AI LITERACY FOR ACCOUNTING STUDENTS
Don’t Trust It — Test It
5 exercises in being AI’s toughest critic — with real Excel and general ledger examples — every accounting student should complete before the fall semester begins.
Here’s the part of AI literacy nobody puts on the syllabus: the accountant is liable for the error, not the AI vendor. When a model invents a citation, misreads a reconciling item, or quietly assumes something you never told it, the output still looks fluent and confident — that’s what makes it dangerous. The AICPA’s current guidance is blunt about this: AI-generated analyses, summaries, and computations must be verified against primary source data before they go anywhere near a client deliverable, and that verification has to be documented.
So the skill worth practicing before the semester begins isn’t “how to prompt AI.” It’s how to read AI output the way an auditor reads a client’s schedule — assuming it’s wrong until you’ve checked it. Each exercise below has you use the AI, then hunt for exactly where it’s likely to fail. All five run on free tools and take about two hours total.
1. Give the AI a formula to write — then try to break it
Getting Excel or Copilot to write a formula from a plain-English request is the easy part now. The exercise is what you do next: feed the formula the exact inputs most likely to expose a flaw, and see if it survives.
Excel example
You have a Chart of Accounts tab (Account #, Account Name, Type) and a Trial Balance tab (Account #, Debit, Credit).
Prompt: “Write an Excel formula that looks up the account number from the Trial Balance tab in the Chart of Accounts tab and returns the account name.”
You’ll likely get something like:
=XLOOKUP([@Account], ChartOfAccounts!A:A, ChartOfAccounts!B:B)Now stress-test it yourself, without asking the AI first: add an account number that doesn’t exist in the Chart of Accounts. Add a duplicate account number. Format one account number as text instead of a number. Leave a cell blank.
What you’re checking for
Does the formula return a clean error or a silent wrong answer? #N/A is honest; a formula that quietly returns the wrong row on a duplicate is the dangerous failure mode. Silent failures are particularly insidious because they pass a surface-level review but contaminate downstream analysis.
Then ask the AI to fix whichever case broke it, and verify the fix against the same test cases — don’t just take its word that it’s fixed. Run the corrected formula through the identical stress tests. Document any cases where it still fails.
Why it counts: ICAEW’s 2026 guidance on spotting AI errors in financial models makes the same point — errors hide in the edge cases a confident-sounding formula never mentions.
2. Score the AI on false alarms, not just catches
Ask AI to review a general ledger for anomalies and it will almost always find something — the real test is whether what it finds is right, and whether it misses what it should have caught. That means you need to know the answer key before you grade it. This two-dimensional scoring (catching real issues + avoiding false positives) is what separates a usable tool from a liability.
General ledger example
Build a 40–50 line GL export (Date, Account, Description, Debit, Credit, Entered By). Make sure it’s realistic: include some routine transactions, some legitimate unusual entries, and the problems you’ll plant.
Plant real problems: a duplicate entry, an entry where debits ≠ credits, a weekend-dated posting, or a transaction that violates your company’s normal approval workflow.
Also plant a decoy: a legitimate transaction that looks suspicious but isn’t — e.g., a large round-number entry that’s an actual quarter-end accrual true-up, properly memoed and approved. This is the test that catches overconfident AI and overconfident humans alike.
Prompt: “Review this general ledger export and flag anything that looks unusual, with your reasoning for each flag.”
What you’re checking for
False negatives: real problems the AI didn’t catch. Note which ones it missed and why — was the entry buried in similar legitimate entries? Did it fail to check for math errors?
False positives: the legitimate accrual it wrongly flagged, because it pattern-matched on “round number” instead of reading the memo. This reveals whether the AI actually reads source documentation or just reacts to surface patterns.
Score it: (issues correctly caught) / (total real issues), and (false flags) / (total flags). A model that’s wrong in either direction and unsupervised is a liability, not a shortcut. Aim for 100% catch rate with zero false positives — that’s the standard you’d hold a junior accountant to.
Why it counts: this is the exact failure pattern FINRA’s 2026 oversight report flags for firms using generative AI — hallucination and bias sit right next to the model’s most confident-sounding claims.
3. Reconcile it yourself first — then find where the AI’s version disagrees with you
Do the reconciliation cold, on your own, before you ask AI anything. Then ask AI to build the same one and compare line by line. Working independently first means you can’t unconsciously anchor on whatever the AI says — you’re checking it against your own work, not against your memory of what sounded right. This is the hardest habit to build because anchoring on AI output happens automatically the moment you see it.
Excel example
Bank statement ending balance: $18,420.16. Book (cash account) balance: $17,955.40.
Outstanding checks: #1042 for $610.25, #1049 for $390.00. Deposit in transit: $1,200.00. Bank service charge not yet recorded: $35.51.
Build the reconciliation yourself first. Then prompt: “Build a bank reconciliation from this data and explain each adjustment.”
What you’re checking for
Do the two adjusted balances match? If not, trace the AI’s reasoning line by line against source data — the bank statement and the check register — not against how convincing its explanation sounds. Work backwards from the numbers to the supporting docs.
A classic AI failure here: putting an adjustment on the wrong side (treating a deposit in transit as a book adjustment instead of a bank-side one). Fluent doesn’t mean correct. Another common error: arithmetic mistakes buried in a confident narrative. A third: missing an item entirely but writing a plausible-sounding explanation anyway.
Document which side each item should affect and why. Then check if the AI got it right, and whether its explanation of the reasoning was accurate or just sounded good.
Why it counts: verifying against primary source data, not against the model’s own explanation of itself, is exactly the standard the AICPA’s 2026 guidance sets for any AI output in a deliverable.
4. Find the assumption the AI never told you it made
Give AI an accounting scenario with a missing fact — one where the correct treatment genuinely depends on information you withheld. A good model should flag the ambiguity or ask a clarifying question. A model that just picks an answer and states it with total confidence is showing you exactly the risk the profession is worried about. This exercise trains you to spot confident-sounding answers that are actually guesses.
General ledger example
Scenario: On January 1, your company pays $12,000 cash for a 12-month software subscription.
Prompt: “Draft the journal entry and a memo explaining the accounting treatment under U.S. GAAP.”
It will likely assume straight-line recognition over 12 months and a standard prepaid asset — reasonable, but an assumption, not a given. The model will present this with equal confidence whether or not it’s the right call.
Now add the missing fact and ask again: the contract is cancellable with a full refund of unused months at any time. Does the AI’s answer change? Did it even notice the assumption it made the first time? Ask a third time with a different twist: the company is in financial distress and this subscription is critical to operations — does that change the treatment?
What you’re checking for
Whether the AI states its assumptions up front or buries them inside a confident-sounding answer. A trustworthy model will say “assuming standard U.S. GAAP and no special contract terms” before it answers. A less reliable one will just answer.
Whether a fact you add later actually changes the treatment — and whether the AI catches that on its own or only when you point it out. If it changes its answer without acknowledging that the first answer was conditional, that’s a red flag.
Whether the model can distinguish between facts that matter and facts that don’t. Financial distress by itself doesn’t change revenue recognition — so if the AI’s answer changes there, it’s reasoning by pattern match, not by principle.
Why it counts: Forbes’ 2026 coverage of AI error responsibility in accounting is direct about this — the professional is on the hook for judgment calls the AI made invisibly. Surfacing those calls is the job.
5. Keep a verification log, not a prompt log
The habit firms are actually asking for isn’t a record of what you asked AI — it’s a record of what you checked and what you found wrong. That’s the documentation the AICPA’s verification standard describes, and it’s a real working paper, not busywork. This is the artifact that protects you when something goes wrong.
Verification log example (one row per exercise above)
Columns: Date | Tool | What I asked | What it gave me | What I checked it against | Errors or assumptions I found | Trust score (high/medium/low)
Example row: “8/12/26 | ChatGPT | Reviewed 45-line GL export for anomalies | Flagged 6 entries | Checked each against source memos and approvals | 1 false positive (legitimate accrual), missed 1 real duplicate | Medium — needs a second pass”
By August 24 you’ll have a one-page record that shows the AICPA’s required verification step, not just AI usage — the difference between “I used AI” and “I audited what it gave me.” This log becomes part of your working papers. It’s also the record that shows you did your job if the AI output later turns out to be wrong.
Why it counts: when AI makes the error, the accountant is still liable — a documented verification trail is the only thing that protects you, and it’s a habit worth building before it’s required of you.
Before you trust any AI output, ask two questions
Can I verify this against a primary source — not against how confident it sounds? And what did it assume that it never told me? If you can’t answer both, you haven’t finished the exercise yet — you’ve just used the AI. Build these two questions into every deliverable, every deadline, every time you’re tempted to ship an AI-generated answer without testing it first.
Leave a Reply