Should we build our own AI or buy something off the shelf?
Neither, exactly. AI systems have three layers, and the right answer differs per layer: rent the models, own the integration layer, and configure the workflow layer. Companies that build all three usually rebuild commodity plumbing. Companies that buy all three end up with reporting that cannot answer their specific questions. The durable position is owning the layer where your data is joined and your business rules live.
Break the question into layers before answering it
"Build or buy" is unanswerable as posed, because an AI system is not one thing. It is at minimum three: the model that does inference, the integration layer that moves and reconciles data between your systems, and the workflow layer where your rules, thresholds and outputs are defined.
- Models: rent. They improve monthly and get cheaper. Building your own is justified only in rare, narrow cases, and even then usually on top of an existing base model. Keep this layer swappable — see model integration.
- Integration layer: own. This is where your identity resolution, your field mappings and your history live. Whoever owns it controls whether you can change any other vendor.
- Workflow layer: configure. Rubrics, alert thresholds, brief recipients and routing rules should be settings a competent operator can change, not code changes that require a developer.
The eighty-twenty test
In practice most service businesses find a productized platform handles the large majority of what they need — the connectors, the scheduling, the storage, the reporting shell — while some meaningful fraction is genuinely specific to how they operate. That fraction is where custom work belongs.
The useful framing is: if the platform solves most of the problem, customize the remainder. If the technology to solve it does not exist yet, that is a custom development conversation, and it should be an explicit decision rather than a slow drift.
Honest costs of each direction
Building everything means you also own monitoring, retries, credential rotation, API deprecations and the person who understands it all. That last item is the real risk; one engineer holding the whole mental model is a fragile arrangement in a company that does not otherwise employ engineers.
Buying everything means accepting someone else's definitions. If a vendor decides a lead is any call over thirty seconds and your business defines it differently, you will argue with your own dashboard forever. Ask what is configurable and what is hard-coded before signing anything.
A rule that holds up over time
Own what is specific to you. Rent what is specific to nobody. Your competitive knowledge lives in how you define quality, which jobs you value and how you route work — not in the code that authenticates against an ad platform.
That principle is why our work pairs a configurable platform with custom integration rather than treating them as alternatives.
Topics: build vs buy · platform · architecture · decisions
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.