As of 13 August 2026, AI can only partly predict which leads will become customers.

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 CRM export and write down the exact prediction point, the customer outcome, the sales-cycle window and the columns that were available at that point.
   2. Remove names, email addresses, phone numbers and other unnecessary personal data, then export a representative sample containing both converted and non-converted leads.
   3. Paste the business context and de-identified sample into a chatbot with the prompt above, and ask it to identify missing fields, leakage and unsuitable features before proposing a model.
   4. Save the proposed baseline, feature list and time-based testing plan, then ask the chatbot to produce code or spreadsheet instructions that match your actual column names.
   5. Run the analysis on a later holdout period that was not used to design the score, and record predictions alongside the eventual customer outcomes.
   6. Compare the results with your existing lead-priority method, checking precision, recall, conversion-rate lift and calibration rather than relying on a single ranking.
   7. Ask a sales-operations or data colleague to inspect the data split, excluded fields, subgroup results and recommended actions before putting the score into the CRM.
   8. Run the score as decision support, review false positives and false negatives with the sales team, and set a date to retest it against new outcomes.

   Prompt

   Act as a sales-operations analyst. Help me design and test a model that ranks leads by the likelihood that they will become paying customers.
   
   Business context:
   - Product or service: [describe it]
   - Typical sales process and stages: [describe them]
   - Definition of a lead: [define it]
   - Definition of a customer: [define the conversion event]
   - Prediction point: [for example, when the lead first enters the CRM]
   - Sales cycle length: [state the usual range if known]
   - Available fields: [list the CRM columns]
   
   I will paste a de-identified CRM sample below. Do not invent missing values, outcomes or performance figures. First identify data-quality problems, class imbalance, duplicate records, time leakage and fields that would not have been known at the prediction point. Then propose a simple baseline and a more useful scoring approach. Explain which fields to include and exclude, how to split the data by time, and how to test the model on leads it has not seen.
   
   Return:
   1. A plain-English plan.
   2. A data-cleaning checklist.
   3. A suggested feature list with the reason for each field.
   4. Pseudocode or reproducible code, clearly labelled if it needs adaptation.
   5. Evaluation measures including precision, recall, conversion-rate lift and calibration, without fabricating results.
   6. A table template for recording actual results after the prediction period ends.
   7. Practical safeguards against over-prioritising one type of lead or using personal data unfairly.
   8. The limits of the prediction and the decisions that must remain with the sales team.
   
   Do not claim that correlation proves causation. Do not recommend automatically rejecting or suppressing leads. Ask questions where the data or business definition is insufficient.
   
   CRM sample:
   [paste de-identified rows here]

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