Models that anticipate which customer leaves, who is worth more, when and what they'll buy. Serious predictive analytics turns your CRM into a commercial operating system — and reduces decision cost without doubling your team.
Enough transaction volume to train models: propensity, repurchase, customer value, churn. An online SME with traction is an ideal candidate.
If your CRM has structured activity with sufficient data, scoring and propensity models pay back from the first quarter.
Recurring businesses where churn is a critical metric. A 30-90 day cancellation prediction model enables proactive retention.
Active loyalty programme with individual purchase history. Predictive models for dynamic segmentation, recommendations, basket size.
I review CRM, GA4, ERP, platforms. Data quality, gaps, biases, freshness. If the data doesn't support modelling, it must be cleaned before training anything.
Prioritisation of models by impact and viability: churn, purchase propensity, lifetime value, upsell propensity, dynamic segmentation. Not all at once.
Model construction with rigorous validation (train/test split, cross-validation, accuracy metrics). Comparison against baseline to ensure real improvement.
The trained model only adds value if integrated into operations: dashboards in your BI, alerts for sales, automation with marketing automation. Without integration, it's paper.
More than a ‘tech buzzword’, predictive analytics properly implemented is a commercial efficiency lever. Concretely:
Churn models alert 30-90 days ahead. Sales acts with margin — discount, session, call — instead of reacting to the cancellation.
Predictive scoring prioritises the 20% of leads that generate 70% of sales. Sales stops burning time on what doesn't pay back.
A lifetime value model lets you treat each customer according to their potential — not their past spend. Advertising investment, service, retention.
Product recommendations based on behavioural patterns, not fixed rules. AOV and conversion rise without manual effort.
Dynamic segments based on real behaviour and evolution prediction. Each customer treated their way, without requiring manual work.
Demand prediction by product, moment, region. Procurement, logistics and marketing aligned on the same projection — not on different intuitions.
Online store with large base but flat conversion. Propensity model + segmented automation. Revenue per user +24% in four months.
El marketing del cerebro es más predictible que el marketing de la opinión. — Ángel Ortega Castro
B2B SaaS platform with high churn. Predictive model trained on 18 months of data. Proactive retention programme. Annual churn down 18 points.
Retail chain with 80K loyalty members. Dynamic segmentation by value, frequency, propensity. Personalised promotions with 3× redemption.
CRM, events, content, campaign history.
Trained or generative algorithm guided by rules.
Lead prioritised, content published, decision taken.
Predictive analytics makes sense when there is sufficient data, minimum analytical culture and willingness to govern the model. Four typical scenarios:
Without enough history, the models don't converge. If you have less than 12 months, starting is premature — better to invest in data setup first.
Sales attends to whoever shouts loudest, not who is worth more. Marketing decides segments by feeling, not by data. The model changes that from the root.
Month-by-month departures with no clear pattern. The model identifies the early signals of leaving — and allows intervention before they materialise.
Stock, HR, logistics, marketing — all operational functions improve with serious prediction. The difference from traditional forecasting is granularity and continuous adjustment.
As a practical rule, minimum 600 conversions per month over 12 months for propensity and churn models. For LTV or segmentation you can train with less but with adjusted expectations. Without sufficient volume, I recommend investing first in basic analytical setup.
It depends on the case. For SME: Python + scikit-learn / XGBoost for training + integration with your BI (Looker Studio, Power BI). For large organisations: more mature stacks (BigQuery ML, Databricks, AWS SageMaker). What matters is the use case — the technology fits.
Your team, with maintenance guidance documented at the end of the project. Models are retrained every 3-6 months to avoid ‘model drift’. If you don't have an internal profile, we set up a light retainer for periodic maintenance.
A combination. For many use cases, classic statistical models (regression, clustering) are as good as AI and more interpretable. I only scale to advanced machine learning when the problem justifies it. AI isn't the best solution by default.
Critical from day one. Anonymisation, aggregation, consents, transparency towards the customer about the use of predictive models in commercial decisions. European regulation (GDPR, called RGPD in Spanish/French) is strict and best complied with without shortcuts.
First 45-minute session, free of charge and no commitment. If we fit, I send you a detailed proposal within 5 days. If we don't, you take away a useful initial diagnosis.