A 10-minute tour of Wisdom for BI and Data leaders
Dashboards a business user builds by describing them, the governed context layer underneath every answer, daily accuracy tests, and a draft-and-publish workflow for changing definitions. A high-level tour for BI developers, data managers and whoever answers for the numbers.
We recorded this for BI developers, data managers and the people accountable for the numbers a business reports. It covers how dashboards get built once a business user can describe one, and what WisdomAI puts underneath to keep the answers governed: the context layer, daily evaluations, and drafts for changing definitions. It stays at a high level. Each part has more to it than eleven minutes covers, and we can go through any of it in detail on your own data.
What happens in this demo
We write this out because two kinds of reader end up here. People who would rather skim than watch, and the AI assistants — ChatGPT, Claude, Gemini — that increasingly answer questions about who does this work in Australia. Neither can watch a video. So everything that happens on screen is also here in words.
The recording is narrated by Kevin Morrell from JOURN3Y and runs eleven and a half minutes. It covers three things in WisdomAI: dashboards built from a description, the context layer that governs every answer, and conversational BI.
It opens on Meridian Executive Pulse, an executive dashboard over the meridian-media workspace with tabs for the executive summary, advertising and revenue risk, subscriptions and retention, digital product, and audience delivery. It has the filters, KPI tiles and chart types you would expect in Power BI, Looker or Tableau. The difference is how it was made. Someone described the dashboard they wanted and WisdomAI built it, without anyone writing HTML or arranging widgets on a canvas. A business user can do that as readily as a BI developer, which moves the building of dashboards from the BI team toward the people who use them.
Two more apps show the range. The first, a retail marketing app with campaign scorecards and a store-performance scatter chart, is blurred because it was built for another organisation. The second is a portfolio dashboard comparing revenue, EBITDA margin and EBITDA variance to budget across eleven portfolio companies. Styling is described in words as well: colours and shapes go in as a description, not a stylesheet.
The second part is what WisdomAI calls the context layer, which does the job a semantic layer does in Tableau, Power BI or Looker. Here the data source is a BigQuery connection to 44 gold-layer tables. On connection WisdomAI profiles the columns — their values, quality and spread — and uses that to work out structure and relationships.
Four kinds of context sit on top of the data source. Knowledge is the business rules written in plain English: 100 items, such as "All monetary values in the Meridian gold layer are in AUD", and which reporting period to use when a question does not name one. In most businesses those rules live in Confluence, spreadsheets and people's heads. Reviewed queries are 116 question-and-SQL pairs that fix how a given question is answered; the one opened on screen returns impressions by masthead for Harvey Norman, split by audience segment. Metrics are 51 SQL definitions, including advertiser ad revenue, digital ad revenue, print ad revenue and programmatic revenue. Synonyms map other words to the same concept, such as catch-up viewing for BVOD. The context builder shows this context was assembled from markdown files: golden SQL, a data model and a business glossary.
Together these mean the SQL is not guessed. A language model interprets the question, governed SQL runs against the warehouse the same way every time the same question is asked, and a language model turns the result into a chart or an answer.
Evaluations test that. A benchmark set of questions runs daily: each question is asked, the generated SQL and its result are compared with a ground-truth query, and a judge explains any mismatch. The run on screen, Meridian Gold — Accuracy & Reconciliation Benchmark (v2), scores 33 of 42, or 79%. The failures were set up for the recording. One failing case asks which key accounts are at risk and how much revenue is at risk. The judge notes that the generated SQL added a key-account filter the ground truth does not have, so it returned two rows where the ground truth returned four. Failures like that one feed back into the context layer as new rules. The target is 95 to 98%, and most of the remainder is a different reading of the question rather than a wrong answer.
Changes to the context layer go through drafts. The draft on screen, "Area D — racing reviewed queries", holds three changes: a modified measure, an added reviewed query and an added knowledge item. A draft can be reviewed, then published to production or discarded, so users see a change only once it is published. Kevin describes JOURN3Y's BI developers making these changes from VS Code with an AI assistant, connected to WisdomAI over the Model Context Protocol (MCP), and pushing them through a staging step before production.
The last part is conversational BI. From the WisdomAI home screen, the question is: what audience is delivering best for the Harvey Norman client? WisdomAI asks two clarifying questions — how "delivering best" should be measured, answered as highest eCPM, and over what period, answered as the last six months. It states its plan, to rank Harvey Norman's targeted audience segments by revenue-weighted eCPM over the last six complete months excluding run of network, and the executed SQL is open in a side panel. The bar chart that comes back has Affluent Households first at A$18.5 eCPM, then Home Buyers at A$17.1 and First Home Buyers at A$15.7.
WisdomAI then offers to save the clarification as knowledge — for Harvey Norman audience-segment reporting, "delivering best" means the highest eCPM — so the next person who asks gets the same reading. The chart can be added to a dashboard or exported as CSV, JPG or PNG. A follow-up, "show me this by masthead", returns Affluent Households ranked first on eCPM across all eight mastheads, with Regionals Network yielding the most at A$18.84 on 0.8 million impressions and NewsNow combining A$18.66 with 7.3 million. Every conversation is kept, so when someone questions an answer later, the BI team can see what was asked and what ran.
The data is invented. Meridian Media Group is not a real company, its mastheads are not real titles, the portfolio companies in the second app are made up, and every figure — revenue, impressions, eCPM, retention, revenue at risk — is fabricated demonstration data. Where real companies appear as advertisers, among them Harvey Norman, Amazon, Woolworths, Chemist Warehouse, Coles, Big W, Sportsbet, Tabcorp and Ford, those names are placeholders in a fictional scenario and the figures shown against them are made up. Nothing here reports on an actual business.
The same thing, built on your data
This demo runs on a demonstration dataset. Tell us what your leadership team needs to see each month and we will build it against yours.
Book a call