The Monday meeting has one line for net revenue last quarter. Growth's slide has one number. Finance's slide has another. Both numbers came from an assistant, and both assistants were asked the same sentence.
Growth's query subtracts discounts and buckets the calendar quarter in UTC. Finance's query subtracts refunds and tax and uses the fiscal quarter in the company's timezone. Neither query is on the slide. The meeting spends its time choosing a number. It does not spend its time on what to do about the quarter.
Until someone can point at the definition that was used, neither figure can be the one the business trusts. "Net revenue last quarter" is not one sentence in one place. It is a slogan two teams have filled in differently, and two assistants will fill in differently again the next time they are asked.
A metric is a query with the arguments written down
People talk about metrics as if they were facts waiting in the database. Revenue is rows. Net is a decision about which rows to subtract. Last quarter is a decision about which clock and which boundary. The grain is a decision about whether a refund in this quarter reduces this quarter or the quarter of the original invoice.
Leave any of those out and you have two defensible numbers. Growth is not wrong that discounts change what you collected. Finance is not wrong that tax was never yours and that a refund reverses a sale. The database can return both. The business can use only one of them in the meeting, or it can use both and label them. It cannot use both under one unlabeled name.
The same split shows up without any AI. Two dashboards, two analysts, two extracts. Assistants make it faster and more polite. Each one produces a clean answer and does not know the other session exists. The collision moves from a back-and-forth between analysts, which at least surfaces the formula, to a pair of slides that both look finished.
Why the usual tools allow two copies of one name
A SQL client has no shared measure. The definition is the last query someone ran. If they save the file, the definition exists for them. The other team does not open that file. They open their own client and write a reasonable query. Both files can be named net_revenue.sql. Names do not reconcile.
A BI tool can hold one measure called Net Revenue, and a disciplined team keeps it that way. The tool will also let a second author publish a second measure with the same words on the chart title. It does not notice the collision. When the assistants are pointed at the warehouse tables instead of at the published measure, they never see the measure. They invent a third.
Excel holds a cell with a label. The formula is in another cell, or in a sheet that was not attached to the email. The label travels. The formula stays. By the time the slide is built, "net revenue" is text next to a pasted integer. There is nothing to diff.
Two text-to-SQL sessions are two inventors. Each one will pick a plausible set of subtractions because the sentence "net revenue" is underspecified and the model has seen many versions of it in training. They will not agree with each other, and they will not warn you that they guessed. The guess is invisible if the statement stays behind the prose.
The tools store queries, charts, and cells. They do not, by default, store the decision that this query is the one the company means by those words. That decision is a sentence someone has to write and a place other people have to look.
One definition, pointed at one query
Write the metric as a sentence that names the query. Include the clauses that caused the Monday collision:
Net revenue for a quarter is invoice totals, minus discounts, minus refunds recorded against those invoices, excluding tax. The quarter is the company fiscal quarter, in the company timezone, by invoice date. The grain is one row per invoice, not per line.
That sentence will be wrong for some other company. It has to be specific enough to be wrong, or it will not settle anything.
Put the query next to the sentence. A query without the sentence is a formula nobody can audit in a meeting. A sentence without the query is a policy nobody can run. You want both, in a place both teams open. The repository that already holds the reporting SQL is a better place than a slide deck.
MeaningGraph is an open registry of shared definitions people keep in Git. A metric sentence can live there with a revision, so "net revenue" has a source someone can point at, and a change is a contribution rather than a quiet edit in a dashboard. The registry works whether or not you use DataTug. datatug.app marks DataTug's integration with MeaningGraph as planned. A definition on meaninggraph.io is not, today, a metric DataTug looks up for you. Link the sentence from the query, in the repo you already have.
When two numbers appear, put both queries on the table before the debate about the business. The difference is usually a filter, a clock, or a subtraction. Find that line. Then decide which definition the business will use, and who is allowed to change it. A tool will not make that decision. A meeting can, once the two statements are visible.
What you should not expect this week is a governed approval workflow: a queue, a sign-off, a lock that stops a second dashboard from using the words. That workflow is real work, and it is a step beyond writing the definition down. Write the definition first. The workflow has nothing to approve until the sentence exists.
Where DataTug fits
datatug.app includes "save, share and reuse queries." DataTug helps by saving the query with the result, so two people can open the same question again in the web app, the CLI and terminal UI, or AI chat. If both assistants are pointed at that saved query, the second one does not have to invent a second formula. If they are not pointed at it, they will invent one, in DataTug or anywhere else.
Saving a query is not an approval workflow, and this article does not claim one. There is no committee inside the product that declares a metric official. There is a query two people can open, with the rows it produced, and a sentence you keep beside it. That is the part that stops the unlabeled collision.
The sources you can open in that project, as the site marks them working today, are SQLite, PostgreSQL, MySQL, Oracle, SQL Server, BigQuery, inGitDB, and REST APIs. Personal and public projects are free. Put the fiscal-quarter query there, with the sentence in the name or the notes, and send the link to the Monday meeting instead of a pasted integer. The demo is a place to see a saved question opened again, if you would rather watch that on public data first.
Pain pointsNext: The central database that never stays built