When you actually need a custom ML model, and when you don't

Most businesses that think they need a custom model actually need an off-the-shelf API wrapped around their own data. It's faster, cheaper, and usually good enough.

The signal that you genuinely need something custom is specificity: a prediction or classification problem that's unique to your business, with enough historical data to actually train on, where a general-purpose model consistently gets it wrong.

Before committing to a build, we run a short feasibility pass — is there enough clean data, is the problem actually learnable, and is the accuracy bar realistic. Skipping this step is the single biggest reason ML projects stall out.

If it does make sense, the model is only half the project. Deployment, monitoring for drift, and a retraining schedule matter just as much as the training itself — a model that's accurate on day one and ignored after that quietly gets worse.