For thirty years, the thing that sat between people and company data was a table. Analysts learned its columns, wrote queries against it and built dashboards on top. When two people disagreed, they compared SQL.
That arrangement worked because the reader was human. A person can look at a table called fct_opps_v3_final, ask a colleague what it contains and work out which of two similar columns is the right one. An agent can’t do any of that, and over the next few years, most of the questions about your business will come from agents.
So the contract between software and data needs to be something more precise than a table. This series has argued, from eight directions, that it should be a definition. This post pulls the argument together and links the whole series so it can be read from one page.
What a definition has that a table doesn't
A table tells you what is stored. A metric definition tells you what something means, and it carries that meaning with it wherever it goes.
A governed definition has five properties, and each one answers a failure we covered in the last two months.
- It has a name. An agent can ask for win_rate without guessing which columns to combine. The name is the interface.
- It has a stated basis. The denominator, the filters and the time handling are written down, so "win rate" can't quietly mean three different things on one slide.
- It carries its caveats. A closing balance isn't summed across days. A ratio can't be averaged. The rule travels with the metric rather than being rediscovered by whoever writes the next query.
- It has a history. A definition lives in a repository, so changing it is a reviewed change with a name and a date on it.
- It can be refused. If a question can't be answered from what is defined, the system says so, instead of improvising a plausible number.
None of those can be bolted onto a table afterwards. They come from deciding, early, that the definition is the thing everything else passes through.
How it fits with series one
Series one argued that a model should do two jobs, reading what someone meant and writing about the result, and that everything in between should be deterministic code. Series two supplies the missing piece: deterministic according to what?
The answer is a governed definition. The trace, the computing statement, the year-to-date rule and the permission scope all rest on the same thing. A trace can name the metric that was used. The computation is repeatable because the definition is fixed. The YTD rule is a line of configuration instead of a convention. Permissions apply at read because the metric says what is being calculated and entitlement says for whom.
In Equative Insights OS, MetricFlow makes that definition portable and reviewable, and Equative makes it answerable by both people and agents, with the trace attached.
Series two on one page
If you’re arriving here first, this is the series in order. The Tuesday posts read the market; the Thursday posts show what we built.
01 · After Dreamforce: the agents are ready. Are your numbers?
Salesforce’s announcements moved the conversation from what agents can do to how to run them. Observability tells you what an agent did; it doesn’t tell you whether “win rate” meant the same thing in the agent, the forecast deck and the finance close.
02 · We moved our metrics into MetricFlow. Here's why.
Our decision record. Every measure is now a dbt semantic model and metric, versioned in git, and MetricFlow generates the SQL. The model picks a defined metric; it never writes the query.
03 · Every vendor has a semantic layer now. Whose is yours?
A fair map of Tableau Semantics, Snowflake Semantic Views, Databricks Metric Views and MetricFlow. Each is consistent inside its own platform, and almost nobody runs on one platform. Three questions to ask any vendor about exporting, consuming and leaving.
04 · Define once, answer anywhere
No. Ossie is a format for describing metric definitions so they can move between tools. You still need a semantic layer, or a metric engine, to store those definitions, compile queries from them and apply access control.
05 · An open standard for what a metric means
The Open Semantic Interchange, now Apache Ossie, and what changes when definitions become portable. Adoption is early, and a standard is only as useful as the number of tools that read it.
06 · Your CRM is a source, not the definition
Salesforce records what your commercial team did, not what the business earned or risks. Model it as a source, join it to finance and risk data, and define cross-system metrics once, with entitlement still enforced at read.
07 · Soon, most of your metric queries won't come from people
Long-horizon and multi-agent systems mean software will query metrics far more often than analysts do. Agents given raw tables each invent their own definition, and nobody reviews the SQL. What they need is a small governed catalogue to call by name.
08 · The metric is the interface
You’re reading it.
This series picks up from the first one. If you haven’t read it, AI where it counts, deterministic everywhere else links all eight posts of Product & Pitfalls from a single page.
The standard we'd ask you to hold any vendor to
Series one ended with four questions to ask in any demo, ours included. Series two adds a fifth.
- Run the same question twice. Do you get the same number?
- Show me the statement that computed this figure. Who wrote it?
- Could this answer have drawn on data I'm not cleared to see?
- What did this answer cost, measured and not estimated?
- Show me where this metric is defined, and who last changed it.
That last question is the one that separates a semantic layer from a label. If the answer is a file, a commit and a name, the metric is an interface you can build on. If the answer is “it’s in the dashboard”, it isn’t.