Module 12The six AI workflows3 min read
Management pack commentary: three reports to a one-page pack
How to turn period reports from Xero or QuickBooks into a one-page commentary in minutes instead of an afternoon, with every variance computed by a script and every unexplained movement labelled as unexplained rather than guessed at.
Management pack commentary turns period reports into a one-page narrative: what moved, by how much, and the two or three things worth a sentence. The draft takes minutes instead of an afternoon, and a human edits and signs it. The rule that makes it safe is that the ledger shows what moved, never why, so any cause the data does not support is written as unexplained rather than invented.
Key takeaways
Never invent a cause. A confident wrong explanation loses the client's trust in the entire pack, not just that line.
One file covering both periods. Two separate exports invite a period mismatch nobody notices until the meeting.
Three to five points, not fifteen. A pack is not a list of every account that moved.
No budget variance without a budget. The ledger does not contain one, so the workflow does not pretend to.
Getting the data out
Export both periods in a single file. In Xero: Accounting, Reports, Account Transactions, set the date range wide enough to cover both months, export to CSV. In QuickBooks Online: Reports, Transaction List by Date, same approach.
One file matters. The script reports how many rows landed in each period, and that count is the check that catches a wrong date format before the number reaches a client.
What earns a sentence
Three tests, and a movement has to pass all three.
- Size. The movement is large enough to change a decision.
- Surprise. It is not what the reader would already expect.
- Actionability. Somebody could do something about it.
An account that moves four per cent on a small base fails all three. Leave it out. The discipline of leaving things out is what makes the page readable.
Install it: skills/management-pack-commentary
Computes account-level movement between two periods with the percentage change, sorted by absolute movement, and reports how many rows fell into each period so a date-format error cannot reach the client.
Get the six workflows on GitHubWorks in Claude, Claude Code, Codex and Cursor. Copy the folder into .claude/skills/ and ask in plain language.
The shape of the page
| Section | What goes in it |
|---|---|
| Headline | One or two sentences. The single thing the reader needs to know |
| What moved | Each point: the figures, the movement, the percentage, and one sentence on why |
| What to watch | The thing that is not yet a problem |
| Not explained | Movements the ledger does not account for, named openly |
The "not explained" section is the one that builds trust. A pack where everything has a tidy reason is a pack somebody made up reasons for.
Why the model must not compute the variance
A commentary built on a hand-totalled number is a commentary that gets withdrawn in a board meeting. Every figure and percentage comes from the script, and the prose matches the output exactly, with no rounding to make a sentence read better.
The reasoning behind that rule, and why it applies to every workflow in this set, is in AI for accountants.
What it cannot do
It does not explain causes. It identifies movements. A human who knows the business explains them.
It does not do budget variance unless a budget is supplied as data.
It does not sign the pack. A named person does.
Common questions
How long does a pack take once this is built?
Minutes to draft, then however long the editing takes. The saving is in the first draft, which is the part that used to consume an afternoon.
Can it compare against budget?
Only if you supply the budget as data. The ledger export does not contain one, and a budget variance calculated against an assumed budget is worse than no budget variance.
What if a movement has no explanation in the ledger?
It goes in the "not explained" section by name. That is a question for the client, and often the most useful line on the page.
More in The six AI workflows
The monthly query chase: one email instead of six
Find every transaction you cannot code and turn it into one client email instead of six.
The client email desk: first drafts of the replies that eat an afternoon
First drafts of the technical replies, with every position marked for a human to verify.
Client onboarding: one intake, three documents that agree
One intake producing an engagement letter, information request and AML checklist that agree.
Fee scoping: price a client from their data, not from a guess
Price a client from measured volume and measured messiness instead of from a guess.
The year-end review pass: a second set of eyes that is not yours
A second set of eyes that runs the same checklist on every file, at the same standard.