Win on the Workflow, Not the Wrapper
A thin wrapper is a thin layer of prompt and interface on top of a model anyone can access. It feels like a product, but it has nothing the model provider or the next builder cannot replicate in an afternoon. The AI businesses that last go deep into a specific vertical workflow: the messy, end-to-end way a real job actually gets done in a real industry, with all its edge cases, integrations, and judgment calls.
Vertical means narrow and deep. Instead of 'AI for writing', it is 'AI that drafts and files a specific type of legal document for small firms', or 'AI that handles intake and scheduling for dental clinics'. The narrower the wedge, the more of the customer's actual workflow you can own, and the harder you are to replace with a general tool. Depth in one vertical beats breadth across many.
The model is one component of the product, not the product itself. Your value is the workflow you automate around it: the integrations into the tools the customer already uses, the steps you handle that the model cannot, the way you fit into how the work really happens. Solve the whole job, not just the part the model is good at.
- A thin wrapper has nothing competitors cannot copy in an afternoon.
- Go vertical: own a specific industry's full workflow, not a generic capability.
- The narrower the wedge, the more of the customer's job you can own.
- Treat the model as one component, not the entire product.
Solve the Moat Problem Early
The hard truth of AI businesses is that the smartest part of your product is rented from someone else, and so is everyone else's. The model is not a moat. If your only advantage is access to a model, you have no advantage. So the central strategic question, from day one, is: what do I own that a competitor with the same model access cannot easily copy?
The defensible assets in AI are usually not the model at all. Proprietary data from your own customers that makes your output better over time. Deep integrations into the systems an industry runs on, which are painful to rebuild. Workflow ownership so complete that switching means rebuilding how the customer works. Trust and a brand in a specific niche. Distribution and a customer relationship you control. Stack a few of these and you have something durable.
Plan for the model to commoditize, because it will. Capabilities that feel magical today will be a cheap, ordinary feature soon, available to everyone. Build so that improving models make your business better rather than making it redundant. The businesses that survive are the ones where the model is a rising tide that lifts their real moat, not the moat itself.
- The model is not a moat. Everyone can rent the same one.
- Real moats: proprietary data, deep integrations, workflow ownership, trust, distribution.
- Stack several weak advantages into one durable position.
- Build so better models strengthen your business instead of replacing it.
Respect the Cost of Inference
Unlike traditional software, where serving one more user costs almost nothing, AI products pay a real, recurring cost every time the model runs. This is the cost of inference, and it changes your economics. If you charge a flat subscription but heavy users run up large inference bills, you can lose money on your best customers. Many AI startups discover their unit economics are upside down only after they have scaled the problem.
Do the math early. Know roughly what a typical user costs you in inference per month and make sure your pricing covers it with margin to spare. Usage-based pricing, sensible limits, caching repeated work, and using smaller or cheaper models where they are good enough are all levers to keep inference cost under control. The goal is healthy gross margins even as usage grows, not just revenue that looks good until the bill arrives.
Inference cost also shapes what you should build. Workflows where the AI output is high-value and used in moderation tend to have far better economics than ones where users generate enormous volume for low-value output. Choose problems where customers happily pay far more than the inference costs you, because the result is worth it to them. The best AI businesses sell expensive outcomes, not cheap tokens.
- Every model call costs money. Flat pricing on heavy users can lose you money.
- Know your inference cost per user and price with margin above it.
- Use usage-based pricing, limits, caching, and cheaper models to protect margins.
- Favor high-value, moderate-volume workflows over cheap, high-volume output.
Validate Demand Before You Build the AI
AI does not exempt you from the basic law of business: you still have to prove someone will pay before you build. In fact the temptation is worse, because the technology is exciting enough to seduce you into building first and checking demand never. Resist it. Confirm that a specific group has a workflow painful enough to pay you to automate, before you sink time into models and pipelines.
You can validate an AI product without building the AI. Run the workflow manually or half-manually for your first customers, delivering the result by hand or with simple tools while charging real money. This concierge approach proves the demand, teaches you exactly where the value is, and tells you which parts are even worth automating, all before you invest in building the automated version.
Push the test toward money. An AI demo gets oohs and aahs and zero dollars. A customer who pays for the outcome, even when you are quietly doing the work by hand behind the scenes, has told you the truth. Charge first, automate second, and let real payment decide what you build.