You asked your AI for “active users” three times and got three numbers. Which one is right?

A number is only as trustworthy as the definition behind it. If the definition lives in someone’s head, every new question, tool or assistant picks its own. Certifying a metric moves the definition out of heads and into one approved place.

One question, three answers

Take “How many active users did we have last month?” Three reasonable people can answer it three ways.

Nobody made an error. Each count is correct for its own definition. The trouble starts when a meeting treats the three as one number, and compares this month’s figure from one tool to last month’s from another.

A certified metric answers the question before it is asked: “Active user means a customer who placed an order in the last 90 days, counted by order date, test accounts excluded. Owner: the head of growth. Approved on this date.” Every report, sheet and assistant then uses that sentence.

What this means for your decisions

• A change in the number should mean a change in the business. If the definition shifts between two reports, the gap is the definition, not your customers.
• Write the word, not just the formula. “Revenue” needs to say whether refunds are taken out. “Churned” needs to say how many quiet days count.
• Name an owner and a date. A definition nobody owns drifts. When it must change, change it once, say so, and note which past numbers are no longer comparable.

How large companies certify a metric

Big organisations treat this as infrastructure. A data team defines the metric once in a semantic layer that sits on top of the warehouse, marks it approved, and every dashboard and assistant is pointed at that layer instead of the raw tables. Snowflake has semantic views, Google Looker has LookML, Microsoft has Power BI semantic models, ThoughtSpot has Models, and Tableau has Tableau Semantics.

The news hook is Databricks. On 2026-11-11 it launched plugins that put its Genie data agent inside ChatGPT, Claude, Microsoft Copilot Cowork and Grok Bot, and a Genie MCP server for other assistants. Databricks describes a “Genie Ontology” in which a team can “define, govern, and certify” business context, with certified assets generally taking precedence over what Genie learns from usage. Databricks says answers come back with citations. These are Databricks’s own statements; the post gives no accuracy figures for the launch, and the percentages in it are illustrations, not measured results.

The pattern is the same across vendors, and it is worth knowing: each one answers over a definitions layer that someone built first. On how much that layer matters, two outside sources point the same way. An independent comparison of Genie and other tools, run by a competing vendor, noted that its scores partly measure how well the data model was prepared. And users in Databricks’s own community forum report that accuracy rose most from preparing the data, not from rewording prompts.

This works well when you have a warehouse, a data team and an admin. Databricks’s own requirements include data registered in Unity Catalog and a SQL warehouse, set up by a person with admin rights.

What to do if you have no data team

You do not need a semantic layer to get most of the benefit. You need the habit it enforces.

1. Write each definition down in a sentence. One shared page or note is enough: the metric, what counts, what does not, the time window, the owner, the date.
2. Give the definition to the assistant every time. Paste it into the question, or keep it in the project notes your assistant reads, and ask it to say which definition it used. If the answer does not name one, treat the number as unchecked.
3. Rerun the same method on new data. The next month’s number is comparable only if it was computed the same way. Keep the method, and the code if there is any, next to the definition, and run it again on the new file instead of asking a fresh question.

That third step is where we sit. MCP Analytics runs a named, tested method on your own file or connector, hands back the method and its code, and lets you rerun the same analysis on next month’s data. We are not an enterprise semantic layer, and we do not certify your definitions for you. If your company already has one, use it for the governed number and use a tested method on top of it.

What a certified metric is not

• It is not a correct number. It is a number with an agreed meaning. A bad definition, certified, is a consistently bad number.
• It is not a permission or security system. Those are separate, though some products ship them together.
• It is not the same as a tested statistical method. A certified metric says what you counted. A method says whether a change in it is real or noise.
• It is not only for big companies. The enterprise tooling is, the idea is not.

Sources

• Databricks blog: Bring your business context to ChatGPT, Claude, Microsoft Copilot Cowork and Grok Bot (2026-11-11), a vendor statement
• Databricks documentation: Genie Ontology, a vendor statement
• Databricks documentation: Genie MCP server, a vendor statement
• The semantic-layer products named above are described from each vendor’s own documentation. The note on benchmark scores is from an independent comparison that was itself run by a competing vendor. All read 2026-11-12.