Most finance people using AI are using it badly — not because the tools are weak, but because they’re asking one-line questions and getting one-line answers.
“Summarize this P&L” produces a summary you could have written yourself. The same tool, given the role, the context, the format, and the specific question you actually care about, produces something worth reading. The difference isn’t the model. It’s the prompt.
Below are nine prompts we’ve refined through daily use in a working accounting practice. Each includes what it’s for, the prompt itself, and — more usefully — where it goes wrong and what you still have to check.
Read the section on confidentiality first. It’s the part most guides skip, and it’s the part that matters most if you handle other people’s financial information.
Before Anything Else: What Never Goes Into a Chatbot
General-purpose AI tools are not a secure environment for confidential information by default. Consumer tiers of most products may retain inputs, and some use them for training unless you’ve explicitly opted out or you’re on a business plan with different terms.
DON’T PASTE
✗ Client or company names alongside financial detail
✗ SIN, business numbers, or CRA account numbers
✗ Banking details, account numbers, or credentials
✗ Employee names attached to compensation
✗ Customer or supplier lists
✗ Anything covered by an NDA or engagement letter confidentiality clause
DO INSTEAD
✓ Anonymize before pasting — “Client A,” “Vendor 1,” “the company”
✓ Scale figures proportionally if the absolute numbers are sensitive
✓ Use business or enterprise tiers with contractual data terms
✓ Turn off training on your inputs where the setting exists
✓ Keep a written policy so your team isn’t deciding case by case
If you’re a professional handling client information, this isn’t optional caution — it sits inside your confidentiality obligations. Worth confirming the specifics with your professional body, since expectations differ by province and are still developing.
Every prompt below is written to work on anonymized data.
Why Most Finance Prompts Fail
A weak prompt asks a question. A strong one supplies four things first:
Better prompts won’t fix weak underlying numbers
If your close runs late or your books need review before anyone can analyze them, that’s the constraint worth fixing first. Book a free 30-minute call for an honest assessment.
Nine Prompts Worth Keeping
1. Variance commentary from a trial balance
Turns half a day of writing into a first draft in minutes.
You are a financial controller preparing month-end commentary for an owner-managed business. Below are current month actuals, prior month actuals, and budget by account.
For each account with a variance over 10% or $5,000 (whichever is greater), write one sentence stating the variance and the most likely driver based on the account’s behaviour over the periods shown.
Rank by dollar impact. Maximum 8 items. Plain language, no accounting jargon.
At the end, list separately any variance you cannot explain from this data alone and state what information would be needed.
[paste anonymized trial balance]
Where it goes wrong: it will speculate about causes it cannot know. The final instruction is what keeps that speculation in a labelled box instead of woven through your commentary. Always read that section first.
2. Stress-testing your own forecast assumptions
The most valuable prompt on this list, because it’s the one nobody in your organization is incentivized to ask.
You are a skeptical lender’s credit analyst reviewing the forecast below. Your job is to find the weakest assumptions, not to be encouraging.
Identify: (1) assumptions that appear unsupported by the historical data shown, (2) any place where growth is projected without a corresponding cost increase, (3) working capital implications that appear unmodelled, (4) the single assumption that, if wrong by 20%, would do the most damage.
For each, state the specific concern and what evidence would resolve it. Be direct.
[paste anonymized forecast with 24 months of history]
Where it goes wrong: it can only see what you paste. If the growth assumption rests on a signed contract you didn’t include, it will flag a real assumption as unsupported. That’s still useful — it tells you what a lender will ask about.
3. Board narrative from a reporting pack
Write a one-page board update from the financials below.
Structure: (1) headline — the single most important thing that happened this quarter, in one sentence, (2) three things that went well, (3) three things that need attention, (4) two decisions the board should be prepared to discuss.
Audience is non-financial directors. No jargon. Every claim must trace to a number in the data. Do not editorialize about strategy.
[paste anonymized quarterly figures]
Where it goes wrong: it will pick the largest number as the headline, which isn’t always the most important one. You know which story matters this quarter; the draft saves you the writing, not the judgment.
4. Interrogating a contract for revenue recognition implications
Read the contract terms below and extract everything relevant to revenue recognition timing.
Identify: performance obligations and when each is satisfied, payment timing versus delivery timing, any variable consideration or contingencies, termination and refund provisions, anything creating a financing component.
Present as a table. Flag any clause whose accounting treatment is genuinely ambiguous rather than guessing at it.
[paste anonymized contract terms]
Where it goes wrong: this is extraction, not conclusion. It surfaces the clauses that matter so you spend your review time on the hard parts. The actual treatment decision is yours, and for anything material it belongs with your accountant.
5. Explaining a variance to a non-financial manager
Rewrite the finance commentary below for an operations manager with no accounting background.
Rules: no accounting terms without a plain-language gloss, lead with what it means for their decisions rather than the number itself, maximum 150 words, end with the one action they should take.
Keep every figure accurate. Do not soften bad news.
[paste your own commentary]
Where it goes wrong: simplification sometimes loses a necessary caveat. Check that nothing material got smoothed away in the translation.
6. Building a close checklist from your actual process
I will describe how our month-end close currently works, informally and out of order.
Convert it into a structured checklist: sequence tasks by dependency, assign a target business day to each, identify which tasks could run in parallel, and flag any step where a review or approval appears to be missing.
Then list what you’d expect to see in a close process of this type that I haven’t mentioned.
[describe your process]
Where it goes wrong: the final instruction is the valuable half — it surfaces gaps rather than just tidying what you already do. Cross-check the output against your actual controls before adopting it. More on this in our guide to month-end close optimization.
7. Sanity-checking a spreadsheet model
Below is the structure and formula logic of a financial model. Review it as an auditor would.
Identify: circular references or logic that could produce one, hard-coded values sitting where a formula should be, inconsistent formulas across a row that should be uniform, sign convention errors, and any calculation whose result wouldn’t reconcile to a related figure.
List issues by severity. For each, state what would go wrong and how to test for it.
[paste formula logic and structure, not client data]
Where it goes wrong: it reasons about logic you describe, not a live file. It won’t catch a broken reference. Useful as a design review, not a substitute for testing the model with known inputs.
8. Preparing for a lender conversation
Based on the financial summary below, generate the fifteen hardest questions a commercial lender would ask in a credit review.
Prioritize questions about debt service capacity, revenue concentration, working capital, and anything in the numbers that looks inconsistent.
For each question, note what a weak answer would sound like — so I can tell whether mine is one.
[paste anonymized summary financials]
Where it goes wrong: it generates plausible questions, not your specific lender’s questions. Treat it as rehearsal, and pair it with proper cash flow projections before the meeting.
9. The devil’s advocate on a decision you’ve already made
I am about to make the following financial decision, and I have already decided in favour of it. Argue against it.
Give me the three strongest objections, the financial conditions under which this becomes a serious mistake, and the leading indicator I should watch that would tell me early if it’s going wrong.
Do not balance this with reasons it might work. I have those.
[describe the decision and the relevant numbers]
Where it goes wrong: forced disagreement sometimes manufactures weak objections. Discard those. The value is in the one or two that make you uncomfortable — and in the leading indicator, which is often the most practical thing it produces.
constraint — every time
from the data provided
data into a consumer tool
Want a second opinion on what these prompts tell you?
Bring the output of prompt 2 or prompt 8 to a free 30-minute call. We’ll tell you which flags are real and which are artifacts of incomplete data.
Four Things No Prompt Will Fix
- ✗Tax positions. These carry professional accountability and consequences for being wrong. Research is fine; conclusions belong with a qualified advisor.
- ✗Anything you’ll sign. Financial statements, filings, and lender submissions carry your name. A draft is a draft until a qualified person has verified every figure in it.
- ✗Bad underlying data. A well-prompted analysis of unreconciled books produces an articulate, well-structured wrong answer.
- ✗Knowing what’s missing. These tools reason about what you give them. Recognizing that an entire category of cost never made it into the file is judgment, not analysis — the failure modes are covered in our piece on where AI in accounting falls short.
Frequently Asked Questions
The Bottom Line
The finance people getting real value from these tools aren’t using better software than everyone else. They’re asking better questions, on cleaner data, and checking the answers.
- ✓Start with prompt 2: nobody in your organization is paid to attack your own forecast
- ✓Add the constraint every time: “tell me what you can’t determine from this data”
- ✓Fix the data before the prompts: good questions on bad numbers still give bad answers
Good prompts, better numbers
These tools work best on books that close on time and reconcile properly. Book a free 30-minute call and we’ll tell you honestly where yours stand.