoice. Ask one thing: does it run on unstructured data — text, audio, video, images — or on a simple, single-purpose database? If the answer is yes, your team can replicate it in a week.
Not “eventually.” Not “with the right engineering talent.” A week.
This is not a provocation. It’s a technical reality that crystallized somewhere between 2023 and 2025, and most founders, investors, and executives haven’t fully absorbed what it means.
What Happened
Foundation models solved the comprehension of unstructured data. Not incrementally — structurally. The ability to read, understand, summarize, classify, generate, and reason over text, audio, video, and images went from a hard engineering problem to an API call. Inference costs dropped over 80% between 2023 and 2025. Quality crossed human-level on most tasks that matter commercially.
The consequence is simple and devastating for a large portion of the software market: if your product’s core value proposition is “we understand [type of unstructured information] and give you [insight, action, output],” that proposition is now table stakes. Anyone with a developer, an API key, and a week of work can deliver the same result.
Category One: Unstructured Data as the Source
Gong is the clearest case. A $7 billion company built on a genuinely hard problem: understanding what actually happens in sales conversations. In 2018, that required significant engineering effort and proprietary models. In 2026, it’s Whisper for transcription, any frontier LLM for analysis, a few hundred lines of glue code, and a week of work. The value Gong holds today is not in the technology — it’s in the enterprise contracts, the integrations, the sales motion, the decade of trust built with revenue teams. The moat is commercial. The technology stopped being a moat the moment anyone could replicate it in a sprint.
Grammarly has a $13 billion valuation and does one thing: make your writing better. That is now a default feature of every IDE, every email client, and every operating system on the planet. For free.
Jasper raised $125 million at a $1.5 billion valuation to generate marketing copy. Today, that is the textbook use case of any frontier model. A system prompt and an afternoon.
Otter.ai, Fireflies, Fathom — meeting transcription and summaries. These products didn’t lose to a competitor. They were absorbed as native features by Zoom, Google Meet, and Microsoft Teams. The capability became infrastructure.
Rev.com built a real business on accurate transcription. OpenAI’s Whisper is more accurate and costs fractions of a cent per minute.
Intercom’s AI agent, Drift — “understand customer questions and respond intelligently.” This is a frontier LLM with your knowledge base as context. A weekend project.
The pattern holds across recruiting tools that parse resumes (HireEZ, Beamery), sentiment platforms that read review text (Brand24, Mention), video editing tools (Descript), async video layers (Loom’s AI summary feature), and content generation broadly. Every single one of these products was built on the same implicit premise: processing unstructured information is hard. That premise became false.
Category Two: The Simple Single-Purpose Database
This is the category most people miss — and it’s equally important.
It’s not only AI-native products. Any SaaS that solves a problem with a trivially simple database schema is also replicable now. Not because AI made the problem easier to think about — because AI coding tools made the engineering cost collapse.
Calendly is worth $3 billion. What it does: stores your availability, syncs with your calendar, and generates a booking link. That’s one table and three API integrations.
Linktree reached a $1.3 billion valuation with a product that is, at its core, a single database table. A list of links with a frontend.
Typeform raised $135 million to build beautiful forms sitting on top of a database.
Humand, 15Five, the basic layer of Lattice — employee engagement apps built on a database of people and periodic form responses. The “intelligence” is text sentiment analysis. The core is a CRUD app.
Doodle: scheduling polls. One table, options, votes.
Buffer and Hootsuite, at their core: a database of scheduled social posts with a publishing connector.
DocSend: file storage with a view counter.
None of these are stupid products. They solved real problems and built real businesses in an era when software development was expensive enough that companies preferred to pay for someone else to build and maintain these tools. That era ended. The cost of custom software collapsed the moment AI-assisted development became real. A developer with Claude Code today ships in a week what took a team of five months to build in 2019. The math on build-vs-buy permanently shifted.
The Devin Moment
In March 2024, Cognition launched Devin — the “world’s first fully autonomous AI software engineer.” The demo video got 30 million views on X. Cognition has since raised over $1 billion at a $26 billion valuation, and is running close to $500 million in annualized revenue.
Devin reads code (text), understands requirements (text), writes and tests code (text). The underlying information type is, at its core, unstructured.
At Teramot, we built our own internal version — an autonomous coding agent embedded in our engineering workflow — in one month. We’ll be open-sourcing it shortly.
This is not a claim that Devin has no value. Devin has real, compounding value: Goldman Sachs, Mercedes-Benz, and the U.S. Army as enterprise customers; a full IDE through the Windsurf acquisition; distribution built through years of enterprise sales. Those things are genuinely hard to replicate.
But the core technical capability — an agent that reads and writes code — took one Teramot engineer one month to build for internal use.
That is the point. The moat, when it exists, is commercial. The technology, in almost every case built on unstructured data, stopped being one.
The Nuance
Some companies do have real moats today. Let’s be precise about what that means.
Salesforce has moat because its data model is structurally embedded in the operational workflows of thousands of enterprises. Switching costs are real, not psychological. Workday has moat for the same reason. ServiceNow because its integrations are load-bearing infrastructure in IT operations. Stripe because financial compliance and payment trust are genuinely difficult to replicate at scale.
These moats share something important: they’re built on structured, operational data deeply embedded in how companies actually run — not on the ability to process unstructured information.
For the products that do have commercial moats — distribution, enterprise relationships, multi-year contracts — those moats exist today. But they’re fragile in a specific way. They won’t be attacked by a competitor. They’ll be eroded from within, as enterprises realize they can build the equivalent internally. That process has already started. The CIOs who in 2024 were saying “we need to evaluate vendors” are in 2026 saying “we need to evaluate what we’re still paying vendors for.” The question has shifted.
What Survived
There is one category of problem that did not become trivial. Not because the models aren’t capable enough — but because the nature of the problem is structurally different.
The structured, operational data that lives inside companies — in ERPs, transactional databases, data warehouses, internal systems — is not unstructured. You can’t point a foundation model at it and get reliable insight. This data has schemas specific to each company’s history. Semantic definitions that were never written down. Business rules that live in the head of someone who’s been in the role for seven years. Fields that are deprecated but still populated. Tables that only join correctly if you know the story behind why they were created.
revenue in this ERP means net after returns and taxes. revenue in the adjacent system means gross billing. client_id in the sales database is not the same client_id in the finance database. No foundation model resolves this without infrastructure built to make it knowable.
This is not a processing problem. It is a context problem. And context — real, operational, company-specific context — cannot be replicated in a week. It accumulates. It compounds. It becomes more valuable the longer a system runs and learns from production data.
The same AI wave that commoditized everything built on unstructured data did not touch this layer. If anything, it made it more valuable: as the marginal cost of building on unstructured data dropped toward zero, scarcity shifted entirely to the layer that understands structured operational context.
The unstructured problem was always going to be solved. It was solved at scale and at cost by a handful of foundation model labs in less than three years. The structured operational problem — the one that knows what your company’s numbers actually mean — remains unsolved for most organizations on the planet.
That’s the only frontier left worth building on.