The Bar Moved

A few years ago, adding AI to a product meant hiring people who could train and serve models. Today, the strongest models are one API call away, and the hard part is product work, not machine learning.

Most companies I talk to do not need a data science team. They need someone who can wire a good model into a real workflow and handle the messy parts.

The Stack I Reach For

• A frontier model API for the intelligence (Claude, and others where they fit)
• A vector store only when the feature actually needs retrieval
• A normal database and app framework for everything else
• Evals, so I can tell if a change made things better or worse

That is it. No training, no GPUs, no MLOps platform.

Where the Real Work Is

The model is maybe twenty percent of the effort. The rest is product:

• Designing the prompt and the surrounding context so the model has what it needs
• Handling the cases where the model is unsure or wrong
• Making the feature fast enough that people actually use it
• Deciding what happens when the API is slow or down

None of that requires a PhD. It requires someone who ships.

Start Small and Measurable

Pick one feature where AI clearly helps, give it a clear success measure, and ship it behind a flag. Watch real usage, fix what breaks, then expand. You will learn more from one live feature than from six months of planning a platform.

The companies winning with AI right now are not the ones with the biggest teams. They are the ones shipping.