Good decisions start with data. They don't end there.
In a world of constant optimization, it's natural to trust the numbers. Data makes decisions cleaner, faster, and often more accurate. Whether pricing risk, forecasting demand, or deciding who to hire, data provides structure. The signal.
But don't mistake data for the full picture.
In the rush to scale decision-making, perspective is often lost that every model is a trade-off. You collapse complexity into variables, compress behavior into probabilities, and surface patterns for predictions. That's the magic—and the limitation. Behind every input is a story that didn't make it into the dataset. Over time, those exclusions shape outcomes.
Say you're building a risk score. You train on past defaults. The model flags short job tenures, missed rent, lower education. It's statistically sound. But it misses the reasons, which might be volatility due to illness, caregiving, or a broken system. The prediction is right on average, but blind to exceptions. But the real world doesn't follow clean curves. People switch industries, bounce back, start over. Markets shift. Incentives change. The signal gets noisy. If your model can't adapt to that noise, it fails subtly—quietly missing the people or moments that didn't look right but mattered in the long run.
When deciding with data, you're also including what the data doesn't say. A strong model will outperform human intuition most of the time when it comes to overall outcomes. But models aren't neutral. They're shaped by their training data and constraints. When we reduce a person to a single probability, the decision process also becomes more rigid.
It's easy to assume the model got it right, especially when it confirms our priors. But smart operators test, challenge, and question their models. They evolve because conditions change—and so does the cost of being wrong. The best decisions are rarely fully automatic. Data should guide judgment, not replace it. Strong decision systems don't just predict—they allow for deep-dives, understanding of biases, and well-judged overrides. They leave space for context to make things smarter, and fairer.
The answer isn't to humanize every decision but to be thoughtful about when automation makes sense. Build systems that handle the routine but flag the edge cases. That recognize the difference between patterns worth trusting and those that deserve a second look.
Because data is great at telling you what's likely. It's your job to decide what's worth doing.