Putting it together
Five lessons of theory are worth nothing if they live in your head as good intentions. What follows is a month-end routine you can copy into a document today, with the verification stitched into each step rather than bolted on at the end.
That distinction matters more than it sounds. A single review at the end of a process catches almost nothing, because by then you have lost track of which figures came from where and everything has acquired the comfortable patina of work already done. Checks belong next to the step that created the risk, while you still remember what you did.
Build this outside the site. A document in your firm's system, a note file, a laminated card by the monitor. Nothing here is stored anywhere, and there are no accounts to create. It is yours to keep and yours to change.
In plain English
- Workflow:
- A written sequence of steps with a named check attached to each one, so the checking is not left to memory.
- Working papers:
- Your evidence of what you did. An AI output is never evidence. Your check of it might be.
- Human in the loop:
- A person approving each output before it has any effect. In this workflow, that person is you at every single step.
- Prompt pack:
- A small reusable set of prompts you keep and refine, rather than rewriting from scratch every month.
The routine
MONTH END: [client] [period]
STEP 0. BEFORE ANY AI TOUCHES ANYTHING
[ ] Is this tool approved by my firm for client data?
[ ] Have I redacted names, account numbers and references?
[ ] Would the amounts and dates alone identify this client?
If yes, stop and work without the tool.
STEP 1. CATEGORISE THE BANK DATA
[ ] Pasted my actual chart of accounts, exact names
[ ] Instructed UNSURE instead of guessing
[ ] Applied this client's house rules
[ ] Reviewed EVERY row marked UNSURE
[ ] Spot checked 10 confident rows: largest amounts,
possible capital items, VAT-sensitive items, unfamiliar merchants
[ ] Nothing auto-posted. Every entry approved by me.
STEP 2. THE NUMBERS
[ ] No figure in this pack came from a chat reply
[ ] Every total produced by a formula or by the accounting system
[ ] Row count checked before and after every filter or paste
[ ] Recalculated by hand: biggest item, one exclusion, one credit note
STEP 3. RECONCILE
[ ] Reconciliation done in the system, not by AI
[ ] AI used only to list candidate causes of unexplained differences
[ ] Every flag traced to a source document
[ ] Flags that came to nothing recorded too
STEP 4. ANOMALY SWEEP
[ ] Asked for hypotheses, not conclusions
[ ] Every flag has a named check
[ ] Checks performed, outcomes written down
STEP 5. COMMENTARY
[ ] Drafted with placeholders, no generated figures
[ ] Every figure typed in by me, read off the source
[ ] Direction words checked: up, down, improved, fell
[ ] Any figure that appeared during a rewrite: deleted or verified
[ ] Causal claims removed unless I know them to be true
STEP 6. BEFORE IT LEAVES MY DESK
[ ] For every number, I can name the cell or report it came from
[ ] Nothing filed or signed rests on unverified output
[ ] Judgment on treatment is mine and documented as mine
Checkpoint
Copy that checklist somewhere you will actually see it during month end. A checklist in a folder you never open is a document. A checklist next to your monitor is a control.
The prompt pack
Keep these five in a file and paste them in each month. Refine them as you learn what your clients throw at you. A prompt you have tuned over six months is worth far more than a clever one you wrote once.
-
CATEGORISE
Use ONLY these categories, exact names: [paste chart of accounts].
Never invent a category. If ambiguous or you are not confident,
write UNSURE and name the candidates. Never guess.
Client house rules that override your reasoning: [list them].
Return: description, category, UNSURE yes or no, one-line reason.
-
FORMULA, NOT ANSWER
Do not calculate anything. My columns are [list them],
data runs row [2] to row [N].
Give me: the inclusion rule in plain English, any rows you are
unsure about, a formula that applies the rule, and a second formula
that counts the matching rows.
-
ANOMALY HYPOTHESES
Flag what LOOKS unusual. You are generating hypotheses, not conclusions.
Do not say anything is confirmed, correct or an error.
Group as: possible duplicates including near-matches, unusual amounts,
timing oddities, description and coding mismatches.
For each: row reference, what caught your attention, and the exact
check I should run. Do not flag anything you cannot give me a check for.
Rank by amount at stake.
-
COMMENTARY WITH PLACEHOLDERS
Write commentary for a reader who runs the business and has no
finance training. Write NO figures. Use placeholders FIG1, FIG2 and so on.
At the end, list each placeholder with a plain English description of
the figure I need and which statement it comes from.
Tone: plain, warm, second person. Explain any term you use in the
same sentence. About 250 words.
The story in my words: [describe it, no figures].
-
HOW COULD THIS BE WRONG
I calculated [figure] as [amount] using this method: [describe it].
Do not recalculate and do not give me a number.
List, most likely first, the ways this method could be wrong:
wrongly included or excluded rows, sign errors, date boundaries,
duplicates, anything else.
For each, give me the specific check to run in my spreadsheet.
Prompt five is the one people underuse. Asking how a figure could be wrong costs one minute and uses the tool for the thing it is genuinely brilliant at, which is being inventive about failure modes. It has read more accounting horror stories than any of us.
Two failure modes to watch for in yourself
Creeping trust. Month one you check everything. Month four the categorisation has been right so often that you skim it. Month seven you stop spot checking. The tool has not become more reliable, you have simply become more comfortable, and comfort is not evidence. Keep the sample size fixed no matter how well it has been going.
Verification theatre. Ticking a box without doing the check is worse than having no box, because now there is a written record saying you checked. If you did not read the figure off the source, do not tick the line. An honest blank is defensible. A false tick is not.
โ Weak prompt
Prompt
Here is my month-end pack. Review it and tell me if anything is wrong.
Output
I have reviewed the pack. The figures appear consistent and the commentary reads well. One minor suggestion on wording in paragraph two.
It cannot review a pack. It did not recompute anything, it cannot see your source documents, and the reassuring first sentence is generated text like all the rest. You have been told your work is fine by something with no way of knowing.
โ Good prompt
Prompt
Here is my month-end pack with figures redacted to placeholders. Do not tell me whether it is correct, because you cannot know that. Instead give me a list of the checks a reviewer would run on a pack of this shape, ordered by how much damage an error there would cause, and tell me what evidence each check needs.
Output
A prioritised review programme: which figures to trace to source, which totals to recompute, which cut-off tests matter most, and what evidence each one requires.
You get a genuinely useful review programme, and the reviewing is still done by the only party capable of doing it, which is you.
This course is general education and not professional, tax, legal or regulatory advice. It makes no claim about what any jurisdiction, regulator or professional body requires of you, and this checklist is a starting point rather than a compliance programme. Your obligations come from your professional body, your regulator and your own judgment. And permanently: do not paste client data into a consumer AI tool unless your firm has approved that specific tool for that specific purpose.
The Balanced Books badge
Finish this lesson and the Balanced Books badge is yours. It stands for something quite specific: that you can get real speed out of these tools without ever letting an unverified figure reach a client, a return or a signature.
Which brings us back to the sentence this whole course was built around. AI predicts plausible text, and a plausible number is not a correct number. Everything else is detail. You already knew how to do the accounting. Now you know exactly where the machine helps and exactly where it must not be allowed anywhere near the work.
๐ Quiz
Question 1 of 4Why should verification be built into each step rather than done once at the end?