Training a custom AI model and integrating an existing one (GPT-4-class models, retrieval-based systems) are different projects with different costs, timelines and risk profiles. Most SMEs answer this question correctly by asking what problem they actually have.
What "buy" actually looks like
Integrating an existing model means building the application layer around a capability that already exists: a chatbot grounded in a company's own documents via retrieval, or a phone agent that uses an established speech pipeline. The model itself is not being trained from scratch; the engineering work is in retrieval, prompt design, integration with existing systems, and testing against real conversations. This is the route that gets a working pilot in two to four weeks.
What "make" actually requires
Training or fine-tuning a custom model is justified only when the task is specific enough, and valuable enough, that an off-the-shelf model's general capability genuinely falls short, and even then, most of what looks like a "custom model" need is actually solved by grounding a general model in the company's own data via retrieval, which is far cheaper and faster than training. Real custom training requires labelled data, ongoing retraining as conditions change, and specialist skill to maintain: a different cost structure entirely from integration work.
The decision framework
Three questions decide it. Is the task genuinely specific to the business, or is it a version of a problem (answering questions from documents, transcribing calls, classifying email) that general models already handle well when grounded in the right data? Is there a realistic budget and timeline for ongoing model maintenance, or does the business need something working and stable within weeks? And is the expected volume high enough to justify the fixed cost of custom training against the marginal cost of an API call to an existing model?
For the large majority of SME use cases (customer service, document processing, phone handling), the honest answer is buy-and-integrate. Custom training earns its cost at a scale and specificity most SMEs have not reached yet, and reaching for it early usually means months of work to match what a well-grounded existing model already does.
Takeaway
Default to integrating an existing model grounded in the company's own data: it is faster, cheaper, and covers most SME use cases. Reach for custom training only once a specific, high-volume task has proven that general models genuinely fall short.