Grounding Is the New Formatting

There was a time when the person who could make a report look right was the most valuable person in the room. Conditional formatting, a clean measure name, the right number of decimals, a title that read like a sentence. It was craft, and it was visible — you could point at two dashboards and say which one was built by someone who cared.

That craft didn't disappear. It moved. The same instinct that used to make a report legible to a human now makes a model legible to an agent — and this time, legibility isn't polish. It's whether the answer is correct.

Grounding is the new formatting. The descriptions, synonyms, and linguistic metadata you attach to a model do for AI exactly what formatting used to do for a reader: they turn a raw number into a meant one.

The two columns

Picture two versions of the same field.

The first is called F1. No description, no synonyms. It's a currency value, but nothing in the model says so. An analyst who's been here three years knows F1 is fiscal-year-to-date revenue, net of refunds, excluding the two test accounts finance keeps forgetting to close. None of that is written down.

The second is called revenue_ytd. Its description reads: "Fiscal-year-to-date net revenue. Excludes refunds and internal test accounts. Resets on the fiscal calendar, not January 1." It carries synonyms — "sales," "top line," "YTD rev" — and it's tagged as currency.

Point an agent at the first and ask "how's revenue tracking this year?" It finds a column that looks like money, sums it, and answers with total confidence. It might even be close. It will also silently keep the refunds and the test accounts, and it will use the calendar year, because nothing told it otherwise. You get a number. You don't get your number.

Point it at the second and it doesn't have to guess. The description is the grounding. The synonyms mean it recognizes the question even when the user says "top line" instead of the column name. The currency tag means it answers in dollars, not a bare float. It's the same query it would have written anyway — now it's the right one, phrased the way you'd phrase it.

That's not a smarter agent. It's a better-described model.

The platforms already bet on this

This isn't a style preference; it's where the tooling went. Microsoft's Copilot grounds its answers on the semantic model — the tables, measures, and relationships, plus the linguistic metadata: the synonyms, descriptions, format strings, and data categories you set on the model. Set them well and it answers in your language. Leave them blank and it falls back to guessing from column names — which is exactly the failure mode a fluent guesser is prone to.

Microsoft is even retiring the older report-surface Q&A feature — the one that read questions against a specific report — in favor of that model-grounded path. Read that as the signal it is: the industry decided the place to teach a machine what your data means is the model, not the report sitting on top of it. Which is the whole argument.

The pattern holds from the other side too: a model shipped without descriptions and synonyms underperforms one that has them — not because the math is different, but because the machine is working from less. A well-grounded field is one where the meaning travels with the number instead of living in someone's head.

Formatting was always about meaning

Here's what the old formatting craft understood that we're relearning: presentation was never really about looking nice. A well-formatted report was one where a reader could tell, at a glance, what a number meant — that this was a percentage and that a count, that this figure was final and that one still running. Good formatting was compressed meaning.

Grounding is the same job for a different reader. An agent can't see your bold headers or your color scale. What it "reads" is the metadata: the description, the synonym, the data type, the relationship. That's its formatting. Skip it and you've handed it the spreadsheet equivalent of a wall of unlabeled gray numbers — and asked it to be confident anyway.

The craft didn't get less important when the audience became a machine. It got less forgiving. A human squints at an ambiguous label and asks someone. An agent doesn't squint. It commits.

The test

Open your most-used model and read three column descriptions out loud — actual descriptions, not the column names. If someone who started last week couldn't tell you what the number means, what it excludes, and when it resets, your agent can't either. It'll do what it always does with a gap: fill it, fluently, and move on.

The fix isn't a bigger model or a cleverer prompt. It's the least glamorous work there is — writing down what your fields mean, one description at a time, until the meaning lives in the model instead of in the tenure of whoever built it. That used to be how you made a report worth trusting. Now it's how you make an agent worth pointing at your business.