# Zevian — Performance investigation case study > From AI scoring to evidence-led investigation, with findings, supporting signals and manager control. - Reza A. Work About Contact Direction - Evidence - Context - Earlier work - Foundations # Helping sales managers understand performance changes I moved the product from AI scoring toward evidence-led investigation, keeping the manager in control of the decision. My role Solo Founder Ownership Discovery, design, AI-assisted build Evidence 10+ conversations, React prototype Links Current prototype ↗ Previous prototype ↗ Previous case study ↗ Finding, evidence, and manager context ## Outcome I designed and built the earlier scoring product with payment and email integrations, reaching two paying customers. Manager feedback led to a new investigation workflow. That React prototype remains in development; its outcomes are not yet validated. ## Problem A performance change needs explanation. A score alone does not provide the context needed to act. ## My responsibility Challenged the scoring model, set a new direction, and designed a workflow from finding to human decision. ## Product direction Manager feedback exposed two gaps: dashboard numbers lacked context, and AI scores were hard to trust. I removed the dashboard from the current workflow and brought findings, source evidence and manager context together. ## Assumption I changed The early idea used employee reports and KPIs to produce AI scores. The new direction gives managers findings, evidence, and room to add context. Early assumption Set KPIs Set Employee reports AI Scores Coaching In discovery conversations, one manager said they did not trust AI scores. Another described investigating why those patterns occurred as the bigger challenge. New direction Collects data Add context Manager decides Findings The system prepares the evidence. The manager judges what it means. ### AI-assisted research Fireflies captured the interviews, Google Drive held the transcripts, and ChatGPT helped group feedback. I reviewed the themes against the interviews before changing the product. Research method and design changes I organised the transcripts in a dedicated Google Drive folder connected to ChatGPT. AI supported synthesis; I checked the grouped feedback against the interviews and decided what to change. Unclear dashboard numbers led me to remove the dashboard and narrow the current workflow. Investigating patterns was already part of one manager's job and a bigger problem for another. Distrust of AI decisions informed visible evidence, missing context and manager control. These are qualitative findings and design decisions. They do not establish measured validation of the current workflow. ## Evidence and control A finding should help a manager investigate. The interface separates what changed, possible reasons, and the decision. ### Evidence Show the change and supporting signals before suggesting a possible cause. ### Uncertainty Label missing context so a possible reason is not mistaken for an established fact. ### Human context and judgment Let the manager add context and choose what happens next. ### Traceability Store the action alongside the evidence and context used to make it. ## Edge cases A finding changes when a manager or employee adds context. The record keeps the evidence, context and decision together. Default state Employee context and decision made Both contexts added and summary updated Earlier version · Evaluation and patterns ## Earlier prototype: employee scoring These screens show the scoring product before I changed direction. They are separate from the current investigation prototype. Caption: Ask for one decision at a time Managers had to define a goal and its scoring rules in one form. I split setup into two steps and added presets, so they could start with the goal instead of a blank form. The trade-off: one more step, with less to decide on each screen. Caption: Review the evidence behind a score The report brought targets, written evidence and coaching notes together, so the manager could review what the score was based on. This screen belongs to the earlier evaluation product. Caption: Give the dashboard a clear starting point The earlier layout repeated status and metrics across several cards. I gave the team score the most weight and brought reporting status into one bar. Detail moved below the summary or a click away. This improved the hierarchy, but it did not answer why performance changed. That became the focus of the current prototype. Caption: Keep AI answers connected to their sources The manager chose the people and dates before asking a question. The answer linked back to the reports it used, so the manager could check the source material. The current prototype takes this further with evidence, missing context and a manager decision. 1 / 4 KPI setup Report Dashboard Ask AI Earlier RAG decisions In the earlier scoring product, I designed the retrieval-augmented generation (RAG) assistant around a selected set of reports, people and dates. Answers linked to source reports, and the assistant was set up to decline questions outside the retrieved evidence. Choose where AI judgment belongs Numeric criteria used targets and deterministic checks. Judgment criteria used weighted rules and report quotes to support the score. This kept numeric checks separate from AI interpretation. Keep a human review path Managers could override a score with a reason while preserving the original score. Employees could see the criteria and score a draft before submitting it. These choices belong to the earlier scoring product. They describe the approach, not measured safety or validation of the current investigation prototype. Explore the earlier case study ↗ ## Design foundations I defined tokens and interaction states in Figma, then used Claude Code and Codex to build and revise the React investigation prototype. Primitive and semantic tokens ### AI-assisted design and build I used ChatGPT with Figma to create tokens, then used Claude Code and Codex to build and revise the React prototype. I set the product direction and interaction rules, and iterated on manager feedback. AI-assisted working method I used Claude and Codex for requirements refinement, architecture and roadmap planning, feature KPI definition, backlog generation, prototypes and code. I chose between them according to the task. Figma tokens, components and interaction states gave the build a design reference. The current prototype was built with Claude Code and Codex, then revised as manager feedback changed the workflow. The investigation prototype remains in development. This describes my working method; it does not establish measured speed gains or product validation. Notion informed the drawer pattern and Attio informed the visual language. I am exploring a Gmail-style list to make findings easier to scan. ## What changed and what comes next Manager feedback changed the problem I was solving: a score was not enough to explain a performance change. The current prototype brings findings, evidence and context together. Next, I need to test whether managers can use it to choose a useful next action. The earlier product’s paying customers do not validate this new workflow. Next project · Jayga → --- [View the visual case study](/zevian)