Generic AI is a commodity, expert AI isn't

Every AI demo you've seen in the last year looks the same. Someone opens a chat, asks a reasonable question, and the model gives a plausible answer in well-formatted markdown. The answer could have come from GPT, Claude, Gemini, or any of the models that will be released next quarter. The demo works because the underlying model is good. It also works because the underlying model is interchangeable. Swap the vendor and the demo still runs.

This is the thing I keep coming back to when I think about what's worth building right now. The raw capability of frontier models is going to keep getting better and keep getting cheaper, and the gap between vendors is going to keep shrinking. That's a commodity trajectory. Commodities are fine to use and bad to build a company on. If the only reason anyone would pick your product over a bare ChatGPT window is that you pre-wrote a prompt and added a logo, you're not building a product. You're building a thin client against someone else's thick investment.

I've spent the last year building in and around this space, and I've landed somewhere specific. The moat isn't the model. The moat is everything around the model. That's the bet I want to walk through.

What a generic model can't know

A frontier model can write you a contract clause, explain a tax treaty, draft a setlist, summarise a PDF, or roleplay as almost any professional you can name. It can do all of this because it was trained on a huge, messy, public corpus that includes examples of all of those things. What it cannot do is know what happens in your specific world.

Take a concrete example from my own work at Stagent. When a DJ plays a show in Berlin, there are at least a dozen downstream obligations the artist, the agent, and the booking agency have to fulfill inside the next two weeks. The setlist has to be reported to a rights organization. The invoice has to include a withholding tax line if the promoter is in certain jurisdictions. The per-diem owed to the artist depends on the length of travel in a way that varies by country. The accommodation and transport have to be matched against the original advancing document to catch discrepancies before the final payment. The agency commission has to be deducted, converted, and accounted for differently depending on whether the artist is on retainer or on a deal basis.

Ask a generic model to 'help me with the paperwork for a Berlin show' and it will give you a plausible checklist that misses at least six of these. Not because the model is stupid. Because the model has never seen inside a real booking operation, doesn't know the conventions, doesn't know which rights organization operates in which territory, doesn't know what the default withholding rate is in Germany for foreign performers, and doesn't know what 'advancing' even means. These are not things a model can reason its way into. They're things you learn by spending years inside the industry, and they're invisible from the outside.

This is the shape of every real domain. Medicine has this. Law has this. Construction has this. Hospitality has this. Accounting has this. And the music industry has it in a particularly sharp form because the conventions vary per genre, per country, per deal type, and per decade of the industry. A generic model is a generalist. Every industry is built on specialists. The bet isn't whether a generalist can ever become a specialist. It's whether you can build a layer on top of a generalist that encodes enough of the specialist knowledge that the combined system behaves like a specialist from the user's point of view.

What the moat actually is

I've come to think of the moat in three layers, roughly in order of how easy they are to copy.

The first layer is the prompt. This is what most AI products are. You write a strong system prompt, you add some few-shot examples, you ship a wrapper. Prompts are the cheapest layer and they're also the most copyable. Someone will reverse-engineer your prompt by next Tuesday, or a competitor will write a similar one from their own domain knowledge, or a future model will make your prompt unnecessary. The prompt layer is real but it's not defensible.

The second layer is workflow. This is where you take the prompts and wire them into the actual operations of a specific industry. Not just 'summarise this contract', but 'ingest this booking contract, extract the fee and the rider, cross-reference the rider against the venue's capabilities, flag any missing technical requirements, generate the advancing document the tour manager actually uses, route it to the venue, and remember the whole thing so the next booking at the same venue is faster'. That's not a prompt. That's a product. It requires you to know what the user does all day, what data flows between which tools, where the friction lives, and which edge cases will embarrass you in front of a promoter if you get them wrong. Workflow is harder to copy than prompt because you need domain experience to even describe it.

The third layer is data. Not training data. Proprietary runtime data. The stuff that accumulates inside your product as users use it. After three years of agencies managing their bookings through your platform, you have a data set that describes the industry better than any public source: who books whom, at what fees, through which promoters, with which riders, for which genres, across which territories, with what outcomes. That data set isn't just an asset for retraining models. It's the backbone of every judgment call your product will ever make about what to recommend, what to warn about, and what to autocomplete. Data is the hardest layer to copy because time is the only input and there's no shortcut.

Sit above a commodity model with enough workflow and enough data, and you stop competing on model capability. You compete on how well you've encoded the industry. The model is the engine. The product is the chassis. Nobody buys a car for the engine alone.

Why this is the right moment to build vertical

The reason I think this is an especially good moment for vertical AI is that the commodity layer is getting good enough, fast enough, that the interesting questions have moved up the stack. Two years ago, the hard problem was getting the model to generate coherent paragraphs. Today that's free. The hard problem is getting the model to generate the right paragraph for a specific operational context, with the right data fetched from the right place, without hallucinating the parts that would cost a user real money if they were wrong.

That's a vertical software problem, not an AI research problem. The people best positioned to solve it are not the people training the models. They're the people who already know the industry. If you've spent a decade watching booking agencies fall over on the same operational problems, you can build a specialist AI for them faster than any generalist can, because you know what the right answer looks like and they have to guess.

I'd bet, over a five-year horizon, that most of the durable value in applied AI will sit in vertical products with three characteristics. One, they're operated by a team with deep domain experience. Two, they sit on top of commodity models rather than trying to build their own. Three, they accumulate proprietary runtime data as a side effect of being used, and that data compounds into a moat the commodity layer can't reach.

This is the bet I'm making with the work I'm doing through Ryde. Not building a better model. Building the infrastructure that lets small teams with deep domain expertise deploy AI for their industry without having to become AI companies themselves.

Onoma, the first product out of Ryde, is built on exactly this thesis. The runtime is provider-agnostic by architecture. Claude, GPT, Gemini, any of them work. The middleware handles the API keys; the core never touches them. What Onoma owns is the layer above: per-user memory that compounds with every interaction, domain-specific agents that encode industry workflow, and a context pipeline that routes information based on patterns no single prompt could capture. Each user's data lives in its own isolated runtime. No shared data stores. The privacy isn't a policy document. It's the architecture. Swap the model and nobody notices. Remove the context layer and you're back to asking ChatGPT to 'help with the paperwork for a Berlin show'.

The model is the raw material. The domain is the product. The data is the moat. I keep those three sentences written on a sticky note above my desk because it's easy to forget in a news cycle full of benchmark announcements.

The test I use now

When I see a new AI product, I run a simple test. If you replaced the underlying model with a different one tomorrow, would anyone notice? If the answer is 'no, the user experience is basically the same', the product is a thin wrapper and its pricing power will follow the model's pricing power toward zero. If the answer is 'yes, because the product depends on a specific workflow, a specific data set, or a specific domain knowledge that's encoded into the layer above the model', then the product has a chance to be durable.

Generic AI is a commodity. That's not a criticism. Commodities are useful. They're just a bad place to plant a flag. The flag goes on the hill where the specialists live. That's the one I'm walking up.