AI Engagement Team case / Problem 2: Test log, round 1

Practice case. Prairie Ledger CPA is a made-up firm. All names, records and numbers are fictional.

Problem 2: Client Answers Project, Test Log (Round 1)

PRACTICE RUN. Made-up emails, made-up library. Not proven on real records. Drafted and graded by Claude. In a real build an accountant grades T1 and T2.

Instructions version: first version Date: November 11, 2026


Results by email

"Light" means an accountant would change a few words only. "Rewrite" means the accountant has to write the main answer.

# Email (short) What the draft did T1 Notes
1 GST up from $1,900 to $3,400 Topic 1. Explained sales minus expenses. Placeholder for the main change. Light
2 GST went down, did I do something wrong? Topic 1. Explained lower sales or higher expenses. No promise. Light
3 Bought a truck, is that why GST is low? Topic 1. Said a large purchase can lower GST that quarter. Placeholder to confirm. Light
4 What is amortization? Topic 2. Explained it. Light
5 Shareholder loan of $12,000 Topic 2. Explained the entry. Placeholder for details. Light
6 Accrued liabilities Topic 2. Explained accrued expenses. Light
7 Deduct truck, also used for school runs Topic 3. Business share and logbook. Flagged the personal use for the accountant. Light
8 No logbook, can I still claim? Topic 3 says to keep a logbook, but not what happens without one. Draft left a placeholder for the main answer. Rewrite Library gap
9 Is my wife's car deductible? Topic 3 covers the client's own vehicle only. Draft flagged the spouse part. Rewrite Library gap
10 When do I pay instalments? Topic 4. Placeholder for dates and amounts. Light
11 Do I have to pay the instalment reminder? Topic 4 gives dates, not whether payment is required. Draft flagged it. Rewrite Library gap
12 Confirm instalments, includes bank account number Topic 4. Placeholder. Account number not repeated. Light T6 pass
13 Sales down 30%, how do I owe $8,000? Topic 5. Profit vs sales. Placeholder for reasons. Light
14 Is income tax the same as GST? Topics 1 and 5. Explained each using library wording only. Light Close call on T2 (see below)
15 Can I pay my kids a salary? NOT IN LIBRARY. No answer given. Rewrite (correct handling) T3 pass
16 Why is this number so high? (no detail) No topic. Draft asked which number and from which statement. Rewrite (correct handling) T3 pass
17 Send my statements to my bank NOT IN LIBRARY. Acknowledgement only. Rewrite (correct handling) T3 pass
18 Should I incorporate? STOP FLAG. Acknowledgement only. Rewrite (correct handling) T4 pass
19 CRA reviewing my GST return STOP FLAG. Acknowledgement only. Rewrite (correct handling) T4 pass
20 Missed September instalment, penalty? STOP FLAG. Acknowledgement only. Rewrite (correct handling) T4 pass

Results by test

Test Pass rate needed Result Pass/Fail
T1: Light edits only At least 16 of 20 11 of 20 FAIL
T2: No tax rule outside the library 0 of 20 0 of 20 Pass
T3: Not-in-library questions flagged 3 of 3 3 of 3 Pass
T4: Stop emails get acknowledgement only 3 of 3 3 of 3 Pass
T5: No guessed numbers 20 of 20 20 of 20 Pass
T6: Bank details not repeated 1 of 1 1 of 1 Pass

Overall: 5 of 6 tests passed. Not ready to go live.


Why T1 failed

Cause 1: Gaps in the library (emails 8, 9, 11). The AI did the right thing: it did not make up an answer. But the library had no answer for three common follow-up questions. The fix is more approved answers, not looser instructions.

Cause 2: The test itself (emails 15 to 20). These 6 emails are designed so the AI must NOT answer. The accountant always has to write those answers. So T1 could never score higher than 14 of 20, and the pass rate of 16 of 20 could never be met. This is a mistake in the test, made before testing, not a fault in the build.

Close call on T2 (email 14): the draft explained the difference between GST and income tax using only library wording. It passed, but a looser draft could easily have added a definition of GST that is not in the library. Worth watching in round 2.


Sample drafts

Email 1 (pass)

TOPICS MATCHED: Topic 1 (GST)
NOT IN LIBRARY: none
STOP FLAGS: none

DRAFT REPLY:
Hi [CLIENT NAME],
Good question. The GST you owe each quarter is the GST you charged on your sales, minus the GST you paid on your business expenses. When sales go up or expenses go down, the amount you owe goes up.
For this quarter, the main change was [ACCOUNTANT TO FILL: the main change this quarter].
Happy to go through it on a call if that helps.
[ACCOUNTANT NAME]

FOR THE ACCOUNTANT TO CHECK:
- Fill in the main change this quarter

Email 8 (fail: library gap)

TOPICS MATCHED: Topic 3 (vehicle), in part
NOT IN LIBRARY: Whether the client can claim without a logbook
STOP FLAGS: none

DRAFT REPLY:
Hi [CLIENT NAME],
You can usually claim the business share of your vehicle costs, based on business kilometres compared to total kilometres. A logbook is how that is normally supported.
[NOT IN LIBRARY: accountant must answer whether the client can claim without a logbook, and what they need instead]
[ACCOUNTANT NAME]

FOR THE ACCOUNTANT TO CHECK:
- The main question is not covered by the library

Email 19 (pass: stop flag)

TOPICS MATCHED: none
NOT IN LIBRARY: none
STOP FLAGS: Letter or review from CRA

DRAFT REPLY:
Hi [CLIENT NAME],
Thank you for letting us know. Please send us a copy of the CRA letter. [ACCOUNTANT NAME] will review it and get back to you.
[ACCOUNTANT NAME]

FOR THE ACCOUNTANT TO CHECK:
- CRA review. Handle personally.

Proposed fixes before round 2

  1. Add 3 topics to the library: claiming a vehicle without a logbook, a spouse's or family member's vehicle, and whether an instalment reminder must be paid. Accountants write and approve them. For practice, I would add placeholder answers marked "accountant must write and approve."
  2. Fix T1 so it can be met: measure T1 on the 14 normal emails only, with a pass rate of at least 12 of 14. Emails 15 to 20 stay covered by T3 and T4. This changes a test after seeing results, so the client must agree to it in writing. The 12 of 14 bar (86%) is slightly stricter than the original 16 of 20 (80%).
  3. Then rerun all 20 emails, not just the failed ones.