Variance Commentary the AI Drafts and You Approve

Nov 18, 2026By Yvon Descieux
Yvon Descieux
Cover graphic of a ledger with computed variance columns and a drafted comment beside each flagged line awaiting approval, with the words Code computes every variance and Your team approves

Variance analysis automation lets code compute every flux and lets the AI draft the commentary. Your team approves each line before anybody else reads it.

Where the days go

When we rolled up 17 discovery calls with finance teams, 6 companies named the same monthly grind. Board deck, flux and PVM assembly from a trial balance, often handled by a third party, with days of churn each cycle.

The math is not what takes the days. The variance on an account is one subtraction. The time goes into everything around it. Pull the trial balance. Line it up against the comparison period. Decide which lines cross the threshold. Ask the account owner what happened. Turn their answer into a sentence. Paste the sentence into the deck. Then do it again when a late entry moves three numbers.

That is a workflow problem, and most of it splits cleanly between code and a language model, with a person still holding the pen at the end.

What code does

Code reads the trial balance export for the current period and the comparison period. It computes the variance for each account in amount and in percent, flags the lines that cross thresholds your team sets, and where the data supports it, breaks a change into price, volume and mix. The same structure drills down by plant or branch, so the commentary follows the business the way your leadership reads it.

Every figure is calculated by deterministic code. The same file goes in, the same numbers come out, every run. Headline totals are summed from the detail underneath them, never copied from text a model wrote. When a late entry lands, you rerun, and every flux updates together instead of one cell at a time.

On calls we describe the split as roughly 98 percent Python and 2 percent AI. Variance work fits that split well, because almost all of it is arithmetic and lookup.

What the AI drafts

The model receives finished numbers, never raw data to add up. For each flagged line it drafts commentary that says what moved, by how much, and which accounts or units drove it. That part comes straight from the computed results.

What the model cannot know is why. A price change, a delayed shipment or a reclass lives in someone's head, not in the trial balance. So when the reason needs business context, the draft says so and leaves a clear question for the account owner, instead of inventing a plausible story.

If the model's wording and the computed numbers ever disagree, the computed numbers win, and the report says so.

Your team approves every line

Nothing reaches the deck without a person signing off. The FP&A lead reads each draft and approves it, edits it, or sends it back to the owner with the question attached. The job shifts from writing commentary in a blank cell to editing a draft that already has the right numbers in it.

We have built a PVM sample, executive views and a variance design for finance teams working through exactly this. Each follows the same split. Code produces the numbers and the drill downs. The model writes the first draft. People approve.

Expect the IT questions early. On 5 of 6 recent finance calls, security and upkeep came before any build. Who maintains the Python, where it runs, CSV or API, what it costs each month, and how to turn it off. Those deserve plain answers. Ours are that it runs inside your walls, starts from exports you already produce, and you own the code with a runbook and no subscription.

To start, bring three things.

1. Trial balance exports: The current period and the comparison periods, at account level, in the format your ERP already produces.

2. Thresholds: The amount and percent rules that decide which lines get commentary today.

3. Owners: One named person per account group who answers the questions the draft leaves open.

You can see how this fits with the rest of our finance systems work. Let code compute the flux. Let the model write the first draft. Keep approval with the people who sign the deck.