Make Retention Decisions You Can Explain
DecideX reads your account notes, predicts churn with a calibrated model, and returns a typed recommendation with its confidence and evidence. Your team reviews every call, and every outcome makes the model better.
Illustrative Example
Northwind Dental Clinic
- Signals From NotesComplaint detected · Competitor mentioned
- Calibrated PredictionHigh churn probability from account data
- Typed DecisionIntervention · Escalate · Handle today
- Human ReviewApprove, override, then record the outcome
How It Works
Four Engines, One Accountable Decision
Each part does the job it's best at. AI reads the notes, a prediction model estimates the risk, a decision engine recommends the next step, and your own rules decide what gets flagged.
- 01UnderstandAI Note Reader
Read the Unstructured Evidence
Support notes, emails and call summaries become structured signals: sentiment, complaints, competitor mentions and cancellation intent. Only quotes that appear word for word in your notes are kept as evidence.
- 02PredictPrediction Model
Estimate Churn Probability
A model trained on the public IBM Telco dataset scores the account. Isotonic calibration means a 70% prediction really does churn about 70% of the time, and each prediction lists the features that moved it.
- 03DecideDecision Engine
Answer Typed Questions
The decision engine weighs the prediction and the notes together, answering four clear questions: the action, the owning team, the urgency and whether to escalate. Every answer comes with probabilities and a confidence score.
- 04ReviewYour Team
Keep Humans in Charge
Low-confidence or conflicting calls are flagged for review. Reviewers approve or override, then record whether the customer actually left. Those outcomes retrain the model.
Features
Built for Decisions That Need to Hold Up
Everything a retention team needs to act quickly and explain every call afterwards.
Calibrated Probabilities
Predictions are calibrated and graded on a held-out test set with ROC-AUC, PR-AUC, Brier score and expected calibration error.
Evidence You Can Check
Every decision shows the quotes, signals and feature impacts behind it, so reviewers can see why a call was made.
Typed, Bounded Answers
Every recommendation is one of four clear actions and one of four teams, never vague advice, so your team knows exactly what to do.
Human Overrides
Reviewers can override any decision. DecideX tracks how often AI and humans agree and where they differ.
Outcome Feedback Loop
Record whether each customer churned. Real outcomes measure live accuracy and feed the next training run.
Drift Monitoring
Population stability checks warn you when the customers you score start to look different from the training data.
Graceful Degradation
If an AI service is down, simple backup rules step in and the decision is flagged for a person to check.
Private by Default
Your decisions stay in your workspace. You can export or delete them at any time, and gender and age are excluded from the model.
The Model
Real Metrics From the Model
Measured on customers the model never saw while it was learning. As you record real outcomes, your own accuracy appears on the AI Performance page.
Churners Caught
86%
Share of customers who really left that the model flagged in advance
Ranking Quality
0.84
How often a leaver is ranked riskier than a stayer
Churn Detection Score
0.65
How precise the model stays as it catches more churners
Probability Error
0.02
How far a stated chance is from what really happens. Lower is better
Model Comparison
Four candidates trained on the same split. The winner is chosen by validation PR-AUC, then calibrated.
| Model | Accuracy | Precision | Recall | F1 | ROC-AUC | PR-AUC | Brier |
|---|---|---|---|---|---|---|---|
| Baseline (Class Prior) | 73.5% | 0.0% | 0.0% | 0.000 | 0.500 | 0.265 | 0.195 |
| Logistic Regression | 80.2% | 64.8% | 55.6% | 0.599 | 0.841 | 0.630 | 0.138 |
| Random Forest(selected) | 79.2% | 64.4% | 48.4% | 0.553 | 0.841 | 0.644 | 0.138 |
| LightGBM | 79.8% | 65.0% | 51.6% | 0.575 | 0.841 | 0.644 | 0.138 |
Candidate metrics use a 0.5 cut-off before calibration. The headline figures above are for the calibrated model at its cost-tuned threshold.
Trained on Public Data
IBM Telco Customer Churn: 7,043 customers with a 26.5% churn rate, released under Apache-2.0. Protected attributes are excluded so recommendations do not depend on gender or age.
Thresholds Set by Cost, Not Habit
A missed churner is treated as 5x as costly as an unnecessary intervention. That puts the churn threshold at 21% instead of the usual 50%, trading some false alarms for far fewer missed churners.
Why DecideX
More Than Asking a Chatbot
A single prompt can sound confident without being right. DecideX splits the job so each answer can be measured.
| Question | Asking a Chatbot | DecideX |
|---|---|---|
| Where the probability comes from | Generated text | A calibrated model graded on held-out data |
| Answer format | Free-form prose | Typed answers from fixed options |
| Confidence | Rarely stated | A probability distribution for every answer |
| Evidence | Can be invented | Verbatim quotes and measured feature impacts |
| When a service fails | No answer | A transparent fallback policy, flagged for review |
| Learning from outcomes | No | Outcomes measure accuracy and retrain the model |
FAQ
Questions Worth Asking
How DecideX reaches a recommendation, and what happens when something goes wrong.
Do I need an account?
No. The first time you open the dashboard, your browser gets a private workspace key. There are no passwords or emails. To use the same workspace on another device, copy the key from Settings and paste it there.
What does DecideX actually decide?
For each customer, DecideX recommends one of four actions (intervention, retention offer, monitor or no action), the team that should own it, how urgently to act, and whether to escalate to a senior manager. Each answer includes probabilities and a confidence score.
Where do the churn probabilities come from?
From a machine learning model, not a language model. DecideX compares a baseline, logistic regression, random forest and gradient boosting on the public IBM Telco churn dataset, picks the best by validation PR-AUC, and calibrates it so the probabilities are meaningful.
What happens if an AI service is down?
The decision still completes. If the decision engine can't be reached, simple backup rules recommend an action based on the predicted risk, and the decision is marked Partial and flagged for a person to check. If the notes can't be read, the decision continues without them and says so.
Can my team override a recommendation?
Yes. Reviewers can approve or override the action and the escalation call on any completed decision, and add a note. The dashboard tracks how often AI and human reviewers agree.
How does the model improve over time?
When you record whether a customer churned or stayed, that outcome measures live accuracy and becomes training data. An administrator can retrain the model at any time. A new version only goes live automatically if it performs at least as well on the same held-out test set.
Is my data shared with other users?
No. Your decisions live in a private workspace tied to a random key stored in your browser. Anyone without that key can't see them. Recorded outcomes are used, without names or notes, to retrain the shared model. You can export or permanently delete your workspace from Settings at any time.
Your Next Retention Call, Explained
No sign-up needed. Load a real customer from the dataset and see every stage of the decision in under two seconds.