As of 13 August 2026, AI can only partly predict which customers may leave.

What to actually do

1. Hand it to a person

   The route this page recommends

   A person who owns the outcome does this end to end, worth it when the failure is dear.

2. Use a tool built for this

   Second choice

3. Do it yourself

   The distant third

   A chat interface, power-user skill, and roughly 1 hour until you can act on the result.

   How to actually do it

   1. Open the customer data export and a separate document containing the business definition of churn, the prediction period and the actions your team can take.
   2. Remove unnecessary personal fields, assign a stable non-identifying customer ID, and record which fields were known before each prediction date.
   3. Gather historical records that show both customers who left and customers who stayed, then attach a data dictionary explaining every column, date and missing-value code.
   4. Paste the business context, data dictionary and a representative sample into the prompt, or upload the prepared file to an AI analytics tool such as Akkio.
   5. Ask the model to check leakage, duplicates, missing values and time ordering before it builds a model, and reject any feature that became known only after churn.
   6. Compare the reported predictions with later known outcomes, checking precision, recall, false positives, missed leavers and performance across relevant customer groups.
   7. Give the risk list to a colleague who understands the customer process, agree a limited contact test, and record whether the intervention changes retention without treating the prediction as proof about an individual customer.

   Prompt

   Act as a cautious customer-analytics analyst. I want to predict which customers may leave using the dataset and business context below.
   
   Business context:
   - Product or service: [describe it]
   - Definition of churn: [state exactly what counts as leaving]
   - Prediction window: [for example, leaving within the next period]
   - Actions available to the business: [list them]
   - Important exclusions or constraints: [list them]
   
   Dataset:
   [paste a data dictionary and a representative sample, or attach the file]
   
   Do the following:
   1. Check whether the data contains a usable historical churn outcome, customer identifier, prediction date and features available before that date.
   2. Identify missing values, duplicate customers, inconsistent dates, class imbalance, possible leakage and fields that should not be used.
   3. Propose a time-aware train and test split. Do not use information that became available after the prediction date.
   4. Build or describe a baseline and a churn model, then report precision, recall, the confusion matrix and how performance changes at practical contact thresholds. If you cannot calculate a measure from the supplied data, say so instead of estimating it.
   5. Explain the main factors associated with higher predicted risk, clearly distinguishing association from cause.
   6. Recommend a threshold and customer-contact test plan based on the available actions and their capacity. Do not recommend automatic penalties, service withdrawal or decisions based on protected characteristics or their proxies.
   7. State what the model cannot establish, how it could fail, and what checks a colleague should complete before using any score.
   
   Return: data-quality findings, modelling approach, validation results, a table of customer IDs with predicted risk if justified by the data, recommended next steps, and a short plain-English summary for management. Never invent data, performance figures or customer reasons. Flag any conclusion that is not supported by the supplied dataset.

   Open it prefilled in ChatGPT or Claude, or copy it into Gemini, which takes no prefill link.