Capacity 01
Conversational UI
I design copilot conversations for commercial bankers that start from the task on screen and end in a completed action.
Enterprise banking copilot, live at 150+ financial institutions with 29,000+ active users.
- Context stays with the task. Coaching appears inside the workflow after a redesign removed the option to jump into copilot chat mid task, which had created user friction.
- Uncertainty is shown only when it matters. Confidence appears as a warning when it is low, after a debate with product over the credibility risk of a permanent percentage.
- Every conversation has a way out. Users can edit, reject, or escalate to a person, with 12% overridden, 8% corrected, and 5% escalated.
Impact
- 68% weekly adoption among eligible users
- 4.2 vs 3.8 feature level CSAT against the product overall
- 74% of users return within a week after a poor recommendation
Capacity 02
Agentic Design
Agentic AI for back office teams at financial institutions, built alongside a copilot used at 150+ institutions.
- Autonomy is set per action. Each action sits at one of three levels, suggest, ask for approval, or complete alone, based on the cost of an error.
- Boundaries come before screens. Scope and exclusions are agreed with product owners and business analysts first, so the interface reflects what the agent is allowed to do.
- Handoff to a person is a designed state. Escalation carries context forward, and 5% of copilot sessions end in a handoff.
Impact
- 5% of sessions end in human escalation
- 12% of recommendations edited or rejected before action
- 60% of low confidence warnings lead users to seek a second opinion
Capacity 03
Predictive Design
Recommendations surfaced inside commercial banking workflows at 150+ financial institutions.
- Recommendations arrive at the point of decision. They appear while the banker works on the task, so no separate search is needed.
- The accountable person keeps the final call. Every recommendation can be edited or rejected before action, and 12% are.
- Appropriateness is tracked next to accuracy. Override rate and return after a poor recommendation show whether predictions stay useful when they miss.
Impact
- 20% lower loan processing time over six months
- 12% override rate on recommendations
- 74% of users return within a week after a poor recommendation
Capacity 04
Ambient AI
AI guidance placed inside existing banker workflows, with no separate tool to open.
- Guidance lives inside the screen the banker already uses. Coaching appears in the workflow, which removes the need to open a separate assistant.
- Timing is designed as carefully as content. A redesign removed the option to jump into copilot chat mid task, which had created user friction.
- Quiet by default. Guidance stays out of the way until the task state calls for it.
Impact
- 25% faster task completion for loan officers
- 15% faster for relationship managers
- 10% faster for customer care teams
Capacity 05
Generative AI
AI generated content drafted for commercial bankers and reviewed before it reaches a record.
- Every output is a draft. Users review and edit generated content before submission, and 8% of outputs are corrected.
- Control stays at the point of submission. Nothing is committed until the person approves it.
- Corrections are treated as quality signals. Edit and rejection rates show where generated content needs to improve.
Impact
- 8% of outputs edited before submission
- 12% of recommendations edited or rejected before action
- 4.2 vs 3.8 feature level CSAT against the product overall
Capacity 06
Systems Thinking
Interaction patterns applied across a copilot at 150+ institutions and a design system built from scratch.
- Patterns are built once and reused. A design system built from scratch gave teams shared components and reached 100% adoption.
- Changes are traced across the whole workflow. A decision on one part, such as escalation rules, is checked against workload, cost, and handoff paths before it ships.
- One experience serves every role. Loan officers, relationship managers, and customer care teams use the same copilot and see different task outcomes.
Impact
- 56% workflow time reduction from the design system
- 100% adoption of the design system
- 150+ institutions and 29,000+ active users on one copilot
Capacity 07
Trust & Explainability
Confidence and reasoning shown to commercial bankers who remain accountable for each decision.
- Uncertainty prompts a pause. When confidence is low, 60% of users stop and seek a second opinion, and 40% proceed on their own judgment.
- Reasons are given in plain language. Each recommendation states what it is based on, so the banker can judge it.
- Trust is measured after errors. 74% of users return within a week after a poor recommendation, which shows trust holds through a miss.
Impact
- 60% pause and seek a second opinion on a low confidence warning
- 40% override and proceed
- 74% return within a week after a poor recommendation
Capacity 08
Failure & Recovery
Recovery paths built into a copilot used in daily banking workflows.
- A wrong recommendation has an exit. Users edit or reject it before action, and 12% do.
- A person can take over. Sessions can end in a manual handoff, and 5% do.
- Design failures get redesigned. A mid task chat option that created friction was removed after real usage exposed the problem.
Impact
- 74% of users return within a week after a poor recommendation
- 5% of sessions end in human escalation
- 8% of AI outputs corrected before submission
Capacity 09
Responsible AI
Accountability design for AI recommendations in commercial banking.
- The accountable person stays in the loop. Bankers edit or reject 12% of recommendations, so the final decision remains with the person responsible for it.
- Uncertainty is disclosed at the point of action. Low confidence appears as a warning at the moment of the decision.
- A person can take over. 5% of sessions end in a manual handoff, so accountability stays with people.
Impact
- 12% of recommendations edited or rejected by the user
- 60% of low confidence warnings lead to a second opinion
- 8% of AI outputs corrected before submission
