After years of building AI systems in production, I changed how I start.
I used to start from the model. What could be predicted, optimized, classified. What the data would support.
Now I start from the decision.
What is the person actually trying to do? Who owns the outcome? What context needs to persist? What happens after the first answer?
That shift sounds small. It is not.
Consider retail inventory planning. A good forecasting model is useful. But forecasting demand is only one part of the problem.
You still have to decide how much to order, where to allocate it, how to react to stockouts, how to adjust when forecasts change, and how to balance cost, service level, and space constraints, all at once, all dependent on each other.
Each step depends on the previous one. The context keeps evolving.
When you start from the model, you build a system that answers the forecasting question well. When you start from the decision, you realize the forecasting question is downstream of a dozen others, and that the system needs to hold across all of them, not just one.
You can have a very good forecasting model and still end up with empty shelves, excess inventory, or decisions nobody trusts.
The model was not wrong.
The system was not designed for real decision flow.
The more I worked on real systems, the more it felt like this was the core problem. Not the quality of the models. The fact that the systems were not designed for how decisions actually work, multi-step, context-dependent, evolving.
Starting from the decision changes what you build. It changes what you measure. It changes what counts as working.
That idea is at the center of what I am building at Toutami: AI-native systems designed for multi-step decisions and real workflows, not just isolated predictions. Systems that stay with the problem as it unfolds.
Once you start looking at problems this way, a lot of current AI products start to feel incomplete. Not because the models are bad, but because the layer between the model and the decision was never built.
That layer is what we are building.