Turning fragmented operational data into story-driven decisions

01. Overview

OLIVAI automated quality control across a network of managed sites. In doing so, it generated 10× more data than the existing dashboard was built to handle. Critical operational signals were buried in noise, managers were slow to act, and the CEO had a 90-day deadline to fix it with AI.

My job was to figure out what "fix it with AI" actually meant for real users in real environments, and to ship something that worked within every constraint we had.

02. The Problem

OLIVAI's quality control automation was working. The problem was everything downstream of it. As the system scaled, it produced exponentially more data, but the existing dashboard wasn't built for that volume. There was no prioritization, no narrative structure, and no clear path from signal to action.

Managers were spending more time trying to interpret the data than acting on it. They'd open the dashboard, scan a wall of metrics, and leave without a clear next step. The tool was generating information but not enabling decisions.

• Business problem: signal buried in noise. A 10× increase in automated data volume overwhelmed the existing dashboard. Critical signals (staffing issues, quality flags, financial variances) were invisible until it was too late to act.
• User problem: no path from data to decision. Managers operated in high-pressure, time-constrained environments. They needed to move from "something happened" to "here's what I should do" quickly. The current system made that nearly impossible.

03. The Ask

The CEO's vision was clear and ambitious: managers should be able to get answers and direction through conversation alone, no charts, no manual analysis, no dashboards. I facilitated early alignment with leadership to translate that vision into measurable goals:

• Shorten issue-to-resolution time: once a problem was identified, the tool needed to surface clear next actions, not surface-level alerts that sent managers on a separate investigation.
• Establish OLIVAI's standalone product identity: beyond the feature, the CEO wanted OLIVAI to have its own brand and design language, a foundation that could scale across future AI-driven capabilities.

04. Discovery

Before a single wireframe, I needed to understand what was actually feasible and where the real risks were. I kicked off a cross-functional alignment workshop to pressure-test the CEO's vision against engineering realities, user behavior, and timeline constraints.

• Technical constraint: engineering could support roughly 100 predefined question-and-answer scenarios within the 90-day window. The underlying data was inconsistently tagged, making true open-ended querying infeasible. The AI had to be guided, not open-ended.
• Behavioral risk: the same managers who couldn't find signals in the existing dashboard were being asked to switch to a conversational interface. We had no evidence they would, and this became the central bet we needed to validate.
• Environment risk: managers weren't sitting at a desk with time to explore; they were on-site, handling escalations, checking data in 30-second windows. Any solution requiring mental effort before delivering value would be abandoned.

05. Execution Model

A 90-day timeline with this level of ambiguity meant sequential handoffs would kill us. I designed an execution model that let design, research, and engineering move simultaneously, each informing the others without blocking them: directional research set early constraints, engineering validated feasibility in parallel, and iterative testing refined assumptions before we committed to build.

06. Research & Findings

Early research suggested chat could work. Managers' natural behavior mapped to a consistent mental model: trigger → scenario → action. But the real-world environment told a different story.

• Consistent mental model: 91% of managers applied the same trigger → scenario → action pattern, before being shown any structured alternative.
• Chat slowed decisions: in noisy, high-pressure conditions, unstructured chat required cognitive effort before it delivered value, adding friction at exactly the wrong moment.
• Speed matters more than flexibility: managers didn't need the ability to ask anything; they needed the most relevant answer immediately, without having to formulate the question first.

07. Concept Exploration

Research findings put me in conflict with the CEO's vision. Rather than presenting a single recommendation that risked a standoff between design judgment and executive instinct, I designed three concepts, each a different position on the spectrum from vision-led to behavior-led:

• Concept A: Vision-led, chat-first interface. Conversation as the primary entry point. Maximum flexibility, but required users to know what to ask.
• Concept B: Hybrid, chat + structured stories. Chat as the entry point, with pre-built story cards that collapsed into view as users scrolled.
• Concept C: Behavior-led, story-first dashboard. Pre-structured narrative summaries aligned to the user's mental model, no open-ended input required.

08. Validation

I designed the validation experiment to measure the quality of decisions, not whether users liked the interface, observing speed, confidence, and friction across identical scenarios using each concept.

• Unmoderated scenario modeling: same operational scenario, different concepts; measured time-to-identify, time-to-understand, and time-to-choose-a-next-action per concept.
• Q&A quality checks: same questions, same data set, all three concepts; compared accuracy, response speed, and confidence to act.
• Qualitative cross-checks: side-by-side session review, observing hesitation patterns, backtracking behavior, and decision-loop completion.

Key finding: structured stories won on every metric. Faster decisions, less hesitation, higher completion rates: 80% of users oriented to the pre-built story cards immediately, with no training required.

09. The Solution

I evolved the strongest concept into a production design that kept chat alive as an entry point, without forcing users to rely on it. As managers scrolled, the interface collapsed into pre-built story cards that surfaced context, signals, and next actions in one view.

• Chat as persistent entry point: embedded for constant access within the story dashboard.
• Latest Activity snapshot: a real-time status view built to answer "what's happening right now" in under 10 seconds.
• Structured story cards: narrative summaries built around the trigger → scenario → action mental model, pre-structured and role-aware.
• Needs Attention CTAs: benchmarked, role-specific next actions that replaced the open-ended question "what should I do?" with a direct, contextual answer.

10. Design System

The CEO asked for a standalone brand for OLIVAI. With a 90-day timeline and an active legacy platform in production, rebuilding the design system from scratch wasn't an option, and wouldn't have served users anyway. Instead, I extended 4insite's existing foundation with a new pattern layer: components and interaction models specific to AI guidance, confidence cues, and contextual actions.

12. Reflection

The strongest decisions on this project came from three places: embracing real constraints rather than designing around them, reshaping information around how managers actually reason rather than how we assumed they would, and treating AI as a tool for guided confidence-building rather than full automation.

If I were starting over, I'd push for quantified behavioral baselines earlier, and position the design system contribution as a first-class deliverable from week one rather than a parallel workstream.