Every company adopting artificial intelligence faces the same structural decision, and it is rarely framed clearly. Which parts of this should we own, and which should we rent?
The answer is not the same for every layer of the stack, which is why "should we build or buy AI" is not really one question.
Separate the layers first
It helps to think in four layers:
- The model — the underlying system that produces text.
- The infrastructure — where it runs, how identity and access are handled, where data is stored.
- The platform — how your users reach it, how documents get in, how permissions and history work.
- The process logic — the specific rules of how your business handles your work.
Most confused discussions happen because two people are talking about different layers without noticing.
The model: rent it
Training a competitive general-purpose model requires resources unavailable to almost every company on earth, and the result would be worse than what you can rent for a fraction of the cost. There is no serious argument for building here.
The real decision at this layer is not build-versus-buy but which provider, under what contract, processed where. Those terms matter far more than the model's benchmark scores.
The infrastructure: rent it, but choose deliberately
Running your own hardware for AI workloads is possible and occasionally justified — usually where data genuinely may not leave the building for regulatory reasons. For most companies it means paying substantially more for something you then have to operate yourself.
What you should be deliberate about is where the rented infrastructure sits. Processing region, sub-processor arrangements, and whether the provider will sign the contracts you need are decisions with legal consequences, not just technical ones.
The platform: this is the real question
Here the trade-off is genuine. Building your own platform means a login system, permissions, document handling, conversation history, audit logging, and maintenance of all of it — indefinitely. That is a significant ongoing commitment that has little to do with your actual business.
Buying means accepting someone else's decisions about how it works, and depending on them continuing to exist.
For most companies below a certain size, building this layer is a mistake made for emotional rather than commercial reasons. The work is larger than it looks and none of it differentiates you from a competitor.
The process logic: own it
This is the layer where your actual advantage lives. How your company assesses a claim, prices an unusual order, or decides what makes a supplier acceptable — that knowledge is yours and it is the reason the automation is worth anything.
Whether you write it yourself or have someone build it for you, insist on understanding it and on being able to change it. A process integration you cannot inspect or modify is a dependency you will regret.
A practical position
Rent the model. Rent the infrastructure, choosing the region and contracts carefully. Buy the platform unless you have a compelling reason otherwise. Own the process logic and make sure it is documented.
The most common mistake is the reverse: companies build a platform nobody needed while treating the process logic as something the vendor will figure out.
