Most AI features in production applications don't need a custom-trained model. A well-chosen third-party API covers general text generation, standard image classification, and a long list of common tasks faster and cheaper than training something from scratch. The question worth asking isn't "should we use AI" — it's whether your specific task fits inside what a general-purpose API was actually trained to do.
When an API is the right call
If the task is close to what large, general-purpose models are already good at — summarizing text, answering questions about common knowledge, classifying images into everyday categories — a hosted API almost always wins on cost and time to ship. There's no dataset to collect, no training pipeline to maintain, and the model improves over time without any work on your end.
When custom training earns its cost
Three situations tend to justify training a model instead:
- The data domain is specific. Movement and motion data, specialized imagery, or industry-specific signals often don't resemble anything a general model was trained on. A rehabilitation platform analyzing human movement patterns, for example, needs a model trained on movement data specifically — a general-purpose vision API isn't built for that.
- Data can't leave your infrastructure. Healthcare, legal, and other regulated contexts sometimes rule out sending data to a third-party API entirely. A model that runs inside your own infrastructure sidesteps that constraint.
- Cost or latency don't work at your scale. Per-request API pricing can become expensive at high volume, and network round-trips add latency a locally-deployed model avoids.
What training actually involves
In practice: preparing and cleaning the dataset, choosing an appropriate model architecture (commonly in PyTorch or Keras), training against that data, evaluating on held-out examples, and packaging the result behind an inference API so the production application can call it the same way it would call any other service. The model isn't the deliverable — the deployed, callable version of it is.
The practical middle ground
These aren't mutually exclusive. Many production systems use a third-party API for general capabilities and a custom-trained model for the one task that's actually specific to the business. Scoping AI work around the specific problem, rather than "adding AI" as a category, is what makes the decision tractable.