Approval is the product

There is a moment in every AI data tool where it tells you what it found. Most tools treat that moment as a progress update — something to scroll past on the way to the chart. We treat it as the point of the whole thing.

When the agent has read your files, it produces a proposal: this is the grain, these are the columns, here is how I think these two files relate, here are the measures I would define, and here are the assumptions I made that you might disagree with.

Nothing is loaded. No table is created. No dashboard exists. Until you approve it.

Why this matters more than it sounds

The failure mode of automated data preparation is not that it gets things wrong. It is that it gets things plausibly wrong, silently, and you find out in a board meeting. A join that multiplies rows. A date column read as text. A filter that quietly excludes a region.

A proposal you have to read is a forcing function. It puts the assumptions in front of the one person who can evaluate them — the person who actually knows the business — at the only moment when correcting them is cheap.

Batches, without losing the point

Work does not arrive one file at a time. Fifteen files and ten datasets is a normal Tuesday, and making somebody approve fifteen proposals one at a time would turn a safeguard into a nuisance — which is how safeguards get skipped.

So you can approve all of them at once. The bar sets one flag; each proposal acts on it as though its own button had been pressed, with the same loading state, the same errors, the same instruction to the agent. The reply only continues once every proposal is approved, and then the agent is told to link them and build one dashboard across all of it — because without that word, it reasonably builds one per dataset.

Same safeguard. One click.