For most of the past decade, the report was the product. Executives opened a dashboard, and everything that made the numbers understandable lived on the page: the title, the footnote, the layout, and the analyst who built it and could explain it. The semantic model, the layer that holds the business definitions, relationships and calculations behind every report, was plumbing that only the BI team ever saw.
Generative AI changes that. When a leader asks an assistant such as Copilot in Power BI which region missed plan last quarter, the answer is built from the model itself: its tables, columns, measures and relationships, and whatever descriptions someone took the trouble to write. The model has become the interface, and most models were never designed to be read by anyone outside the team that built them.
What the assistant actually reads
An assistant answering a business question works from the model’s metadata: the name of every table, column and measure, how the tables relate, which fields are hidden, how values are formatted, and the descriptions and synonyms attached to each. That is the whole of its knowledge about your business. Where the model is silent, the assistant has nothing to go on and, worse, may fill the gap with a guess.
It does not read the slide explaining that revenue excludes intercompany sales, the email that settled what counts as an active customer, or the analyst’s memory of why one region appears twice. Context that used to travel with the report, or with the person presenting it, now has to live in the model. If a rule is not written there, the assistant will not apply it, however confident its answer sounds.
Names and descriptions are the new labels
In a report, a cryptic column name never mattered, because nobody saw it; the visual carried a friendly title. In a model read by AI, the name is the label. “Net revenue” tells the assistant what it is looking at. “Amt_NR_Adj” tells it nothing, and “Sales” may mean gross, net or booked, depending on who built the table.
Write descriptions for a reader with no context, because that is exactly what an assistant is. A useful description says what the field means, what it includes and excludes, and when to use it instead of its near neighbours. Short, plain sentences are enough. Then add the synonyms people actually use, such as revenue, sales and turnover, so that a question asked in everyday language lands on the right field.
Relationships decide which answers are possible
Relationships, the links between tables, determine which questions the model can answer and how. A person building a report knows that orders can be analyzed by order date or by ship date, and picks the right one. An assistant faced with two possible paths between the same tables has to choose, and it will not always choose the way finance would.
Keep one obvious path between any two tables. Where a date plays more than one role, give each role its own clearly named field. Remove relationships that exist only to keep an old report working. Complexity that a skilled report author works around every day becomes a trap for an assistant that cannot see the workaround. A simple structure is easier for people to trust and far easier for an assistant to reason over.
Certified measures are the answers
Measures, the calculations such as net revenue, gross margin or days sales outstanding, are where answers come from. Over the years most models collect several versions of the same idea: a margin for the sales team, another for finance, a third left behind by a project nobody remembers. A person picks the one their report uses. An assistant picks whichever seems to match the question, and the CFO hears a number that does not match the board pack.
The remedy is a small set of certified measures: one definition per business concept, owned by the function accountable for it, described in plain language and visible to everyone. Helper calculations are hidden. Duplicates are retired, not just renamed. When the same certified measure feeds the executive scorecard and the assistant’s answer, the two cannot disagree, and the path from source system to answer can be shown on request. Certification also sends a signal to people: it marks the version leadership has agreed to use.
Own the model like a product
If the model is the interface, someone has to own it the way a product manager owns an application. In practice, ownership rests on a few clear roles:
- A business owner for each domain, usually in finance, sales or operations, who decides what each certified measure means and approves every change to it.
- A model steward in the analytics team, who keeps names, descriptions, synonyms and relationships current and treats every rename as a change that can alter answers.
- The people who use the answers, leaders and analysts alike, who report questions the assistant got wrong so that gaps are fixed in the model rather than worked around in individual reports.
Ownership is what keeps the interface stable. A column renamed late on a Friday used to break one report; now it can quietly change answers across the organization. Changes to the model deserve the same review, testing and release discipline as changes to any system customers depend on. Publish the definitions where people can read them too, so that the assistant and the finance team are visibly working from the same glossary.
Testing is where the effort shows. Keep a short list of real questions with known answers and run it whenever the model changes. In our experience, the largest gains in answer quality come from work on the model, not from a newer AI.
The bottom line
Reports are not going away; leaders still need a page to monitor the business. But the most important design work in analytics has moved down a layer. Organizations that treat the semantic model as a product, with clear names, honest descriptions, a simple structure, a short list of certified measures and an accountable owner, will get AI answers they can act on. The others will get fluent answers to questions their model was never able to answer. Budgets should follow the shift: less money for new reports, more for the model that every report and every assistant reads.
The same work pays off twice: a model that explains itself to an assistant also makes every Power BI report built on it quicker to build and easier to trust.