Start with a question you can evaluate.
Turn a broad goal into a prediction task with a clear outcome, usable examples, and inputs that reflect what you will actually know at prediction time.
A useful prediction model starts with a question whose answer you can check. “Improve sales” is a goal. “Will this opportunity close within seven days?” is a task with an outcome and a time window.
Define the outcome first
Decide what counts as a positive outcome and when the answer becomes known. If a deal closes after the seven-day window, it is not a positive example for that question. Consistent labels make training and evaluation easier to interpret.
For a classification task, define the available categories. For a scoring task, write down what each level means. If two people would label the same example differently, resolve the rubric before training.
Use the information you would have had
Build each example from the information available when the prediction would have been made. A final deal status or a note written after closing can reveal the answer. Including it makes a historical test less useful for judging a model on new cases.
Ask of each field: would this value have existed at the time? If the answer is unclear, inspect the source and timing before including it.
Review a small set of examples
Read the inputs and expected answers together. Look for missing context, conflicting labels, and cases that do not belong to the task. This review helps uncover problems before they become a training run.
Tuned Predictions currently starts with supported Attio, Pylon, or Fireflies data. Other sources require a conversation with the team, and initial training and deployment need setup support.
Explore the product workflow or integrations to plan your starting point.