Build vs Buy an AI Feature: Model, API or Off-the-Shelf?
Build your own model, call a hosted API, or buy an off-the-shelf tool? A solo builder's framework for the AI build-vs-buy call — and when each is the right one.
The AI build-vs-buy call has three options, not two — build your own model, call a hosted model API, or buy an off-the-shelf tool — and for almost every solo builder the right default is: buy the intelligence, build the product around it. The mistake I see most is treating "build vs buy" as "train my own model vs use someone else's," when the real decision is where you spend your scarce effort. Here's how I actually make the call.
The three options
- Build your own — train, fine-tune or self-host a model, or write a heavy custom pipeline. Maximum control, maximum ongoing cost.
- Buy the intelligence, build the product — call a hosted model API (the Claude/GPT tier) and put your work into the product wrapped around it. This is the default.
- Buy off-the-shelf — use a mature tool or SaaS that already does the job. Least effort, least differentiation.
| Option | You spend on | Right when… |
|---|---|---|
| Build your own model | Training, infra, ongoing ops | You have a real data moat, cost-at-scale, or a hard latency/privacy/offline need |
| API + your product | The product: problem, workflow, decisions | Almost always — the intelligence is a component, not the point |
| Off-the-shelf tool | Integration only | The job is a commodity and not your differentiator |
Why "API + your product" is the default
A hosted model gives you frontier capability for the price of a call. Spending months to train your own to match it is, for nearly everyone, building a demo of the wrong thing — effort poured into a component instead of the product. The product isn't the model; it's the problem you chose, the outcome you defined, and the human decisions the AI never makes for you. That's the part no API ships, and the part your users actually pay for.
The failure mode on this side is the thin wrapper: you buy the intelligence and add nothing around it, so you've got a feature anyone can clone. Buying the model is the default because it frees you to build the part that's hard to copy — not as a substitute for building it.
When to actually build your own
Building your own model is right when you can name the specific reason an API can't serve you:
- A genuine data advantage — you have proprietary data that makes a tuned model materially better, and that edge is your product.
- Cost at real scale — you're at a volume where per-call pricing genuinely dominates and owning the model pays for its own operation.
- A hard constraint — latency, privacy, on-prem or offline requirements an external API can't meet.
If you can't state which of these applies, you don't have a reason yet — you have a preference, and it's an expensive one.
When to just buy off-the-shelf
If the job is a commodity that isn't your differentiator — transcription, generic summarisation, a standard extraction — and a mature tool already does it well, buy it and move on. You build where the intelligence is the product and buy where it's plumbing. Treating everything as build is how a solo project runs out of time before it ships; the discipline is the same ruthless scope-cutting the rest of the build needs.
The honest answer
Build vs buy for AI isn't "my model vs theirs" — it's where your effort earns the most. Default to buying the intelligence and building the product; build your own model only when you can name the moat, cost or constraint that demands it; buy off-the-shelf when the job is plumbing. Get that split right and the AI is a component in a product with a point — which is the whole operating system this sits inside.
Part of Building AI Products. Pressure-test the idea itself with the AI product idea scorecard.
Frequently asked questions
Should I build my own AI model or use an API?
For almost every solo builder and early product, use a hosted model API and put your effort into the product around it. Training or self-hosting a model is a big, ongoing cost that only pays off with a specific reason — a genuine data advantage, cost at real scale, or hard latency, privacy or offline constraints an API can't meet. Until you can name that reason, building your own model is buying a problem, not a moat.
When does it make sense to buy an off-the-shelf AI tool instead?
When the job is a commodity that isn't your differentiator — transcription, generic summarisation, a standard classification — and a mature tool already does it well. Buying there frees your time for the part only you can build. You build when the intelligence IS the product; you buy when it's plumbing.
What's the default AI build-vs-buy choice?
Buy the intelligence (a hosted model API), build the product (the judgement, the workflow, the decisions around it). The model is a component; the product is what you wrap around it — the problem you chose, the outcome, and the human decisions the AI never makes for you. Get that split wrong and you either reinvent a model you didn't need or ship a thin wrapper with no point.