AI Development in Lisbon, Portugal

Financial Services / Banking

A Portuguese retail bank (2.8M customers, EUR 45B in assets, 350 branches across Portugal) wanted to improve customer retention and cross-sell effectiveness using AI. The problems: (1) Customer churn: annual customer churn rate of 6.2% (174,000 customers leaving annually). Each departing customer: representing EUR 1,200-2,400 in lost annual revenue (interest margin, fees, insurance commissions). Total annual churn cost: estimated EUR 250M in lifetime value lost. The bank's retention efforts: reactive — contacting customers only after they initiated account closure. By that point: 85% of decisions were irreversible. (2) Cross-sell inefficiency: the bank's 1,400 relationship managers made product recommendations based on basic segmentation (age bracket, income band, product holdings). One-size-fits-segment marketing: personal loan offers sent to all 30-45 year-olds in income band B, regardless of individual financial behaviour. Response rate: 1.8% for email campaigns, 3.2% for branch-based recommendations. The low response rate: frustrating customers with irrelevant offers and wasting marketing budget. (3) Portuguese language NLP gap: the bank's customer interactions (call centre transcripts, email correspondence, branch visit notes, complaint records) contained valuable signals — but all in Portuguese. International AI vendors (Salesforce Einstein, Microsoft Dynamics AI) performed poorly on Portuguese text analysis, particularly European Portuguese. Sentiment analysis: 62% accuracy on Portuguese text (vs. 89% on English) — insufficient for reliable customer intelligence. (4) EU AI Act compliance: the bank's legal team flagged that AI used for customer scoring and product recommendations in financial services would likely be classified as high-risk under the EU AI Act — requiring documented risk management, bias assessment, and human oversight. Any AI deployment needed to meet these requirements from day one — retrofitting compliance would be significantly more expensive.

We built a customer intelligence platform over 14 weeks: (1) Churn prediction: a machine learning model predicting the probability of customer churn within the next 90 days. Features: 180+ features derived from customer data — transaction patterns (declining transaction frequency, decreasing average balance, increased ATM withdrawals — potential sign of moving funds elsewhere), product engagement (declining app logins, reducing card usage, stopped direct debits), customer interactions (complaint frequency, call centre contacts, branch visits), and life events (salary deposits stopping — potential job change, large outgoing transfers — potential property purchase at another bank). Model: LightGBM (gradient-boosted trees) — achieving 84% accuracy in identifying churning customers 90 days in advance (AUC-ROC: 0.91). The model: validated against 12 months of historical churn data — correctly identifying 84% of customers who actually churned, while maintaining a false positive rate below 15% (ensuring retention efforts aren't wasted on customers who aren't actually at risk). (2) Portuguese NLP for customer signals: fine-tuned XLM-RoBERTa on 500,000 Portuguese customer interaction records (call transcripts, emails, complaint records, branch notes). The model detecting: churn intent signals ("estou a pensar mudar de banco" — "I'm thinking of changing banks", "não estou satisfeito com as condições" — "I'm not satisfied with the conditions"), product interest signals ("gostava de saber mais sobre crédito habitação" — "I'd like to know more about mortgage", "têm algum plano de poupança?" — "Do you have any savings plan?"), complaint severity (distinguishing serious complaints from routine inquiries — important for prioritising retention interventions), and emotional tone (frustrated, neutral, satisfied — Portuguese emotional expression differs from English, requiring Portuguese-specific sentiment training). NLP accuracy: 87% sentiment accuracy on European Portuguese (vs. 62% with off-the-shelf English-trained models) and 82% intent detection accuracy. (3) Next-best-action recommendation: combining churn prediction, NLP signals, and customer financial profile to recommend: for at-risk customers: the optimal retention action (product upgrade, fee waiver, personal call from relationship manager — the model learning which intervention works best for which customer profile), for engaged customers: the most relevant product recommendation (not just "personal loan for everyone" but "this specific customer shows signs of property purchase intent — recommend mortgage consultation"), and for all customers: the optimal communication channel (some customers respond to app notifications, others to email, others to branch contact — the model learning channel preference from historical interaction data). (4) Relationship manager interface: an iPad application used by the 1,400 relationship managers during branch visits and customer calls. For each customer: the app displaying a "customer intelligence card" — churn risk score (high/medium/low with contributing factors), next-best-action recommendation (with talking points in Portuguese), product affinity scores (which products the customer is most likely to be interested in), recent interaction summary (NLP-processed — key topics, sentiment, unresolved issues), and customer lifetime value estimate. The interface: designed for use during customer conversations — the RM glancing at the card before or during the meeting, not reading a detailed report. (5) EU AI Act compliance: complete documentation package — technical documentation (model architecture, training data provenance, feature descriptions, performance metrics by demographic group, limitation analysis), risk management (identified risks: false positives leading to unnecessary retention offers, potential bias against specific demographic groups — mitigated by bias testing and demographic-specific performance monitoring), human oversight (RMs make final decisions — the AI recommends, the human acts; override logging to track when RMs reject AI recommendations and why), and bias monitoring (model performance tracked across age groups, gender, geographic region, and nationality — ensuring the model doesn't systematically underserve any customer segment).

Churn reduction: customer churn rate from 6.2% to 4.1% in the first year — a 34% reduction. In absolute terms: 58,800 customers retained who would have otherwise churned. At EUR 1,800 average annual revenue per customer: EUR 105.8M in preserved annual revenue. Retention intervention success: when the model identified a customer as high churn risk and the RM contacted them with a personalised retention offer: 42% of at-risk customers retained (vs. 15% retention rate for the previous reactive approach — contacting customers only after they initiated closure). The proactive approach: 2.8x more effective than reactive retention. Cross-sell: next-best-action recommendations improving cross-sell response rates: email campaigns from 1.8% to 5.4% (3x improvement), branch recommendations from 3.2% to 8.7% (2.7x improvement). Annual cross-sell revenue: EUR 12M incremental (additional products sold through AI-powered recommendations). Portuguese NLP: customer interaction analysis providing early warning signals previously invisible. One example: the NLP detecting increasing frustration in a corporate customer's email correspondence over 3 months — the RM intervening with a personalised meeting, retaining a EUR 2.4M commercial relationship. The bank's head of retail: "Before this system, we were flying blind. Now we know which customers are at risk and why — and we can do something about it before they leave." EU AI Act: compliance documentation completed and reviewed by the bank's DPO and legal team. The documentation: serving as a template for the bank's other AI deployments. Banco de Portugal: during a routine supervisory review, the bank demonstrated its AI governance framework — receiving positive feedback on the bias monitoring and human oversight implementation. RM adoption: 89% of relationship managers using the customer intelligence cards daily. RM feedback: "The talking points are exactly right — I sound like I understand the customer's situation, not like I'm reading from a script." Training: 4-hour workshop with each RM team — enough to demonstrate the tool's value and build confidence. Development cost: EUR 280,000 (14 weeks). Annual value: EUR 105.8M (churn-preserved revenue) + EUR 12M (cross-sell improvement) = EUR 117.8M annual value. ROI: development cost recovered within first 21 hours of customer retention.