AI + Excel Exercise Pack

Tuesday, October 27, 2026 · Facilitated by Adam Murphy, ImpactMe.AI

Before you start: a safety reminder

Never paste real client, employee, or company data into a consumer AI account (a free or personal ChatGPT, Claude, Gemini, or Copilot login). Those accounts may keep what you type, and your firm has no control over it.

All data in these exercises is fictional. Ridgeline Electric, every general contractor, supplier, job, invoice, and dollar amount was invented for practice. Paste it anywhere you like today.

Back at the office, follow your firm's policy and check with IT about which AI tools and accounts are approved for real data.

How these exercises work

You'll use six practice workbooks for a fictional electrical contractor, Ridgeline Electric, Inc. Each one is small enough to paste into a free chat AI.

#	Workbook	What you'll practice	Time
01	`01_Messy_GL_Export.xlsx`	Cleaning up a messy job-cost export	20 to 25 min
02	`02_Bank_Reconciliation.xlsx` (plus `02_Bank_Statement_September.pdf`)	Finding reconciling items	15 to 20 min
03	`03_Budget_vs_Actual.xlsx`	Drafting variance commentary	15 min
04	`04_AR_Aging.xlsx`	Pivot planning and aging buckets	20 to 25 min
05	`05_WIP_Schedule.xlsx`	Having AI explain a WIP schedule (and catching what's wrong)	20 min
06	`06_Code_Split.xlsx`	Splitting one cost code into two, and proving nothing broke	20 min

Three habits run through every exercise:

Ask for a plan before you ask for answers. "How would you approach this?" gets better results than "do this."
Run the same prompt in two or three tools and compare. Different tools catch different things. When they disagree, that's your cue to look closer.
Verify with Excel, every time. AI writes answers. Excel computes them. A total that ties out is proof; a confident paragraph is not.

How to get data into the AI

Copy and paste (works everywhere): Select the cells in Excel (include the header row), press Ctrl+C, click in the chat box, press Ctrl+V. The AI receives the data as text, with columns separated by tabs.
Upload the file (if your tool allows it): Most chat tools have a paperclip or + button for attaching files. Free plans may limit uploads.
Copilot in Excel (if your firm has it): Open the workbook, click a cell inside the data, and press Ctrl+T to make it a table first. Copilot works best with tables and with files saved to OneDrive or SharePoint.

Heads up: When you copy cells, Excel pastes the values, not the formulas. If you want the AI to review formulas, see Exercise 05 for how to show them.

The 5-part prompt pattern (quick reminder)

Role: "You are a construction accountant..."
Context: What the data is and what you already know.
Task and format: What you want back, and its shape (table, bullets, word count).
Constraints: What to avoid or not change.
Check yourself: "List your assumptions. Flag anything you're unsure of."

Compare-across-tools scorecard

Use this for any exercise where you run the same prompt in more than one tool. A quick check mark or a 1 to 3 score is plenty.

Question	Tool A: ______	Tool B: ______	Tool C: ______
Did it get the facts right? (checked in Excel)
Did it catch the tricky items?
Did it invent anything?
Did it ask a good clarifying question or flag an assumption?
Was it clear enough to hand to a coworker?
Which would you use for this task next month?

Exercise 01: Clean Up a Messy Job-Cost Export

Goal: Turn a messy GL export into clean, pivot-ready data, with AI planning the work and Excel proving it's right.

Time box: 20 to 25 minutes Workbook: 01_Messy_GL_Export.xlsx (sheet GL_Export)

The situation

It's October 2. You exported September job cost detail for four jobs so you can build a pivot of cost by job and cost code. The export came out a mess: dates in five or six formats, job numbers typed every which way, dollar amounts stored as text, a subtotal row in the middle, blank rows, repeated headers, and some rows that appear twice.

Steps

Look before you ask (2 min). Open GL_Export and scroll through rows 1 to 52. Jot down every problem you notice. You'll compare your list to the AI's.
Take a "before" reading. Click an empty cell (try J2) and type =SUM(H5:H52). Write down the result. Is that the real September job cost? (Hint: Excel's SUM skips any number stored as text, without warning you.)
Ask for a cleanup plan, not cleaned data. Select A4:H52, copy, and paste into your AI tool along with the strong prompt below. Read the plan. Did it find everything on your list? Did it find things you missed?
Ask how to do it in Excel. Follow up with: "Give me Excel formulas I can put in helper columns to standardize the job number, cost code, and amount. Assume the data starts in row 5. Explain each formula in one sentence." Put the formulas in columns J, K, L and fill them down.
Decide on the duplicates yourself. The AI can point out rows that look the same. You decide which ones are true duplicates. Look closely at the reference numbers and job numbers before deleting anything.
Optional, faster path: Ask the AI to give you the fully cleaned table back (40 rows, tab-separated) and paste it into a new sheet. This is quicker, but now every number came from the AI, so step 8 matters even more.
Compare across tools. Paste the same data and the same plan prompt into a second tool. Use the scorecard. Did both catch the subtotal row? The two dates written day-first (like 11.09.2026)? The odd five-digit number in the date column?
Verify it (see below).

Starter prompts

Weak prompt:

Clean up this data.

Why it's weak: The AI doesn't know what "clean" means to you, what the data is, or what it's for. It will guess, and it may quietly delete or change rows.

Strong prompt:

You are a construction accountant helping me prepare data for a pivot table. Below is a September 2026 job cost detail export from our accounting system for an electrical contractor. It has formatting problems.

First, do NOT clean the data yet. Give me a numbered cleanup plan that lists every problem you find (dates, job numbers, cost codes, amounts, extra rows, possible duplicates), the row numbers where each problem appears (the first data row is row 5), and how you'd fix it.

Context: every transaction is dated in September 2026. Cost codes are 100 Labor, 150 Labor Burden, 200 Material, 300 Equipment, 400 Subcontract, 500 Job Overhead. Negative amounts may appear in parentheses.

For anything that might be a duplicate, explain why, and tell me what would make it NOT a duplicate. List any assumptions you're making.

[paste A4:H52 here]

Follow-up prompt for formulas:

Give me Excel formulas (Excel 2016 compatible, no newer functions) for helper columns that convert the Job, Cost Code, and Amount columns into clean numbers. Data starts in row 5. Explain each in one sentence and tell me which rows your formulas will NOT handle.

Verify it

Do at least three of these. They take Excel, not trust.

The export checks itself. Row 20 is the system's own "Subtotal Job 1042." After cleanup, the total for Job 1042 must equal that subtotal exactly.
Row count. The export has 48 rows below the header (rows 5 to 52). Count what you removed (blank rows, subtotal, repeated header, duplicates). Removed rows plus cleaned rows must equal 48.
Totals tie. Once amounts are numbers, total the original column with your helper formula. Then subtract the subtotal row and every duplicate you removed. The result must equal the total of your cleaned data, to the penny.
Dates stay in September. On your cleaned date column, =MIN(...) should be 9/1/2026 or later and =MAX(...) should be 9/30/2026 or earlier. Anything outside September was misread.
Hand-check one row. Pick a row with a text amount like $18,975.00 and confirm your cleaned version reads 18975.

What this teaches

AI is excellent at spotting messy-data patterns and writing the formulas to fix them.
"Clean this up" without context invites silent damage: dropped rows, misread dates, legitimate entries deleted as "duplicates."
Formulas you run in Excel are repeatable and checkable. Data rewritten by an AI must be checked line by line.
SUM ignoring text amounts is a real month-end trap, with or without AI.

Exercise 02: Bank Reconciliation Support

Goal: Use AI to suggest matches and find reconciling items, then prove the reconciliation ties to zero in Excel.

Time box: 15 to 20 minutes Workbook: 02_Bank_Reconciliation.xlsx (sheets Bank_Statement and Book_Cash_Ledger)

The situation

Ridgeline's September bank statement and cash ledger don't agree. August reconciled cleanly with no open items, so both start at $186,420.55. Your job is to find every item that explains the difference.

The bank statement also comes as a PDF (02_Bank_Statement_September.pdf), the way it actually arrives from the bank. Same numbers as the Bank_Statement sheet.

Steps

Find the gap. Note the bank ending balance (Bank_Statement, Balance column, last row) and the book ending balance (Book_Cash_Ledger, last row). In an empty cell, subtract them. That's the difference you have to explain.
Give the AI both sides. Copy Bank_Statement rows 4 through the totals row, paste it into your AI tool, then do the same for Book_Cash_Ledger. Label each paste ("BANK STATEMENT:" and "BOOK CASH LEDGER:").
- Or upload the PDF (if your tool takes files): attach 02_Bank_Statement_September.pdf instead of pasting the bank sheet. First ask: "Read this statement. Give me the opening balance, total deposits, total withdrawals, and closing balance, and check that opening plus deposits minus withdrawals equals closing." Compare its numbers to the PDF yourself before you go on. If the AI misread one line, everything after it is wrong.
Ask for matching and a list of reconciling items using the strong prompt below.
Push back on anything vague. If the AI says "there may be a timing difference," ask: "Which line, which amount, and which side of the rec does it go on?"
Ask about the two $20 charges. On 9/16 there are two $20.00 Spokane Office Supply charges on both the bank and the books. Use the last follow-up prompt below. What does the AI say, and how could it know?
Compare across tools. Run the same prompt in a second tool. Count how many reconciling items each one found, and whether any "found" a problem that isn't one.
Build the rec in Excel and verify it (see below).

Starter prompts

Weak prompt:

Why don't these match?

Strong prompt:

You are a staff accountant at an electrical contractor doing the September 2026 bank reconciliation. Below are the bank statement and our book cash ledger for the same account. Both started September at $186,420.55 and August had no open items.

Match each bank line to a book line (checks by check number, deposits by date and amount, ACH and EFT by description and amount).

List every unmatched or mismatched item in a table with these columns: item, date, amount, bank side or book side, and type (outstanding check, deposit in transit, bank fee, recording error, duplicate, other).

Show the reconciliation: adjusted bank balance and adjusted book balance. They should be equal.

For anything you're unsure of, say so. Do not assume an entry is a duplicate just because the amount repeats; tell me what evidence you used.

[paste bank statement] [paste book cash ledger]

Useful follow-ups:

"Check 10479 is different on each side. Is this a transposition? How can I tell?"
"Draft the journal entries needed to correct the books, with a one-line explanation for each."
"There are two $20.00 Spokane Office Supply charges on 9/16. In our books one is office supplies and one is Job 1042 materials. Which bank line goes with which? How sure are you, and what is that based on?"

Verify it

Opening + In minus Out = Closing. For any statement an AI read from a PDF, check this one line first: opening balance plus total deposits minus total withdrawals must equal the closing balance printed on the statement. If it doesn't, the AI misread something.
Build the rec yourself in Excel. Bank ending balance, plus deposits in transit, minus outstanding checks, equals adjusted bank. Book ending balance, minus bank fees, plus or minus errors, equals adjusted book. The difference must be exactly zero. If it isn't, something is missing or wrong, no matter how confident the AI sounded.
Confirm each outstanding check. Use Ctrl+F on Bank_Statement to search for each check number the AI lists. It should not be there.
The divide-by-9 test. If a difference between two numbers divides evenly by 9, it's very often a transposition (two digits swapped). Try it on any mismatch the AI flags.
Sum check. Add up all the AI's reconciling items by side. They should account for the full difference you calculated in step 1.

What this teaches

AI is fast at proposing matches across two lists, and good at explaining why something doesn't match.
Recurring payments of the same amount to the same vendor are not automatically duplicates. AI tools sometimes say they are.
AI can match the amount, but not the intent. Two identical $20.00 bank lines can't tell you which was the office purchase and which was the job purchase. For the cash rec it doesn't matter; for job cost it does. That answer comes from the receipt or the person who bought it, never from the bank data.
The reconciliation is only done when Excel says the difference is zero. AI can't sign off on a rec; you can.

Exercise 03: Variance Commentary a PM Can Read

Goal: Have AI draft plain-language budget vs actual commentary, then check every number it quotes.

Time box: 15 minutes Workbook: 03_Budget_vs_Actual.xlsx (sheets BvA_Detail, Job_Summary, PM_Notes)

The situation

You need monthly variance commentary on three jobs for the owner and project managers. Your firm's rule: explain any monthly variance that is at least 10% AND at least $2,500, in either direction. Variance = Actual minus Budget, so a positive number means over budget. The PMs sent you notes on some of the big ones (sheet PM_Notes).

Steps

Find the material variances yourself first (3 min). On BvA_Detail, in cell J2, enter =AND(ABS(H2)>=2500,ABS(I2)>=0.1) and fill it down to row 55. Then filter column J for TRUE. Count the rows. Keep that number to yourself for now.
Paste the detail and the PM notes into the AI. Copy BvA_Detail A1:I55, then PM_Notes A1:D6. Use the strong prompt below.
Compare the AI's list to yours. Did it find the same variances? Did it miss any or include ones that don't meet the rule?
Read the commentary like the owner would. Is it clear? Is it accurate? Did it connect the pieces (for example, a variance in August that reverses in September)?
Compare across tools. Run the same prompt in another tool. Look especially for any variance where the PM didn't send a note. What did each tool say about it?
Verify every number (see below).

Starter prompts

Weak prompt:

Write variance commentary for this.

Strong prompt:

You are the controller at an electrical contractor writing monthly cost variance commentary for the owner and project managers. Below is budget vs actual by job, month, and cost code for July to September 2026, followed by notes from the project managers.

Rules:

Variance = Actual minus Budget. Positive means over budget.

Only comment on monthly variances that are at least 10% AND at least $2,500 (either direction).

Group commentary by job. For each variance, give the month, cost code, dollar and percent variance, and a one or two sentence explanation.

Use ONLY the PM notes to explain causes. If there is no note for a variance, write "Cause not yet confirmed. Follow up with PM." Do not guess.

Point out where a variance in one month reverses in another (timing) versus where it's a true overrun.

Plain language, under 300 words total.

[paste BvA_Detail] [paste PM_Notes]

Verify it

Count check. The number of variances the AI commented on should match the count from your filter in step 1.
Every number traces to a cell. For each dollar amount or percent in the commentary, find it on BvA_Detail. Highlight each one as you confirm it. Any number you can't find is a red flag.
Quarter-to-date check. For any "timing" claim, look at Job_Summary. If August and September offset, the quarter-to-date variance for that cost code should be small.
Tie-out. At the bottom of Job_Summary, the tie-out cell should read 0. That's a built-in check that the summary matches the detail.

What this teaches

AI is very good at turning a table of numbers into readable prose. This saves real time every month.
It's also very good at making up a plausible reason when it doesn't have one. Telling it "don't guess" (and checking that it listened) is essential.
The person who signs the commentary owns every number in it. Verify before it goes to the owner.

Exercise 04: AR Aging and Pivot Planning

Goal: Use AI to decide what summary view answers a business question, then build it in Excel with an aging-bucket helper column.

Time box: 20 to 25 minutes Workbook: 04_AR_Aging.xlsx (sheet AR_Detail)

The situation

The owner asks: "Which customers are slow pays, how much do they owe us by age, and how much retainage is tied up with each one?" You have invoice-level AR detail as of September 30, 2026: progress billings (pay apps), T&M tickets, and one retainage release. You know a pivot table is the answer, but which pivot?

Steps

Read the columns. Look at the Read_Me sheet and the headers on AR_Detail. Notice that Retainage_Held is separate from Open_Balance. Retainage isn't due until the job closes out.
Ask the AI to plan the view, not build it. Copy AR_Detail A1:N41 and use the strong prompt below. You want back: which helper columns to add, what goes in Rows, Columns, and Values, and why.
Add the helper columns. Ask the AI for the formulas if it didn't already give them. You'll likely need Days_Past_Due (as-of date minus Due_Date) and Aging_Bucket (Current, 1 to 30, 31 to 60, 61 to 90, Over 90). Put the as-of date 9/30/2026 in its own cell (for example, Q1) and point the formula at it.
Build the pivot. Click inside the data, then Insert > PivotTable. Customers in Rows, Aging_Bucket in Columns, Sum of Open_Balance in Values. Then add Sum of Retainage_Held as a second value (or build a second small pivot for it).
Answer the owner's question in three sentences. Ask the AI to draft it from your pivot results (paste the pivot, not the raw data).
Compare across tools. Give the same planning prompt to a second tool. Did both age by due date or by invoice date? Did either one count retainage as past due? Which plan would you trust?
Verify it (see below).

Starter prompts

Weak prompt:

Summarize this aging report.

Strong prompt:

You are a construction accountant. Below is invoice-level accounts receivable for an electrical contractor as of 9/30/2026. Customers are general contractors. Each row is one invoice. Retainage_Held is withheld by the GC until project closeout and is not due yet. Open_Balance is what's currently owed on the invoice.

My owner wants to know: which customers are slow pays, how much they owe by age, and how much retainage is tied up with each one.

Don't build anything yet. Tell me:

What helper columns I should add (with Excel formulas, Excel 2016 compatible).

Exactly how to set up the pivot table: Rows, Columns, Values, and any filters.

Whether to age by invoice date or due date for this question, and why.

What else in the data might signal a slow pay besides age (for example, partial payments).

[paste A1:N41]

Follow-up prompt:

Here are my pivot results. Draft a three-sentence answer for the owner. Use only these numbers. Name the two customers we should call first and why.

[paste your pivot]

Verify it

Grand total ties. In an empty cell, type =SUM(N2:N41). The pivot's grand total for Open_Balance must match it exactly.
Retainage ties. =SUM(K2:K41) must match the pivot's total retainage.
Hand-check one invoice. Pick one past-due invoice. Count the days from its Due_Date to 9/30/2026 on a calendar. Does your Days_Past_Due match? Is it in the right bucket?
Paid invoices aren't aging. Filter Open_Balance for 0. None of those rows should show up in an aging bucket with a dollar amount.
Check the AI's summary. Every number in the owner summary should appear in your pivot.

What this teaches

"Pivot planning" is where AI shines: going from a business question to the right summary view.
Construction AR has quirks (retainage, pay apps, GC payment terms) that a general-purpose AI won't assume unless you tell it.
The pivot does the math. The AI helps you decide what math to do.

Exercise 05: Have AI Explain a WIP Schedule (and Catch What's Wrong)

Goal: Use AI to explain a WIP schedule in plain language, then build the habit of checking the spreadsheet itself, not just the explanation.

Time box: 20 minutes Workbook: 05_WIP_Schedule.xlsx (sheet WIP_Schedule)

The situation

Ridgeline's September 30 WIP schedule covers eight open jobs using the cost-to-cost method. A new project manager asks you to walk them through it. One thing on this sheet is not right. Can the AI find it? Can you?

Steps

Get the explanation. Copy WIP_Schedule A4:P13 and paste it into your AI tool with the first strong prompt below. Read the explanation. Would a PM understand it?
Ask a follow-up a PM would ask. For example: "Job 1042 shows $38,000 overbilled. Is that bad? What usually causes it?"
Now show the AI the formulas. In Excel, go to Formulas > Show Formulas (or press Ctrl+`, the key left of the 1). The cells now show formulas instead of results. Copy A4:P13 again and paste it with the second prompt below. Turn Show Formulas back off when you're done.
Compare across tools. Give both pastes to a second tool. Did each one find the problem? Did either one "find" a problem that isn't one?
Recompute one row by hand (see Verify).
Talk it through. If the problem is real, what does it do to this job's over/under billing? To the balance sheet total?

Starter prompts

Weak prompt:

What is this?

Strong prompt (explain it):

You are a construction controller explaining a WIP (work in progress) schedule to a new project manager who is not an accountant. Below is our 9/30/2026 WIP schedule for eight jobs, using the cost-to-cost percent complete method.

Explain what each column means in one plain sentence each.

Explain over billing and under billing, and why each matters for cash and for the balance sheet.

Pick two jobs, one overbilled and one underbilled, and walk through the math step by step using the actual numbers.

Keep it under 400 words. Avoid jargon or define it.

[paste A4:P13 values]

Strong prompt (audit it):

Below is the same WIP schedule with formulas showing instead of values. Check every formula column for consistency across all eight rows. List any row where a formula is different from the others, explain what it does differently, and tell me the effect on that job's earned revenue and over/under billing. If everything is consistent, say so. Do not guess.

[paste A4:P13 with formulas showing]

Verify it

Recompute one row by hand. Pick Job 1047. With a calculator: Pct_Complete = Cost_to_Date ÷ Est_Total_Cost. Earned_Revenue = Revised_Contract × Pct_Complete. Over or under = Billed_to_Date minus Earned_Revenue. Does it match the sheet?
Add a check column. In Q5, type =F5*K5-L5 and fill down to Q12. Every row should be 0, because Earned_Revenue should always be Revised_Contract times Pct_Complete. Any row that isn't 0 needs a look.
Look for Excel's own warning. A small green triangle in the corner of a cell can mean "this formula is different from its neighbors." Click the cell to see the message.
Over/under logic check. On each row, only one of Over_Billing or Under_Billing should have a number. Both can't be positive.

What this teaches

AI is a patient, clear explainer of financial schedules. Great for training PMs and new staff.
When you paste values, the AI can only explain the numbers it sees. It can't check formulas it never saw.
The point isn't whether the AI found the error. The point is that you checked. A one-column formula check takes 30 seconds and works every month.

Exercise 06: Split a Cost Code (and Prove Nothing Broke)

Goal: Use AI to propose the rules for splitting one cost code into two, apply them with a lookup in Excel, and then run the checks that prove the split didn't lose, double-count, or rewrite anything.

Time box: 20 minutes Workbook: 06_Code_Split.xlsx (sheets Items, Transactions, Mapping, GL_Control)

The situation

Copper prices keep moving, and the owner wants to see copper and aluminum wire costs separately. Starting with the October 2026 period, Ridgeline is splitting material cost code 26-100 Wire and Cable into:

26-110 Copper Wire and Cable
26-120 Aluminum Wire and Cable

You need to recode the 40 items on the items list and every open October job cost transaction (60 rows across six jobs, September and October). The controller's rules are on the Read_Me sheet: classify by the metal of the conductors (not the armor), recode only the open October period, mark anything unclear REVIEW, and don't change the total.

The split itself is easy. The checks are the lesson.

Steps

Read the rules (2 min). Open Read_Me, then skim Items. Find two items you'd have trouble classifying yourself.
Ask the AI for rules, not answers. Copy Items A1:F41 and paste it with the strong prompt below. You want back a short set of rules, plus the items the rules can't handle.
Push back. Ask: "Which items would your rules send to both codes? Which to neither? Which ones contain the letters AL or CU but aren't about the conductor?"
Build the mapping table. Use the follow-up prompt to have the AI apply its rules and return a tab-separated table. Paste it into Mapping starting at A5 (over the example row). Read every REVIEW row. Change any you disagree with.
Apply it with XLOOKUP. In Transactions cell O2, enter the formula below and fill down to O61. Do the same in Items G2 (fill down to G41) so the items list gets its new codes too.
Run the checks on GL_Control. Every line in the "After the split" block must read OK. When one doesn't, figure out why before you touch anything else.
Compare across tools. Give the same rules prompt to a second tool. Did both catch the "AL ARMOR" item? Did either one try to recode September?

The formulas:

Transactions!O2 (Microsoft 365 or Excel 2021): =IF(D2="Closed","26-100",XLOOKUP(G2,Mapping!$A$5:$A$60,Mapping!$D$5:$D$60))
Same thing for Excel 2016 or 2019: =IF(D2="Closed","26-100",INDEX(Mapping!$D$5:$D$60,MATCH(G2,Mapping!$A$5:$A$60,0)))
Items!G2: =XLOOKUP(A2,Mapping!$A$5:$A$60,Mapping!$D$5:$D$60)

Don't wrap these in IFERROR yet. An #N/A means an item isn't on your mapping table. That's information you want to see. Hiding it with IFERROR (or worse, defaulting it to "26-110") is how items quietly land in the wrong code.

Starter prompts

Weak prompt:

Split these into copper and aluminum.

Why it's weak: the AI will classify every item, guess on the unclear ones, and not tell you which ones it guessed. It will probably send "MC CABLE AL ARMOR" to aluminum.

Strong prompt:

You are helping a construction accountant split one job cost code into two. Material code 26-100 Wire and Cable is being replaced by 26-110 Copper Wire and Cable and 26-120 Aluminum Wire and Cable. Below is the list of 40 items currently on 26-100.

Don't map the items yet. First:

Propose a short set of mapping rules based on the descriptions. Classify by the metal of the conductors, not the armor or jacket.

List every item your rules would send to BOTH codes, and every item they would send to NEITHER.

List every item where the conductor metal isn't stated, or that isn't wire at all. Those get REVIEW. A human decides. Do not guess.

Tell me what I should check after applying the mapping to prove nothing was lost, double-counted, or changed in a closed period.

[paste Items A1:F41]

Follow-up prompt (build the mapping table):

Apply your rules. Give me a tab-separated table with exactly these columns: Item_ID, Description, Old_Cost_Code, New_Cost_Code, Rule_Used, Flag_Reason. One row per item, exactly 40 rows, no header notes. New_Cost_Code must be exactly 26-110, 26-120, or REVIEW. Then tell me how many rows you gave me.

Verify it (this is the lesson)

Every one of these is on GL_Control. They take Excel, not trust.

Row count unchanged. Transactions still has 60 rows. Nothing was deleted or duplicated while recoding.
Total dollars unchanged, and still tied to the GL. October on 26-110 plus 26-120 plus REVIEW must equal the October GL balance for the old code, to the penny.
No item left unmapped. Zero blank New_Cost_Code cells, zero #N/A lookups, and every item on Items has a row on Mapping. (Watch the transactions, not just the items list. Something on a transaction may not be on the list at all.)
No item mapped twice. Every Item_ID appears on Mapping exactly once. A lookup quietly uses the first match it finds, so a duplicate row can put an item in the wrong code and the totals will still tie. This is the check people skip.
Ambiguous items flagged for a human. The REVIEW dollars are real money waiting on a decision. Someone has to own that list and clear it before the October close.
History and prior periods left alone. September is closed. Its dollars must still sit on 26-100 and still tie to the September GL. Recoding a closed period changes financials that were already reported.

What this teaches

AI is good at turning a pile of descriptions into a clear rule set, and at spotting the items that break the rules, if you ask it to.
Keyword rules fail in predictable ways: "AL" in "AL ARMOR," both metals in one description, items that aren't wire at all.
A recode is a data migration. Row count, total dollars, GL tie, unmapped, mapped twice, and closed periods are the same six checks for any code change, vendor merge, or system conversion.
The decision on a REVIEW item belongs to a person, and that decision should be written down as a new rule.

Wrap-up: questions for the group

Which tool did best on which exercise? Was it the same tool every time?
Where did an AI sound confident and turn out to be wrong? How did you catch it?
Which verification check would you actually do every month?
Which one of these tasks could you try on Tuesday, with sample or approved data?