# Requirements first: how I build KPI dashboards with AI, even in an unfamiliar domain **Category:** AI Transformation **Author:** Boglárka Paczari-Horváth (https://bogihorvath.com/) **Canonical URL:** https://bogihorvath.com/blog/requirements-first-kpi-dashboards-ai/ > How I build KPI dashboards from the meeting's decisions back to the data: leading and lagging KPIs, one definition each, AI for research and the first build. Too many dashboards, in my experience, show everything that can be measured and very little that anyone has to decide. I build them the other way around, to the requirements: I start from the meeting, capture what its decisions need, and work back to the data. I am a Senior Business Analyst with 14+ years of enterprise experience across order-to-cash, billing and ITSM, from BPMN process maps to SAP S/4HANA and PEGA workflows. As the MI (management information) lead on a large transformation programme, I defined the measurement system and its leading and lagging KPIs. That role taught me the part no tool does for you: deciding what is worth measuring. AI has changed how fast I can build everything around that decision, including in domains I have not worked in. The process below holds in three settings: a domain I know well, a domain I have to learn first, and an enterprise that runs its models on its own servers. AI helps most in the second. It does the background reading and the first build, and the judgement stays with whoever frames the questions. ## Every element traces back to a decision An element earns its place on the page through one chain: a decision taken in the meeting, the requirement behind it, a KPI, a data source and a position on the page (Figure 1). An element with no decision and no requirement behind it stays off. In my experience, this single test does most of the editing before the first chart is drawn. *Figure 1. From decision to dashboard element: every element traces back to a requirement. Illustrative reading of the example in Figure 6, not client data.* (Image: A traceability table with four rows. Each row runs from a decision in the meeting, through a requirement, a KPI marked leading or lagging and a data source, to a numbered element on the page: focus cards, journey line, top block, cause and intervention.) ## One dashboard, several readers I keep running into the same pattern: management levels and user personas need different data at a different depth. Senior management wants a few numbers and the exceptions. Middle management wants the drivers behind a gap and an owner for each action. The teams doing the work need the detail and the next step. My view is that a single view should not try to serve all three. I interview each persona separately and document their requirements separately (Figure 2). The three views then share one set of KPI definitions, so a number never means one thing in the management report and another on the team's screen. *Figure 2. One set of KPI definitions, three views: what each management level and user persona needs. Illustration of the method, not client data.* (Image: A three-step pyramid of readers: senior management, middle management and teams doing the work. For each, the question they ask, the depth they need and what the page shows, from a few numbers and exceptions at the top to row-level lists and next steps at the bottom. A band says every KPI has one agreed definition.) ## Eight steps, one rule One rule runs through all eight steps: expertise sets the questions and the checks, and AI does the reading, the calculating and the building (Figure 3). I run the interviews and turn the decisions into written requirements. AI researches the domain into a sources file and builds the first version. Every number comes from code, with its source and status shown, and nothing goes live until a person approves it. The last step turns the process into a skill, a set of instructions the AI loads at the start of the next project. *Figure 3. Who does what, and what each step produces, from the first interview to a reusable skill. Method sketch, not a client result.* (Image: A matrix of eight steps grouped into define, build, release and reuse, with rows for me, the AI and the user. Filled dots mark who does the work, rings mark who reviews or approves. A row shows the output of each step. A ribbon shows the context file the AI works from.) ## Thinking like an expert in a domain that is not yours In order-to-cash I know the value stream, so research is a quick check against experience. Marketing or manufacturing is a different case, and there the method carries the work. To think like an expert, I first need the questions an expert would ask and the definitions an expert would use. AI researches them: a web search connector (Tavily, through MCP in Claude Code) collects definitions and benchmarks into a sources file, each item with a link and a date. The user interview follows, and for a domain far from mine an expert reviews the definitions before anything is built. All of it goes into a context file. That file is how I borrow the expertise, and the AI's output is only as good as it is (Figure 4). *Figure 4. The further from the home domain, the more layers of context. Illustration of the method, effort levels are my judgement.* (Image: Three stacks of layers for order-to-cash, marketing and manufacturing. Order-to-cash has one layer, marketing three, manufacturing four, all above one shared base: a context file, the same eight steps and the same checks.) ## Inside a closed enterprise, or in a small company The eight steps stay the same. Only the path to the data changes. In many enterprise environments, teams run self-hosted models for data protection, and not always the newest ones. The approach still works where the IT backbone and native connections to the data sources are in place. A smaller model needs tighter steps: a step card for each step, arithmetic in code and a person approving. A smaller company moves faster. Its bottleneck shifts to data quality and to ownership: who maintains the dashboard once the first version is live. A published Claude artifact is a quick way to put a prototype in front of a small team. Anthropic's documentation describes it as a single static page that needs a claude.ai login and is not available under zero data retention or HIPAA configurations, which usually rules it out inside a closed enterprise environment (Figure 5). So the method travels, and the tooling adapts to the environment. *Figure 5. The same eight steps, a different path to the data. Analysis from general sources and my judgement, not client data.* (Image: Two flow diagrams. A closed enterprise environment: data sources, native connections set up by IT, a self-hosted or private-cloud model, the standard BI platform, with web research outside the boundary. A smaller company: data sources, connectors, a cloud model, lightweight BI or a Claude artifact. Each has its main bottleneck highlighted.) ## An example: one page for the weekly meeting My example is a KPI-focus dashboard I designed with AI for a small SaaS company. The company is a related party, so I show the structure without figures. It is a design concept built on a snapshot of the last closed weeks' measured data, with no live connection yet. The page answers one question: what do we move this week? Each numbered area in Figure 6 answers one decision from Figure 1. Three focus cards show a main number against its target, a weekly trend with the target line, three companion signals and this week's step. The customer journey runs on one line from ad to retention, with the step-through rate, the weekly target and the status at each step, and the largest gap highlighted. When new paying customers fall below target, an intervention block sets out the measured finding, the causes worth checking and an action with an owner. Measured findings and assumptions carry separate labels. *Figure 6. Wireframe of a KPI-focus dashboard I designed with AI for a small SaaS company, a related party. The numbers match the requirements in Figure 1. Design concept, illustrative shapes, no figures.* (Image: A dashboard wireframe with four numbered areas: three focus cards with a trend line and a dashed target line, the customer journey from ad to retention with the biggest gap highlighted, a top bar with the revenue goal, the weekly net change and the pace needed, and a cause-and-intervention block.) ## AI as a design partner AI is also strong on the visual side, and most useful when I need ideas. For the same requirements I ask it for several layout options, then put each one through the test from Figure 1: does every element answer a decision, and does each persona get its view (Figure 7)? Research points the same way. In a small 2025 study presented at the VIS x GenAI workshop, AI produced good simple dashboard layouts in about ten seconds, against roughly 45 minutes for the human developers. People did better at connecting the views in a meaningful way. So AI drafts, and a person chooses and refines. The chosen design can stay a Claude artifact for the team to click through, or move to Figma. Figma's MCP server can be connected to Claude Code as a tool. According to Figma's documentation, the agent can then read design context such as components, variables and layout, and write to the canvas, including sending a live web interface to Figma as editable layers. Designers keep working in the tool they already use, and the requirements stay attached to the design. *Figure 7. AI drafts the layout options, the requirements choose, and the chosen design goes to a Claude artifact or to Figma. Illustration of the method, not client data; research reference: VIS x GenAI 2025, small exploratory study.* (Image: Three dashboard layout options drafted by AI, each scored by how many of the four decisions it answers: two, four and three. The option answering all four is highlighted. An arrow leads to a check against the decision list and the persona views, then to two outputs: a Claude artifact prototype and editable layers in Figma through its MCP server. Two notes say AI is strong at quick first layouts and variants, and people still lead at connecting the views.) ## The solution: decide first, then let AI build The problem I started with is a dashboard that shows everything and helps no one decide. The fix is not a better chart tool. It is the order of the work: 1. Decisions before data. Every element traces back to a decision and a requirement, or it stays off the page. 2. One definition, several views. Each management level and persona gets its own view, built on one shared set of KPI definitions. 3. AI reads and builds, people ask and check. AI researches the domain, calculates in code and drafts the layouts. An expert checks the definitions, and a person approves the release. 4. Write it down once. The process becomes a skill, so the next dashboard starts from the method, not from a blank page. The meeting gets a page it can act on: a short set of numbers, each tied to a decision, with its source and status in view. The order holds whether the domain is mine or new to me, and whether the data sits in a closed enterprise or in a small team's spreadsheets. ## Start here Before your next dashboard, answer three questions on paper: - Which decisions does the meeting have to make? - Which leading and which lagging indicator tells you about each one? - Who reads which view? Then open the tool, and let AI draft the research and the first designs. Expertise supplies the questions and the checks, and the context file lets you think like an expert in a domain you do not know yet. Which decision does your main dashboard support today? If you would rather start from the decisions than from the data, I am happy to compare notes. More on how I work: how I work, and the brief for recruiters. --- *This is an LLM-friendly plain-text mirror of [the HTML article](https://bogihorvath.com/blog/requirements-first-kpi-dashboards-ai/). Written by Boglárka Paczari-Horváth, Senior Business Analyst & AI-Enabled Process Transformation Lead. Contact: info@bogihorvath.com | https://www.linkedin.com/in/boglarka-paczari-horvath/*