Commentary as data: making committee narrative auditable
Risk packs say things. Nobody stores what they said. Here is what changes when the narrative is generated from governed measures, approved by a named person, and written back into the model as a table.
2026-09-30 · 8 min · Erdni Okonov, CFA
The problem with packs
Every risk committee pack has two layers. The numbers, which come from a system and can be traced. And the commentary, which comes from a person, is written the night before in PowerPoint, and vanishes the moment the meeting ends. Six months later, when the examiner asks why an amber rating was left amber for two quarters, the numbers are still there and the reasoning is not.
Most reporting programmes govern the first layer and ignore the second. That is backwards. The commentary is where the judgement lives.
The pattern
Treat commentary as data. A draft is generated from the governed measures in the semantic model, using a language model that can see the numbers, the prior period’s commentary and the definitions. A named person edits and approves it. The approved text is stored as a row: measure, period, author, approver, timestamp, and a staleness flag that trips when the underlying data refreshes. Then the table is exposed back into the same semantic model the pack reads from.
Nothing about the committee’s experience changes. The pack still has a paragraph next to each chart. What changes is that the paragraph now has a history.
What it unlocks
Four things, in the order clients notice them. First, continuity: the next draft starts from what was said last time, so the tone does not reset every month. Second, said-versus-happened checks: the commentary that predicted a breach would clear in Q2 can be read against the Q2 number automatically. Third, queryable narrative: Copilot over the model can answer “what did we say about counterparty concentration in the last four packs”. Fourth, an audit trail an examiner recognises, because every sentence has an author and an approval.
How it is built
A commentary table in the lakehouse or Dataverse; a draft-and-approve workflow in Power Apps with individual identity; a generation step that reads only governed measures and is not allowed to invent numbers; and a relationship from the commentary table back to the measure and period keys in the model. The stack is not the point; we have designed it for Fabric and for Databricks with a semantic layer. The discipline is the point: no free text without a key.
Where it applies
Anywhere a number and a sentence travel together: risk and ALCO packs, finance committee narrative, portfolio-manager monthly letters at a fund, model-validation conclusions. If someone will one day be asked “why did you say that”, the sentence belongs in a table.
Want to see it on your own pack?
Send us one committee pack, anonymised. We will show you the same page with commentary as data.