Three Decisions Where the AI Doesn't Get to Go
| Property | Value |
|---|---|
| 📅 Project | OmniRisk · Financial Workflow Dashboard |
| 🔖 Phase | Design |
| 🧩 Input | Feature set · Three task flows · Two personas · Five-phase journey map |
| 🗂️ Output | Three trust decisions · Hi-fi screens · AI Recommendation modal · Manual Override modal |
The Design phase took the prioritized feature set and the three task flows from Ideate and resolved them into three trust decisions — the load-bearing calls that determine whether OmniRisk earns trust from skeptical operators. Each decision answered a specific tension that emerged from research. Each decision was a hypothesis. Testing was where they got tested.
Every workflow decision — widget-based dashboard, in-table highlighting, automatic audit trail generation, sequential progression — traces to one or more of the three trust decisions. They are documented as Core features in Feature Set; the three decisions here are the design calls those features implement.
The conditions for AI adoption in institutional finance are not soft factors. They are structural: rules trigger the recommendation, the AI narrates the reasoning, the operator authorizes every action, and every action logs to an audit trail. Without those four properties composing, AI gets ignored. With them composing, AI gets used.
Every design decision in this phase resolves to one or more of these properties. The decisions are not stylistic preferences — they are answers to the research finding that all four participants converged on the same trust conditions from different vantage points.
The decision: Put AI confidence score and reasoning inside the AI Recommendation modal — not as columns in the Currency Trade Table.
Why the modal won: The badges in the table are the triage layer — Repatriate, Monitor, or Hold tells the operator which positions need action. These badges are triggered by fund-strategy rules set upstream by committee, against the fund's investment strategy. The badge is a trigger to act, not a recommendation to second-guess. The granular reasoning belongs one layer down: when the operator clicks Execute, the modal opens with confidence, rule attribution, and AI-generated rationale before they commit. The table stays fast for triage; the modal goes deep for the decision the operator is about to authorize.
What the modal contains: Confidence score displayed as a band (low/medium/high) with percentage. AI-generated reasoning paragraph explaining which rule fired and why. Data source attribution. Estimated projected impact. Paired action buttons (Accept AI Recommendation and Manual Override) of equal visibility.
Research connection: The FX trader described wanting to verify AI outputs before acting. The middle-office analyst named full transparency as a condition for trust. The modal serves both without slowing either down at the action layer.
Alternative considered and rejected: Surfacing confidence and reasoning as inline table columns. Rejected because it clutters the triage layer with detail the operator only needs at the decision moment — and because triage speed matters more than triage depth.