Skip to tool
MM/FINIMO
MM/FINIMOAI in FP&A, the controlled way

Reviewing AI output

Find the Five

Here is a draft of September's variance commentary for Helix, the kind an AI assistant produces when nothing constrains it. Five statements in it would not survive a controller's review. Tap the ones you would send back.

Helix Consumer Goods is a fictitious company. The draft was written for this exercise, with the errors planted on purpose.

The numbers the draft was given

Cost centreForecastActualVariance
CC100 Digital CommerceRevenue7,992,0007,733,233-258,767-3.2%
CC200 WholesaleRevenue2,808,0000-2,808,000-100.0%
CC300 MarketingOpex1,226,0161,193,323-32,693-2.7%
CC400 LogisticsOpex1,003,1041,001,596-1,508-0.2%
CC500 G&AOpex613,008618,471+5,463+0.9%

September 2026, forecast against actual, in EUR. Materiality: 5% or 100,000 EUR.

The draft

Which statements would you send back?

variance_commentary_2026-090 flagged

The lesson

Fluent is not the same as sourced

  1. Every mistake here reads well. That is why a quick read is not a review.
  2. The checks are mechanical. Does the number tie. Is the cause in the data. Is a zero really a zero. Is the sign right. Who approved it.
  3. Build the checks into the draft. Then the reviewer is the last control, not the only one.

Where to start

Take last month's commentary and run these five checks on it, whoever wrote it. Note which check fails most often. That is the first control to build.

Free series

AI in FP&A, the controlled way

Short, practical tools that show finance teams where AI really helps, without losing lineage, approvals or the controller's trust. Follow me on LinkedIn for the next one.

Not sure where to start? Take the 10-question AI Reality Check