01Decision Support, Not Advice
Recommendations are suggestions for your team to review. They aren't financial, legal or professional advice, and they shouldn't be the only basis for decisions that significantly affect a person.
02What the Model Learned From
The prediction model is trained on the public IBM Telco Customer Churn dataset, which describes telecom subscribers. Its probabilities are most reliable for customers who resemble that data.
If your customers differ a lot, for example in pricing or contract structure, predictions may be less accurate. The drift monitor on the dashboard helps you spot this.
03Models Make Mistakes
- The churn model produces false positives and false negatives. The model card shows how often on held-out data.
- The language model can misread tone or miss signals in notes.
- The confidence score shows how sure the AI is, not a guarantee that it's right.
04Backup Decisions
When the AI decision engine is unavailable, simple backup rules make the recommendation from the predicted risk. These decisions are labelled Partial, have no confidence score, and are always flagged for a person to check.
05Fairness
Gender and senior-citizen status are excluded from the model. Other attributes can still correlate with protected characteristics, so review decisions for unintended patterns.
06No Guarantee of Outcomes
Following a recommendation doesn't guarantee a customer will stay. Recording real outcomes is the best way to measure how well DecideX works for you.