"Should we build our own AI?" usually hides three separate questions: which model, which product, and which integration. The answers are rarely the same.
Buy the model
Training a foundation model is out of reach for almost everyone, and fine-tuning is needed far less often than people assume. Hosted models from providers such as OpenAI, Anthropic, Google and Mistral are capable, improve constantly, and can be swapped as the market moves. Treat the model as a component you choose per task, not a core asset you own.
Build the product
What makes an AI feature valuable is rarely the model. It's everything around it:
- The workflow: where the AI fits into what people already do
- The data: what the model sees, how it's cleaned, and what it must never see
- The experience: how people understand, correct and trust what the AI produces
- The evaluation: how you know it works before customers find out it doesn't
These are specific to your business, and they're where the defensibility is.
Integrate, don't replace
Most businesses don't need a new system. They need AI plugged into the stack they already run: an assistant that reads from the CRM, an automation that drafts replies, a classifier that routes requests. Each step should be logged, and anything customer-facing should get a human check.
When building more makes sense
Go deeper when the AI behaviour itself is the product, when your data is genuinely unique and plentiful, or when requirements around data handling rule out hosted options. Even then, start with a hosted model to prove the product, and invest in custom work once you know what "good" looks like.
Where to start
Pick one workflow with clear value and a measurable outcome. Build the smallest version that works end to end, evaluate it honestly, then expand. And if AI turns out to be the wrong answer for that workflow, that's a result too.