As of 13 August 2026, AI can only partly predict which customers will leave your business.

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 your customer database or spreadsheet and gather historical records containing customer identifiers, dates, purchases or usage, contact history, cancellations and the reason or status recorded when a customer left.
   2. Write down one operational definition of churn, including which event counts as leaving and the future period in which it must occur, then identify the column that records that outcome.
   3. Remove unnecessary direct identifiers and sensitive fields, document every remaining column, and check that each feature would have been known before the retention decision.
   4. Paste the column descriptions and a representative, suitably minimised extract into the prompt, then ask the model to identify missing data, duplicate records, leakage and inconsistent labels before modelling.
   5. Use Akkio or another suitable analytics tool to import the cleaned historical data, select the churn outcome, and create a baseline prediction model rather than treating the first score as reliable.
   6. Hold back a later portion of the historical data as an untouched test set, compare predicted risk with what happened afterwards, and ask a data-literate colleague to check false positives, false negatives and whether the model relies on unsuitable fields.
   7. Run a limited retention trial using transparent, proportionate actions, record the action and later outcome, and compare results with the agreed validation plan before expanding its use.

   Prompt

   Act as a customer-retention data analyst. Using the customer dataset I provide, help me build a churn prediction approach for [business type]. The dataset contains [describe each column], and a customer counts as having left when [precise definition]. The historical outcome column is [column name].
   
   First, inspect the data description and identify missing values, duplicate records, inconsistent labels, possible personal data, target leakage and features that would not be available before a retention decision. Do not infer facts that are not present. Then propose a reproducible modelling plan using a time-based train and test split where the data permits it. Explain which columns should be excluded and why.
   
   Produce:
   1. A plain-English definition of the prediction target.
   2. A data-quality and leakage checklist.
   3. A suitable baseline and a small set of candidate models.
   4. The evaluation measures that matter for this business, including the trade-off between false positives and false negatives.
   5. A table structure for customer risk scores that includes the predicted risk, the main supporting factors and an uncertainty note.
   6. Retention actions that are proportionate and useful, without making decisions solely from sensitive or proxy attributes.
   7. A validation plan comparing predictions with later outcomes.
   
   Do not claim that a customer will definitely leave. Do not recommend contacting or excluding anyone solely because of a model score. State clearly what cannot be concluded from the data. If the dataset is insufficient, say exactly what is missing and provide a safer analysis instead.

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