Reports that design themselves

For years, getting a new dashboard meant filing a request and waiting. You described what you wanted, it joined a queue, and weeks later something close-ish arrived. The “can you build me a report” backlog was a permanent fixture.

Now the report generates from a sentence. You describe the view you need in plain language and a draft comes back in seconds, not sprints. The queue stops being a queue because the first draft no longer needs a person to hand-build it. What used to be the whole job becomes the starting point.

Why now

• Generation got good enough. Models can turn a description into a real, structured report: sensible chart types, groupings, a layout that holds together. Not perfect, but a credible first draft to react to instead of a blank canvas.
• The semantic layer makes it trustworthy. Pointed at raw tables, generation produces one-off dashboards that each invent their own “revenue” and quietly disagree. Drafted from governed definitions, every view inherits the same metrics as the rest of BI, so consistency is built in, not bolted on. Generation supplies the speed; the semantic layer supplies the trust.

What it looks like

A revenue operations lead types, “build me a weekly churn view by segment.”

A naive tool would guess at a churn formula, grab the first segment column it finds, and produce something subtly wrong. A grounded system resolves “churn” to the governed metric, scopes “segment” to the agreed grouping, and assembles a draft (a weekly trend, a breakdown by segment, a table of the segments moving most) in seconds.

Then the human takes over: excludes trial accounts, renames a confusing label, certifies the result. What landed was a near-complete draft, not a ticket to start from scratch. The minutes spent are editorial, not mechanical.

Where it’s heading

From drafting to curating as the default shape of the work. As generation gets more reliable, the analyst spends less time assembling and more time deciding: which views deserve certifying, which definitions need fixing, what the organization should be looking at that nobody’s asked for yet. The report builder becomes a report editor, and the backlog thins because most requests are answered the moment they’re asked.

How we think about it

Generation removes the grunt work; humans keep editorial control. A draft is not a decision. The system should produce the view fast and from governed definitions, but the choice to certify it, to call it the official number, stays with a person. That’s where domain knowledge and accountability live, the same boundary we draw for governed contributions: propose freely, but keep approval human and on the record.

Hold that line and the request backlog is no longer a fact of life. The machine drafts, the analyst curates, and the view shows up in minutes with a person’s judgment still on it.