The most honest thing you can say about the latest wave of AI decision models is that they work best when you have no data to work with. That's a useful capability, but it is not the universal solution the hype suggests. A recent test from Voyage AI demonstrates this clearly: when you have even a modest set of labeled examples, the old-school approach of embedding your inputs with a cheap model and fitting a simple sklearn classifier significantly outperforms the flashier zero-shot alternatives, by a 37% error reduction. This is not a knock on innovation; it is a reminder that the right tool depends on the problem. As we noted in How Your Agent's Tools Can Make Smarter Decisions, decision models are best understood as tools, not magic. And as Forecasting a product launch without history requires a smarter approach shows, when you lack historical data, you need a fundamentally different strategy.
For most teams, the reality is that you already have labels. You have past customer behavior, historical sales figures, or curated training sets from previous projects. The instinct to reach for the newest zero-shot model is understandable, it feels like progress. But the Voyage test reveals that progress sometimes means going back to fundamentals. Embedding your inputs with an open-source model and training a simple classifier on your own labels is cheaper, faster, and more accurate. It is also more transparent. You understand what your model is doing, and you can iterate on it. The zero-shot model, by contrast, is a black box that generalizes broadly but misses the specific patterns in your data. The lesson is practical: before you adopt a new tool, ask whether you already have the data to make a simpler one work better.
This is not to say zero-shot models have no place. The Voyage team is clear that their approach wins when you have labels and loses when you do not. That is a fair trade-off. If you are launching a product with no history, or entering a market where you have zero prior examples, a zero-shot decision model like Jev is your best bet. But if you are sitting on a database of past outcomes, the most innovative thing you can do is use it. The hype around AI often frames simplicity as outdated, but the data says otherwise. A 37% error reduction is not a marginal gain, it is the difference between a model that guesses and one that knows.
The open question is how many teams will actually take this approach. The code is public, the method is straightforward, and the results are compelling. Yet the temptation to chase the newest name in decision models will persist. What we should watch for is whether the industry begins to treat zero-shot models as a specialist tool for low-data scenarios rather than a general replacement. If the Voyage test holds up across more use cases, the smartest move for most organizations will be to invest in labeling and embedding, not in hype. That is a concrete takeaway: your existing data is your greatest asset, and the simplest model trained on it will often outperform the most sophisticated one that ignores it.