AI/UX Framework / Financial Services
Designing Trust in AI
A Framework for Human-AI Collaboration in High-Stakes Financial Decisions
Across four enterprise AI deployments in financial services, I kept encountering the same fundamental failure: systems that were technically accurate but experientially untrustworthy. The outputs arrived stripped of the contextual scaffolding that domain experts relied on to make judgment calls. After solving the same trust breakdowns across these engagements, I stopped treating each fix as project-specific and built a reusable framework.
My Role
I developed this framework incrementally across four enterprise deployments at Fortune 500 financial institutions, each one revealing patterns I had missed in the previous one. What began as ad hoc fixes to adoption problems became a formalized design system for human-AI trust.
Responsibilities:
- Pattern documentation and internal design standards authorship
- Cross-team workshops to validate patterns against observed user behavior
- Integration into design review processes at two institutions
- Longitudinal usability research focused specifically on trust metrics
The Problem
We built an AI scoring model with 94% accuracy. Underwriters ignored it 67% of the time.
The problem wasn't accuracy — it was that the outputs arrived without the context experienced underwriters relied on to make judgment calls. An AI could flag elevated risk on a policy; it couldn't say why that mattered given the regional regulatory environment or the broker's history. Without that context, experts treated the AI the way they treat unsolicited advice from someone who doesn't understand their world. They ignored it.
""I didn't trust it at first, and there was never a period where I could ease into it.""
This was a trust calibration problem, not a UI problem — and most technical roadmaps had been built around the opposite assumption.
Process
Pattern Discovery
Framework Evolution
Confidence Disclosure
Override Dignity
Design Strategy
Progressive Disclosure
AI should demonstrate its reasoning, accept disagreement gracefully, and earn autonomy through demonstrated reliability. Six patterns came out of this work:
- Progressive Confidence Disclosure — show the recommendation first, evidence on demand. Most users want the answer before the reasoning, if at all.
- Override Dignity — make disagreeing exactly as easy as accepting, capture why, and show users their overrides improved the model. Disagreement becomes collaboration, not friction.
- Calibrated Uncertainty — map confidence to scales people already read (traffic lights, gauges), not raw percentages with no frame of reference.
- Explainability Layering — headline, then contributing factors, then full analysis, each opt-in. In practice, 90% of users stop at layer one.
- Graceful Degradation — when the AI doesn't have enough data, show what it does know and flag the gap. Partial transparency beats silence.
- Trust Over Time — start the AI as a second opinion, shift toward AI-first as it demonstrates reliability, never remove the ability to override.
Design Artifacts
Calibrated Uncertainty
Explainability Layering
Graceful Degradation
Trust Over Time
Impact & Outcomes
80%
Trust Score
User trust in AI recommendations increased from 34% to 80% across deployments
3×
Override Quality
Documented override reasoning improved model accuracy through feedback loops
40%
Faster Adoption
Progressive trust onboarding reduced AI feature abandonment
6
Published Patterns
Framework adopted as internal design standard at two financial institutions
What I Learned
Trust is a UX metric, not a sentiment
We added a trust score to our quarterly usability surveys almost as an afterthought. It turned out to predict feature adoption more reliably than task completion rates, error frequency, or satisfaction scores. The implication is significant: if users do not trust the system, it does not matter how usable it is.
Override flows are data collection opportunities
Every time a user overrides an AI recommendation and provides a reason, you receive free labeled training data. If the override flow is designed well, users are improving the model every time they disagree with it. This reframes disagreement as collaboration rather than failure.
What I'd Do Differently
I should have started codifying this after the second deployment, not the fourth — the first two were solved in isolation before I recognized the structural pattern. I'd also invest in quantitative trust measurement from day one instead of relying on anecdotal evidence; formal trust surveys earlier would have made the business case far more convincing, sooner.
Read the Full Story
Discovery — what we found before designing anything
After the fourth project, I started cataloging the specific language users reached for when describing their frustration. The same phrases kept surfacing across different companies, different products, different user populations:
""I don't know why it's saying this.""
""I feel like I'm fighting the system every time I disagree with it.""
""92% confident, but confident about what, exactly?""
""When it doesn't know something, it just goes blank. That's worse than nothing.""
""Day one I checked everything it recommended. Now I just click approve without looking.""
Each of these pointed to a specific failure mode in how AI-driven interfaces handle the relationship between machine confidence and human judgment. Six patterns emerged.
The six patterns, in full
1. Progressive Confidence Disclosure. Present the recommendation clearly. Make the supporting evidence available on demand, but never force it upfront. Most users want the answer first and the reasoning second, if at all.
2. Override Dignity. Make overriding the AI exactly as easy as accepting it. Capture the reasoning behind the override. Then close the loop by showing users that their overrides improved the model. This transforms disagreement from friction into collaboration.
3. Calibrated Uncertainty. Map model confidence to visual scales that users already have intuitions about, such as traffic light metaphors, risk gauges, heat intensity. A raw percentage like 92% means nothing without a frame of reference for what constitutes high, medium, or low in that specific domain.
4. Explainability Layering. Three progressive layers: the headline (what the AI recommends), the summary (the three to four factors driving the recommendation), and the deep-dive (full model analysis with weights and comparable benchmarks). Each layer is opt-in. In practice, 90% of users stop at layer one.
5. Graceful Degradation. When the AI lacks sufficient data for a confident assessment, show what it does know, flag the specific gaps, and invite human input to fill them. Partial transparency is vastly better than silence.
6. Trust Over Time. Begin with AI functioning as a second opinion, where the user makes their assessment first and the AI confirms or challenges it. Gradually shift toward AI-first recommendations as the system demonstrates reliability. Never remove the ability to override.
More of what I learned
- Explainability has diminishing returns: There is a threshold beyond which additional explanation erodes trust rather than building it. After the third layer of detail, users in our studies began to trust the system less. Their reasoning was intuitive: if the AI needs this much justification, something must be wrong with its recommendation.
- Trust calibration varies by expertise: Power users with deep domain knowledge wanted fewer guardrails and faster pathways to override. New users needed more scaffolding and slower trust-building ramps. The same six patterns apply to both populations, but the default configurations need to adapt.
Next Case Study
When AI Acts Autonomously
When AI acts autonomously, trust isn't optional. It's the architecture. This concept applies the framework.