Create a powerpoint presentation vs Vibecode a presentation

About me

  • 10+ yrs in data
  • Have tried all of the roles a data career can offer
  • Power BI my first love still burning
  • Currently consultant / solopreneur

Why Power BI why now?

Semantic Model Tables Measures Relations Context
  • LLMs have all the capabilities in the world — but they lack context
  • For decades, Power BI has been hardcoding context into its semantic model
  • Power BI is in a unique position — no other BI tool has anything like it
  • Looker comes closest with LookML — but it's different
  • Tableau is struggling to adjust to modern realities (Metric Pulse, etc.)

Copilot / Fabric Data Agents

Schema ? DAX Columns Guess On the fly
  • They guess how to calculate metrics — sees schema, estimates. Works for basic metrics; hallucinates on complex ones (retention, cohort, conversion)
  • Generic AI applied to a specific use case — there will always be room for improvement
  • Steep pricing (~$70/user) — generates DAX on the fly, re-guessing metrics you've already defined
  • No visibility — can't see how users interact with Copilot or Fabric Data Agents

Their weaknesses are your opportunity

Your Opportunity No guess Use what's there Specialized Narrow scope Affordable Easy to beat Visibility Usage metrics
  • Guesswork → No guess. Orgs spent years building semantic models — tribal knowledge is already there. Pick the metric and run its calculation.
  • Generic AI → Specialized. Semantic model understanding — nail down the scope of the agent.
  • Steep pricing → Affordable. $70/user is easy to beat. A 200-user org could fit the whole solution in a few hundred dollars.
  • No visibility → Usage metrics. Treat data as a product. See what users ask, which metrics and dimensions they use — super valuable for future development. Power BI dashboards make this hard.

Points of suffering

  • Discovery engine. I was always the go-to when someone needed a niche metric — "have we ever calculated X?" I'd search my head, find the dashboard where I experimented, redirect them. Could never automate this.
  • Breakdowns not in the report. User: "Can we slice this by that dimension?" It exists in the model but not in the UI. Edit, add chart, save — annoying context switching. Agents can automate this.
  • Documentation. Copying DAX, Power Query, Fabric notebooks into ChatGPT/Claude, asking for docs by template. Dumb process — should be totally automatable. This solution can address it.
Discover Breakdown Document

With the rise of LLMs, it's finally time to end the suffering.

Demo time

Try the demo →

How this works behind the scenes

Extract metadata Embed vector DB LLM Azure DAX + breakdowns
  • Vectorize the semantic model. Extract metadata, embed into a vector DB. Each metric gets its list of breakdown dimensions + definition index. LLM maps user input to existing metrics — no guessing.
  • LLM on Azure Fabric. Fully on-premise — nobody wants to expose data to third parties or log into yet another tool.
  • Flow: User question → LLM finds closest metric → executes DAX → returns value + all possible breakdowns. Solves discovery and breakdown.
  • Documentation: Scrape all metrics and sync to Notion or other documentation systems. Solves the documentation suffering.

Where this can go moving forward

Core Access Analytics Excel Channels
  • Metric access management. Turn on/off metric exposure to the LLM. Highlight which dataset has the core revenue metric — every report has "revenue," but which one is canonical?
  • Product analytics. See what users ask. Logs of LLM questions → build data product roadmap. Users asking for metrics by dimensions that don't exist? Prioritize dataset enrichments.
  • Direct Excel connection. Copy Power Query, paste into Excel — live data, refreshed. No re-export. Semantic model at the core.
  • Agentic flow. Today: single-shot Q&A. Tomorrow: "I need profitability analysis of Baltic states for my board meeting — what drove it last quarter?" Agent plans: query metric, break down by everything, find outliers, drill deeper. Real multi-step reasoning.

Enter MCP

Semantic Model MCP Claude Copilot Others
  • Recent client: management into Claude desktop, wanted to query Power BI from there.
  • Current Power BI MCP edits Power BI files — but doesn't properly query the semantic model.
  • This solution can be packaged as an MCP — semantic model querying as a tool.
  • Copilot Studio lets you build agents with MCP. No need to develop an agent — use Claude, Copilot, or others. Narrow the scope: just the MCP tool.

Thank you for your attention to this matter!