In Short: The Model Is the Prompt
Every Microsoft AI experience that answers data questions - Copilot in Power BI, Copilot in Teams via Fabric IQ, Fabric data agents - works the same way underneath: it reads your semantic model's metadata and translates the user's words into queries against it. The model is, in effect, the system prompt. If its names are cryptic, its relationships wrong or its measures undocumented, no amount of AI capability compensates.
The good news: preparing a model for AI is ordinary modelling discipline, applied with new urgency. Everything below improves human self-service too.
Step 1: Name Things the Way the Business Speaks
AI resolution is a matching problem between the user's vocabulary and your model's metadata. Close the gap from the model's side.
- Rename tables to business entities: Customers, Orders, Products, not DimCust or FactOrderLine
- Rename measures to what the business calls them: Adjusted Revenue, not Rev_adj_v3
- Spell out abbreviations that are not universal in your organisation
- Keep one name per concept: if two measures both claim to be margin, decide which one is real
A useful test: read the field list to someone from finance. Everything they cannot immediately interpret is something Copilot will also struggle to match.
Step 2: Write Descriptions That Carry Meaning
Descriptions on tables, columns and measures are consumed directly by AI features when resolving questions and generating answers. They are no longer optional documentation; they are behaviour.
A good measure description answers three things in a sentence or two: what it means, how it is calculated, and any caveat a correct answer depends on. "Gross margin percentage: gross profit divided by revenue, excluding intercompany transactions" tells both a human and an agent everything needed to use the measure correctly.
Copilot in Desktop can draft descriptions for you. Review them: it describes the formula it sees, not the business intent behind it, and intent is the part worth writing down.
Step 3: Get the Structure Right
Natural-language querying is unforgiving of structural shortcuts.
- Keep a proper star schema: facts with additive measures, conformed dimensions, single-direction relationships by default, per our star schema in practice guide
- Verify relationship cardinality; a mistaken many-to-many silently multiplies numbers in AI answers exactly as it does in visuals
- Prefer explicit measures over implicit aggregations, so the aggregation logic is yours rather than guessed
- Hide what should not be selectable: keys, sort columns, staging fields, deprecated objects. Hidden objects are invisible to Copilot's field selection, which is precisely the point
Step 4: Test Security Under AI Access
Copilot answers and data agent responses respect row-level security and workspace permissions. That statement is only comforting when RLS is actually correct.
Before enabling AI features on a model, test with representative accounts: a user restricted to one region, one legal entity, one cost centre. Ask the questions they would ask and confirm the answers are filtered. With traditional reports a misconfiguration shows up as a support ticket; with conversational access it shows up as a fluent answer containing data the user should never see, as we discuss in our row-level security guide.
Step 5: Endorse, Concentrate, Monitor
AI grounding concentrates traffic onto the models the platform trusts, so tell it explicitly which ones those are.
- Certify the golden models; promote the good ones; leave experiments unmarked
- Retire or hide the duplicates that grew up over the years, because every near-copy of a sales model is a chance for AI to pick the wrong one
- Watch what people ask once Copilot is on: failed and oddly-answered questions are a free, continuous audit of where the model's vocabulary and the business's vocabulary diverge
Performance matters too. Conversational access produces bursts of ad-hoc queries, and Direct Lake or a well-tuned import model keeps answer latency acceptable. A model that takes thirty seconds per visual will not become a good conversational citizen.
The Payoff
Model preparation is the least glamorous line in any AI programme and the one with the highest return. It costs a few weeks on the models that matter, it improves ordinary reporting immediately, and it converts Copilot from a novelty into an interface your organisation can actually trust. Then the same foundations serve whatever agents you build next.
Sources and Further Reading
Frequently asked
Because Copilot does not answer from training data; it translates the user's question into queries against your model. The names, descriptions and relationships in that model are the only business context Copilot has, so their quality directly sets answer quality.
Renaming into business language and adding descriptions. A measure the business calls adjusted revenue must be named Adjusted Revenue with a description saying how it is calculated. Most wrong Copilot answers trace back to the model exposing internal jargon the user never uses.
Start with the two or three models that carry the most decision-making weight. AI grounding concentrates traffic onto endorsed, high-quality models, and that concentration is healthy: it is better to have three models that answer correctly than thirty that answer inconsistently.
All of it does. Business-language naming, correct relationships, descriptions and a clean star schema are exactly what makes self-service BI work for humans. AI preparation is ordinary modelling discipline with a new, more impatient consumer.
Copilot and Fabric data agents respect RLS: users get answers filtered to their permitted data. That is only reassuring if RLS is correct. Test with representative viewer accounts before enabling AI features, because an RLS gap surfaces as a data leak through a chat window.
Endorsement (promoted and certified labels) marks which models the organisation trusts. AI experiences use endorsement to prefer authoritative models when resolving questions, so certifying your golden models steers both people and machines to the right source.
