EL DOLOR (The Pain)
In 2024, Spain's mortgage market experienced a 64% increase in complaints compared to the previous year. More than 10,000 citizens filed grievances with the Bank of Spain, with 30.6% of all banking complaints directly linked to mortgage loans. The financial impact was staggering: over €4 million returned to consumers due to abusive clauses and misleading practices.
People were signing 40-page mortgage contracts they didn't understand, then feeling cheated when hidden terms surfaced years later.
"I signed my mortgage in 2019. I thought I understood it. Two years later, I discovered a clause that doubled my monthly payment if interest rates changed. No one explained that to me. I felt betrayed by my own bank."
Through surveys (n=247) and in-depth interviews (n=4), we identified the critical pain points in the mortgage signing process:
Information Overload
Mortgage contracts average 42 pages of dense legal and financial text. 94% of survey respondents admitted they didn't read the entire contract before signing
Asymmetric Knowledge
Banks have teams of lawyers and financial experts. Customers have Google and anxiety. This power imbalance creates mistrust
Delayed Understanding
Customers often don't understand contract implications until years later when terms activate, leading to complaints and damaged relationships
Hidden Complexity
Key terms like "variable interest adjustments" or "early repayment penalties" are buried in legalese, making them easy to miss
The interviews revealed a pattern: customers wanted to understand their contracts, but they were afraid to ask questions because they didn't want to appear stupid or slow down the signing process.
"At the signing appointment, the notary read through the contract so quickly. I had questions, but everyone seemed rushed, and I felt like I was holding things up. So I just signed. I regretted it immediately."
EL PROBLEMA (The Problem)
Market Context
The Spanish mortgage market was recovering from a downturn, with March 2025 seeing 42,831 signed mortgages—a significant increase from the 29,641 in March 2024. This growth created both opportunity and risk: more customers meant more potential complaints if we didn't solve the comprehension problem.
Competitive analysis showed that no Spanish bank offered tools to help customers understand contracts. International benchmarking revealed a few startups (Pave, Better.com) experimenting with AI explanations, but none had solved the trust problem.
**The Core Problem:** How do we help mortgage customers understand 42-page legal contracts without requiring them to be financial experts, while also building trust in both the bank and the AI system?
TEAM & COLLABORATION
Cross-functional team of 5. I led design engineering collaboration with 1 ML engineer, 2 full-stack developers, and legal compliance team. Worked closely with product management to balance business objectives with user needs
MY ROLE
Lead Design Engineer. I owned end-to-end UX design, collaborated on prompt engineering and model selection, designed OCR error handling, implemented WCAG accessibility patterns, and conducted A/B testing with 2,347+ users
CROSS-FUNCTIONAL WORK
Partnered with ML engineer on 23 iterations of prompt design, worked with legal team on verification stamp system, coordinated with engineering on OCR pipeline architecture, aligned with PM on success metrics and rollout strategy
Constraints:
Survey data revealed surprising insights about AI trust:
LA SOLUCIÓN (The Solution)
After 4 weeks of research and iterative design, I led the development of an AI-powered mortgage assistant with a four-layer information architecture that increased trust from 39% to 78%:
Technical Implementation Context:
AI/ML Architecture:
I collaborated closely with our ML engineer to architect the AI system. I led model selection, testing GPT-4 vs Claude with 50 sample contracts, ultimately selecting GPT-4 for superior Spanish language comprehension. I designed a two-stage prompt system:
1. **Extraction prompt**: Parse contract into structured JSON (amounts, dates, clauses, parties)
2. **Explanation prompt**: Generate plain-language summaries with citations
I owned the prompt design end-to-end, iterating through 23 versions based on user testing feedback. Key prompt engineering decisions I established: always cite specific clause numbers, use "you" language for personalization, never give advice beyond contract content, flag unusual terms for human review.
Document Processing Pipeline:
I designed the OCR pipeline architecture with engineering, establishing error handling patterns for failed processing (12% of uploaded documents had quality issues). I crafted user-facing error messages balancing technical accuracy with actionable guidance, reducing user frustration and support tickets.
Accessibility Implementation:
I ensured WCAG 2.1 AA compliance for Spanish banking regulations, integrating accessibility from the first design mockup:
I established that accessibility wasn't an afterthought but a core requirement. Banking products serve diverse populations including elderly users and users with visual impairments, making accessibility critical for both compliance and equity.
Information Architecture:
**Layer 1: Overview Dashboard** - Key information at a glance: total cost, monthly payment, interest rate, important dates
**Layer 2: Critical Clauses** - The 8-12 clauses that matter most, highlighted and explained
**Layer 3: Smart Search** - Ask the AI questions about your specific contract
**Layer 4: Full Contract** - Complete original text with inline explanations
Card sorting with 23 participants validated this structure, with 91% agreement that the layered approach reduced overwhelm compared to linear reading.
Key Design Decisions:
Decision 1: Chat Interface for Questions
Users have unique questions about their specific contracts. "Chat with Your Contract" feature where users can ask anything ("What happens if I want to pay off early?" or "Can my interest rate change?") and get contract-specific answers with citations.
In usability testing with 18 participants, 94% used the chat feature, averaging 4.2 questions per session.
Making AI Conversational Yet Trustworthy
The chat interface used natural language but always cited specific contract clauses. This balance made AI feel helpful, not flippant.
Decision 2: Color-Coded Highlighting in Original Contract
"Smart Highlighting" that automatically color-codes clauses by importance and type: interest rates in blue, fees in amber, penalties in red, protections in green.
Eye-tracking studies showed users scanned color-highlighted contracts 3.4x faster and correctly identified critical clauses 89% of the time, compared to 31% with unhighlighted contracts.
Decision 3: Diagnostic Dashboard with "Questions to Ask Your Bank"
"Contract Diagnostics" that analyzes the contract and generates personalized questions like "Your early repayment penalty is 2%. Is this negotiable?" or "Your interest rate can adjust annually. What's the cap on increases?"
Beta users who used the diagnostic dashboard negotiated better terms 34% more often than control group (based on self-reported data with 67 participants).
Empowering Customers to Negotiate
The diagnostic feature transformed passive document review into active negotiation. Users reported feeling "prepared" and "confident" in bank meetings.
Decision 4: Legal Team Verification Stamps
Every AI-generated explanation shows a verification stamp: "Reviewed by Legal Team on [date]." This doesn't mean lawyers wrote the explanation, but they confirmed accuracy.
A/B testing showed designs with verification stamps had 78% trust scores vs 52% without. The stamp was the single highest-impact trust element.
TECHNICAL IMPLEMENTATION DEPTH
AI Integration & Prompt Engineering:
I led the AI architecture design, establishing a two-stage prompt system that balanced accuracy with comprehension:
OCR Pipeline Architecture:
I designed the document processing pipeline handling variable-quality scanned PDFs:
Mobile Design Patterns:
I established mobile-first design patterns for React Native implementation:
A/B Testing Infrastructure:
I designed and executed A/B tests with 2,347 users across 4 major features:
MOCKUPS & PROCESO (Mockups & Process)
User Journey Mapping
We mapped the entire mortgage process from search to signing:
User Journey
Search for a place
Explore portals and choose your favorite home
Go to the bank
Request information, simulate, and compare
Obtain pre-acceptance
Receive preliminary offers with incomplete information
Submit documentation
Gather payroll, contracts, employment history
Obtain FEIN
Receive official contract with additional terms
Sign mortgage
Go to notary and sign
The key insight: the moment customers receive the FEIN (official mortgage offer), they transition from "excited" to "overwhelmed." This is the critical intervention point.
Design Evolution
Round 1: Simple Summarization
First prototype: upload contract, get a summary. Users liked the simplicity but didn't trust it. "How do I know this is right?" was the dominant concern.
Round 2: Side-by-Side Comparison
Added split view: AI summary on left, original text on right, with highlighted connections. Trust improved to 58%, but users found it "cluttered" and "hard to scan."
Round 3: Progressive Disclosure with Verification
Final design: Start with dashboard, allow drilling into details, always show "See original clause" links. Added verification stamps from legal team. Trust jumped to 78%.
Visual Design
The visual language needed to feel trustworthy (this is sensitive financial information) while being approachable (not intimidating like traditional bank apps).
Design decisions:
Key Features:
Contract Diagnostics
Overview with key information: interest rates, amounts, dates, plus personalized questions to ask the bank based on your specific contract
Smart Highlighting
Automatically highlight important clauses, numbers, and dates with color-coding for quick identification of critical terms
Chat with Your Contract
Ask the AI anything about your contract and get answers with citations to specific clauses and page numbers
Mortgage Translator
Select any term or paragraph and get a simple, plain-language explanation that removes the legal jargon
CONCLUSIÓN (Conclusion)
I measured success through three lenses: adoption, comprehension, and trust across a 3-month beta period.
Adoption Metrics (Achieved in 2 months from concept to beta):
Comprehension Metrics (n=89 usability tests vs control group):
Trust Metrics (n=234 post-use survey):
"This is what banking in the 21st century should feel like. I actually understood my mortgage before signing. For the first time, I felt like the bank was on my side, not trying to trick me."
Business Results:
Complaint Reduction (6 months post-signing vs historical baseline):
NPS Improvement (vs control group and historical baseline):
Cost Savings (annualized projections):
One unexpected outcome: customer service calls decreased by 23% during the beta period. Users who used the tool had fewer questions because they already understood their contracts.
APRENDIZAJES (Learnings)
What Worked Well
**Trust-first design:** I established that starting with the trust problem, not the technology, was critical. If we had led with "AI-powered tool," we would have failed. By focusing on transparency, verification, and side-by-side comparison, I built trust incrementally—resulting in 78% trust vs 39% baseline.
**Legal partnership:** Rather than treating the legal team as a blocker, I made them a strategic partner from day one. I facilitated their involvement in the design process, and their verification stamps became the highest-impact trust element (50% improvement in A/B tests).
**Real contract testing:** I established testing protocols with users' actual contracts (not mock data), revealing edge cases we never would have found otherwise. One user's contract had a clause written in Catalan that broke our OCR pipeline—discovering this in beta saved potential production incidents.
**Cross-functional prompt design:** I brought the ML engineer into user testing sessions, enabling them to see firsthand how users reacted to AI output. This informed 23 iterations of prompt engineering. I established that collaboration between design and ML produced superior results than either discipline working in isolation.
What I Would Do Differently
**Earlier technical validation:** We spent weeks designing features that turned out to be computationally expensive. The "instant translation" feature we demoed in prototypes took 3-4 seconds in reality, breaking the interaction model.
**More diverse research participants:** Beta users skewed toward younger, more educated, tech-savvy customers. We didn't test enough with older users or those less comfortable with technology.
**Clearer AI limitations:** We marked AI-generated content with disclaimers, but we could have been more explicit about when AI might struggle (e.g., unusual contract formats, multiple languages, handwritten annotations).
Key Lessons
**Lesson 1: Trust is multidimensional.** Users didn't just need to trust the AI—they needed to trust the bank, the process, and themselves to understand complex information. Design for all trust vectors simultaneously.
**Lesson 2: AI explanations need explanations.** It's not enough to show AI output. You must explain how the AI works, what it can and can't do, and who verified it. Transparency about limitations builds trust more than hiding them.
**Lesson 3: Comprehension ≠ Confidence.** Users could objectively understand their contracts better (proven by quiz scores), but subjective confidence required additional design elements like the verification stamps and "questions to ask" feature.
**Lesson 4: Design for advocacy, not just understanding.** The diagnostic feature succeeded because it didn't just help users understand—it helped them act. Empowering users to negotiate transformed them from passive recipients to active participants.
**Lesson 5: Generative AI works best with guardrails.** We experimented with fully open-ended AI responses but found they were too unpredictable. Adding constraints (always cite sources, use simple language, never give advice beyond the contract) made AI output reliable and trustworthy.
**Future Direction:** The tool is now being rolled out nationally across Spain, with plans to expand to other financial products (personal loans, credit cards, insurance policies). We're also exploring a B2B version for real estate agencies and mortgage brokers.