Flux & Variance

When automated flux drafting helps, and when it costs you more

Set a tool to read an account's detail and write a flux draft while you work a different account by hand. It's a real speedup on the right account, and a real cost on the wrong one.

The pattern is simple enough to describe. You have ten accounts to explain. You pull the detail for the first, set a draft running against it, and while that's working you pull the second and do it manually — pivot the detail, isolate by whatever dimension the account is organized around, write the explanation.

What "setting a draft running" actually means

Concretely: a general-purpose AI tool, pointed at the general ledger detail for one account, with instructions to write a flux explanation in a particular style — the same comparison you'd normally run, month over month most often, formatted the way your team writes commentary. You start that running on one account, then turn to a second account and work it by hand while the first one is still going.

Nothing about this requires custom software. It's a general assistant given a specific ledger export and a specific instruction, running in the background of a spreadsheet while a person does something else in parallel. The mechanism is unglamorous. What's underexplored is when it's actually worth doing.

When you finish the manual one, you check the draft. If it's usable, you clean it up and move on. Then you start a draft on the next account and work another by hand. The goal is to always have one running — which assumes you already know which accounts are ready to work. A free tracker handles that part, ticking off as subledgers and entries close.

When it works, it genuinely compresses the day. Two accounts progress at once and the drafted ones need editing rather than building.

What "usable" actually looks like

A usable draft names what moved and why in terms you can check. It's rarely ready to submit as written — the common edits are cutting length and removing explanations of movements that didn't need explaining.

That second one is worth naming as a pattern. Automated drafts tend toward completeness: every movement described, every number restated, nothing left out. Human commentary is selective, because a preparer knows most of the small stuff doesn't warrant a sentence. Over-completeness is the most reliable tell that something was drafted rather than written, more than any particular phrasing.

Where it fails, specifically

The failure isn't a bad prompt. It's an account where the reason for the movement isn't in the data you handed over.

The clearest case is an account carrying a lot of journal entry activity. The draft can see that entries posted and what they netted to, so it produces something technically accurate and useless: accruals increased by some amount month over month. That's the variance restated, not explained. What each of those entries was actually for isn't in the ledger line — it's in the entry itself, and often in what someone knew when they booked it.

The draft can tell you the balance moved and by how much. It can't tell you what five separate entries were each doing, because that isn't in what it was given.

So you open the entries. Five, sometimes more, clicking into each one to work out what it was and how it relates to last month. That's the manual work you were trying to avoid, and you're doing it after having waited for the draft.

The cost isn't zero when it fails

This is the part that makes routing a real decision rather than a free option.

If handing an account over always cost nothing when it didn't work, you'd hand over everything and keep whatever came back. But a failed handoff costs the wait, the time spent reading an output to decide it's unusable, and then the full manual effort you would have spent anyway. The parallelism you thought you had wasn't there.

The worst version is handing over an account you'd have found hard to do manually. The reasoning feels sound — this one's difficult, maybe the tool helps. But difficulty for you and difficulty for the tool often have the same cause: the reason for the movement isn't sitting in the detail. So it fails, and now you're behind on the account you were avoiding, with less time than when you started.

When it does handle one of those well, it's the best result you get all day. Those cases are real. They're also easy to forget relative to the ones that went badly, which is worth knowing about yourself when you're deciding whether any of this is helping.

Partial output is sometimes salvageable

Not every miss is total. If a draft got the shape right and only lacked context on a few entries, supplying that context — pasting in what the entries actually were — is often faster than starting over.

The distinction worth making quickly: is this output wrong, or is it incomplete? Incomplete is worth feeding. Wrong is worth abandoning. Re-prompting a wrong output tends to produce a differently-worded wrong output, because the problem was never the wording.

What this means for which account you hand over

The routing decision is the actual skill here, and it follows from whether the reason for the movement is retrievable from what you're handing over.

Accounts where the detail carries the reason — activity organized by customer, vendor, or category, where reading the detail tells you what happened — draft well. Accounts where the reason lives inside journal entries, or in someone's knowledge of why an entry was booked, don't. Not because the tool is weak, but because the input doesn't contain the answer.

Which means the useful question before handing an account over isn't "is this one hard." It's "could someone who has never seen this account work out what happened, using only what I'm about to hand over." If the answer is no, doing it yourself is the faster path, and starting there rather than after a failed draft is most of the time saved.

Most of this surfaces again at flux review. The free flux template flags which balances moved enough to explain. The CloseOps Flux & Variance System ($79) starts from your trial balance instead: confirm each account's classification and it builds the income statement and balance sheet flux statements, then ranks what to investigate.

Found something wrong or missing? Tell me here — anonymous, thirty seconds.