W William Kim / portfolio ← Back to portfolio
Workshop · Accounting team training

Prompting Copilot for Accounting Work

Reconciliations, exception handling, and making sense of large data. Sixty minutes, two live demos.

§ 01 · The one idea

Copilot is not the reconciliation engine.

The tools that find differences and the tool that explains them are not the same tool. Everything else on this page follows from that split.

01 · Deterministic tools FIND
Same input, same answer, provable totals.
  • Power Query merges
  • INDEX/MATCH
  • SUMIFS
  • Pivot tables
  • Conditional formatting
02 · Copilot EXPLAINS
Judgment, context, and language.
  • Classifying an exception list: timing, coding, accrual, credit
  • Searching email and Teams for why an item is open
  • Writing the matching logic you then run yourself
  • Drafting the narrative and the follow-up to the item owner
  • Reviewing a finished rec for what is unexplained or unsupported

Never ask Copilot to be the matching engine. "It found 12 unmatched items when there were 15" is an audit finding, not a productivity win.

§ 02 · The framework

Structure beats cleverness: R-STAR-QC

You will not use all seven every time. Role, Task, Result, and Constraints carry most of the weight.

R
Role
Who should it act as? A staff accountant reviewing a sales tax rec.
S
Situation
What is the context? Month-end close, period ending 7/31, two extracts.
T
Task
What is the single deliverable? Categorize these 6 variances.
A
Action
What steps, in what order? Group by state, then by root cause.
R
Result
What format? A table with columns State, Amount, Category, Next step.
Q
Question
What should it ask you? List anything you need from me to be certain.
C
Constraints
What are the rules? Do not estimate. Flag anything you cannot source.
The 20 second version

Act as [role]. Here is [context]. Do [one task]. Return [exact format]. Do not [the thing that would ruin it].

Source: Adebayo, Akpan & Dell, "Prompt Engineering for Accountants: Automating Financial Tasks with AI," CPA Practice Advisor, September 2025.

§ 03 · Prompt library

The five prompts from the live demos.

Every prompt ends with a boundary, and that sentence is the accuracy control, not politeness. Copy them, then change the period, the account, and the categories to match your own work.

01 Classify a variance list

You are a staff accountant reviewing a sales tax reconciliation for the period ending 7/31/2026. Below are the state-level variances I calculated in Excel from two extracts: the GL sales tax payable detail and the tax engine export. For each variance, assign one root-cause category from this list: timing, coding error, missing accrual, unrecorded credit, rate error, jurisdiction mapping. Return a table with columns: State, Variance, Most likely category, Second possibility, What I should pull to confirm. Do not calculate or invent amounts. Use only the variances I pasted. If a variance could fit more than one category, say so explicitly.

02 Investigate an aged item with Work IQ

Search my email, Teams messages, and files from the last 90 days for anything about Minnesota sales tax, marketplace facilitator rules, or the Kestrel Nutrition account. Summarize what you find in three bullets, name who said it and when, and link each source. If you find nothing relevant, say so rather than guessing.

03 Identify trends from a pivot

Below is corporate card spend by category by month for May, June, and July 2026. Identify the three categories with the largest change and state the direction and dollar amount of each. Then tell me which changes look like normal seasonality versus something worth asking about, and explain your reasoning. Use only the figures I pasted.

04 Flag outliers with a stated definition

Here is spend by cardholder by month. Flag any cardholder whose monthly spend in any month is more than double their own average for the other two months. Also flag any category that appears for only one cardholder. For each flag, give the cardholder, the month, the amount, and one sentence on what I should check. Do not flag anything that does not meet those two definitions.

05 Draft the summary email

Draft an email to the Controller summarizing the four verified findings below. Under 200 words. Professional and factual, not alarmed. Lead with the total dollar impact, then the findings as a short list, then one clear ask. Do not add findings that are not in my list.

§ 04 · Guardrails

Eight ways people get burned.

None of these are exotic. Every one of them shows up in the first week of using the tool on real work.

01 Asking it to do the math

Treating Copilot like a calculator over a large range.

FixCompute in Excel. Hand Copilot the result to interpret.

02 Leaving the format open

"Summarize this" returns whatever shape it feels like.

FixName the columns, the length, and the audience.

03 Accepting the first answer

The first draft is a starting point, not a deliverable.

FixPush back once: "What did you assume? What is weakest here?"

04 Dumping in a whole workbook

Context gets truncated and it quietly drops rows.

FixSend the summary tab or the exception list, not all 4,000 rows.

05 Asking a leading question

"Confirm this variance is timing" gets you agreement, not analysis.

FixAsk neutrally: "What are the possible causes, ranked by likelihood?"

06 Starting a new chat every time

You re-explain context and lose the thread's accumulated setup.

FixKeep one chat per rec. Add turns instead of restarting.

07 Trusting a cited source blindly

It can attach a real-looking citation to a claim the source never made.

FixOpen the link. Every time you plan to rely on it.

08 Pasting confidential data outside Copilot

Your organization's data protection agreement covers Microsoft 365 Copilot. It does not cover a personal ChatGPT account.

FixCompany financial data stays in M365 Copilot. No exceptions.
The four non-negotiables
  1. Every number that reaches a workpaper ties back to a control total before anyone signs it. Copilot drafts. The accountant ties out and signs.
  2. Every categorized exception needs a source document, not just a plausible story.
  3. If Copilot names a dollar amount you did not give it, that is a red flag, not a finding.
  4. Note in the workpaper where AI assisted. Reviewers should know what to scrutinize.
The data rule

Company financial data stays in M365 Copilot.

No exceptions, and no personal accounts.
§ 05 · Downloads

Everything from the session.

The prompt library above is the whole of it in text form. These are the files.

The slides All 19, including the demo debrief and what the tool misses.
Coming soon
Demo 1 · Sales tax reconciliation 100 rows across two systems, no shared key. Fictional data.
Coming soon
Demo 2 · Corporate card spend 258 transactions across three months. Fictional data.
Coming soon
One page guardrail card The split, the non-negotiables, the data rule, and the template. Prints to one Letter page.
HTML · 12 KB · Open ↗