AI Commentary That Shows Its Work, Line by Line

Every plan page eventually needs a sentence explaining what happened: why revenue missed plan by 12%, why the swing in opex looks the way it does. Someone in FP&A usually writes that sentence by hand, every period, for every line that moved. Rexfin generates a first draft of that sentence automatically, but it builds the draft the same disciplined way it builds a chat answer, not by asking a model to free-associate over a spreadsheet.

The evidence comes first, deterministically

Before any model is involved, Rexfin computes the facts a narrative might reference: plan-versus-actual variance rows, year-over-year deltas pulled from comparative atoms already in the filing, and advisory flags for swings that cross a threshold (a year-over-year move of 25% or more, or a plan gap of 10% or more). None of that is written by a language model. It’s arithmetic over numbers that are already cited elsewhere in the product.

Only once that evidence block exists does a model write the explanatory prose, and it writes from that block alone: the same investigator-then-writer split described in AI chat with citations. The same citation-truthfulness gate that fences chat answers fences this prose too: if the draft contains a numeral that isn’t backed by a row in the evidence, it’s rejected. Rexfin retries once with the same evidence, and if that also fails, it falls back to a deterministic template built directly from the numbers (no free-form prose, but no blank page either). You never see a narrative that snuck a number past the gate, because the two paths that could produce one both stop before publishing.

Every row knows its own confidence

Not every number behind a narrative carries the same weight. A row can be audit-grade (straight from a filed statement), derived-and-cited (a calculation over cited atoms), or synthetic-and-unaudited: most commonly a plan or budget figure someone entered rather than a filed number. That distinction propagates: if a narrative’s premise rests partly on an unaudited plan number, the commentary is labeled accordingly rather than presented with the same confidence as a variance computed entirely from audited actuals. Anywhere this shows up in the interface, it carries clear “AI-drafted, unaudited” framing, and it’s suppressed entirely rather than shown half-working if the product is running offline with no model available.

Where it shows up

The main surface is the “what changed and why” note on a plan page, with citation chips that drill back to the underlying row. The same commentary can also be dropped into a PDF or PPTX board pack, behind the identical re-run of the truthfulness check, so a narrative doesn’t get less scrutiny just because it’s headed for a deck instead of a screen. If you’re assembling a full pack rather than a single page, see board pack export for how the exported document is put together, and cited PDF/XLSX exports for how citations survive the trip into a file.

What it won’t do

If a filing has no citable comparative-period figure to compare against (a first-year filer, or a period with a gap in the data), the narrative refuses to manufacture a trend rather than invent one. That’s a deliberate limit, not a bug: a confident sentence about growth that doesn’t have a prior period to measure against is exactly the kind of plausible-sounding number this whole approach exists to prevent. And the commentary is a draft. It’s meant to save the first-pass writing, not to replace someone reading the numbers and deciding whether the explanation actually holds up: the assurance labels exist precisely so a reader knows which parts to trust at face value and which parts to double-check.

See the full set of related capabilities at Rexfin product tour, including how answer voices can restyle a narrative’s framing without touching a single figure in it.