Gains:
- Understanding the logic of automating current reconciliation, bank reconciliation and three-way matching (order-delivery note-invoice) tests with artificial intelligence support
- Ability to classify and prioritize difference (difference) analysis and the reasons for reconciliation differences with artificial intelligence
- Ability to understand that mismatched items and confirmation differences must be investigated by the auditor and that artificial intelligence only shows the difference.
The backbone of an audit is verifying by comparing it with an independent source that a record actually exists and has been kept accurately. A credit entry is real, with the approval of the debtor party; a bank balance is correct, with confirmation from the bank; You test the authenticity of a purchase by ensuring that the order-delivery note-invoice trio matches each other. These processes are collectively called reconciliation and confirmation. Because they inherently require the comparison of large numbers of items, they are repetitive, time-consuming, and where artificial intelligence shines.
In this unit, we will cover three jobs: current/account reconciliation (comparing the records of two parties), bank reconciliation (reconciling the ledger balance and bank statement), and three-way match (comparing the purchase order, goods receipt/delivery note, and vendor invoice). The basic principle remains unchanged: AI matches and points out the difference; It is up to the auditor to investigate the reason for the difference and come to a conclusion.
What is three-way matching and why is it important?
When a business purchases goods/services, three documents typically occur: (1) Purchase order — what was ordered, how much, at what price; (2) Delivery note/receipt of goods — what, how much was received; (3) Vendor invoice — what, how much was invoiced. In a healthy control environment, these three should match: ordered = received = invoiced. The three-way match test checks whether these three match. Disputes; It could mean over-invoicing, fictitious purchases, duplicate payments, price manipulation or simple data error.
AI can match thousands of purchases across these three documents in seconds. But the list of "unmatched items" that appears is not a list of errors. An item may be mismatched for the following innocent reasons: partial delivery (order 100, delivery 60), timing difference (invoice not yet arrived), price difference within acceptable tolerance, discount or shipping item. The auditor's job is to find the reason for the difference and separate the innocent from the suspect.
Match type
compared
Typical reasons for difference
Current reconciliation
Our registration ↔ counterparty registration
Timing, refund, incomplete registration, exchange rate difference
Bank reconciliation
Ledger ↔ bank statement
Check in transit, bank charge, record not yet processed
three way match
Order ↔ delivery note ↔ invoice
Partial delivery, timing, price difference, duplicate, fictitious
External confirmation: one of the most reliable evidence
External confirmation in auditing is the method of obtaining a written response directly from a third party (bank, customer, supplier) about a balance or transaction, and BDS 505 regulates this. External confirmation is one of the auditor's strongest evidence; because the information comes from an independent source, not from the audited entity. Here, AI is the accelerator in producing draft confirmation letters, comparing incoming responses with ledger records, and listing non-responsive parties (those to follow). But two points are critical: first, the filling and sending of the confirmation letter with the actual addressee information should be under the auditor's control; Secondly, the auditor must be sure that the response is genuine and reaches the auditor directly. The integrity of the verification process rests with the auditor; AI just makes comparison and tracking easier.
Gap analysis: AI classifies, auditor decides
The most productive way to examine consensus differences is to classify the differences: which are timing-related (will shift between periods but are not actual differences), which are permanent differences, and which are unexplained. AI can quickly do this preliminary classification: it puts the difference amount, direction, and possible cause into a table. But beware: AI's "probable cause" column is a prediction, not a determination. Just because it says "probably a timing difference" does not prove that the difference is timing; The auditor must verify this with documentation (next period invoice, return receipt, bank receipt).
Tip: Prioritize differences by amount size and explainability. Skip the small and easily explained differences; Concentrate auditor work on large or unexplained differences. AI produces the prioritization table; You set the threshold and focus.
Step-by-step reconciliation automation
- Prepare and anonymize two (or three) datasets. Determine the matching key (invoice number, order number, amount + date).
- Confirm completeness. Do the total and number of records for each cluster agree with what was expected?
- Define the matching rule. Will there be an exact match or will tolerance (e.g. ±1% price difference) be allowed?
- Match with AI. Those that match, those that do not match, and those that are within tolerance appear in separate lists.
- Classify and prioritize differences. Timing/persistent/unexplained.
- Investigate as an auditor. Verify any significant differences with documentation; Document the conclusion and rationale.
three mini cases
Case 1 — Duplicate payment capture. An auditor three-way matched 12,400 purchases with AI. YZ marked 6 items with two separate payments for the same invoice number. The auditor examined: 4 were legitimate invoices with similar numbers from different years; but 2 of them were actually duplicate payments (the same invoice was paid twice), totaling 88,000 TL. The reinstatement process was initiated and the internal control weakness was written in the management letter. AI asked 6 questions; 2 real problems the auditor confirmed.
Case 2 — Mistaking the timing difference for a mistake. A team member directly wrote off the 40,000 TL difference marked by YZ in the bank reconciliation as a "recording error". The responsible person examined: the difference was a check drawn at the end of the period and not yet presented to the bank; It was a completely normal timing difference and it closed the following week. Lesson: a mismatched pen is not an automatic error; No conclusion can be written without investigating the cause.
Case 3 — Tolerance error. One auditor gave no tolerance for three-way matching; Due to penny rounding differences, YZ 900 marked the item "does not match". The list became unexamined; The real issues got lost in the noise. Once the auditor defined a reasonable tolerance (e.g. ±5 TL or ±0.5%), the list was reduced to 27 significant items. Lesson: the matching rule must match business reality; Too strict a rule produces noise, too loose a rule produces blindness.
Weak prompt / Strong prompt
Weak prompt:
Compare these two lists, tell me what doesn't fit.
Problem: matching key, tolerance and output format are unclear. AI doesn't know what to match based on; the result would be unreliable and unrepeatable.
Powerful prompt:
Your role: you are the reconciliation assistant to an independent auditor. You will match and SHOW the difference; The reason and result of the difference are mine.Data (anonymised):- Set A: Purchase orders (columns: order_no, item, quantity, unit_price, amount)- Set B: Delivery notes (order_no, item, delivered_quantity)- Set C: Invoices (order_no, item, invoiced_quantity, invoiced_amount)Matching key: order_no + item.Tolerance: count as "matched" if the amount difference is within ±5 TL or ±0.5% (whichever is greater).Task:1) Match three clusters. Give the following three lists separately: (a) exact matches (summary number), (b) matches within tolerance, (c) items that DO NOT match (full detail).2) Add a POSSIBLE cause column for mismatches (partial delivery / missing document / quantity difference / amount difference), but state that this is a ESTIMATE, to be verified with evidence.3) Write in plain text the matching and tolerance rule you applied.4) Also give the total reconciliation (A, B, C). totals) so I can check completeness. Don't make up what you can't deduce from the data.
This prompt is powerful because it defines the key and tolerance, asks for three separate lists, positions “probable cause” as a guess, and requires total agreement for completeness.
Common mistakes
- Mistaking a mismatched item as a mistake. Writing a conclusion without investigating the reason (timing, partial delivery, return).
- Not setting tolerances. Inflate the list by pennies and lose the real issues in the noise.
- Leaving the pairing key ambiguous. Not controlling what the AI matches based on.
- Skipping completeness. Relying on matching sets without reconciling their totals.
- Mistaking "probable cause" for detection. Accepting the AI's prediction without verifying it with documentation.
Caution: Describing a difference as a "duplicate payment" or "fictitious purchase" is a serious claim. Verify this with documentation (invoice, payment receipt, goods receipt record) before writing it down; If there is a pattern, evaluate it separately as an internal control weakness.
In summary
Reconciliation, confirmation and three-way matching are among the most repetitive tasks of auditing and most easily accelerated with AI. AI matches thousands of items in seconds, lists the differences, and pre-classifies them. But the mismatched pen is not an error, it is a difference to be investigated; “probable cause” is a guess, not a determination. The auditor determines the matching key and tolerance according to the business reality; The auditor verifies the difference with the document; The auditor writes the result. AI shows the difference, the auditor gives the meaning.
Application task
Consider three small data sets (order, delivery note, invoice) or two sets (ledger, bank). Determine the matching key and a reasonable tolerance. Have the AI match with the powerful prompt pattern above and remove the ones that don't match with the "probable cause" column. Then write an "investigation plan" for each mismatched item: which document do I look at, which question do I ask, which conclusion do I consider this a bug/normal?
checklist
- [ ] I anonymized the data and clearly identified the matching key.
- [ ] I confirmed the completeness of each cluster with the total and number of records.
- [ ] I set a tolerance appropriate to business reality (neither too strict nor too loose).
- [ ] I took the matching / within tolerance / non-matching lists separately.
- [ ] I searched the document for every significant item that did not match; I confirmed why.
- [ ] I did not count "probable cause" guesses as results without confirming them with evidence.
- [ ] I have documented serious qualifiers such as duplicate/fictitious and evaluated the internal control impact.