How to create an account reconciliation with AI

Use the CLEAR reconciliation method

The CLEAR method turns an ambiguous request to “reconcile the account” into five controlled passes: Control objective, Ledger boundary, Evidence tie-out, Adjustments, and Review. It begins with the decision the reconciliation supports and ends with an explicit release disposition.

The PCAOB auditing standards for fiscal years beginning on or after December 15, 2024 describe internal control over financial reporting as supporting reliable records, authorized transactions, and timely prevention or detection of misstatement. They also emphasize risk-based control selection, sufficient evidence, and the difference between a control's design and its operation. CLEAR borrows those evidence disciplines as practical educational guidance; it does not turn a reconciliation into an audit opinion.

• Control objective: name the assertion being checked, such as completeness of cash activity or validity of a clearing balance.
• Ledger boundary: record entity, account, currency, source system, period end, cut-off rule, and ledger extract identifier.
• Evidence tie-out: compare the ledger closing balance with an independently obtained source balance and preserve both source references.
• Adjustments: list reconciling items one by one, with owner, evidence, age, treatment, and due date. Never let AI silently net unrelated items.
• Review: have an authorized reviewer reperform the arithmetic, inspect material exceptions, and choose release, repair, or hold.

Sources: [1]

Prepare a source-and-authority card before prompting

The source-and-authority card is the first artifact. It prevents a model from inventing the basis of comparison or treating a plausible explanation as evidence. Complete it before uploading or quoting any data.

If sensitive data cannot be put into the approved AI environment, work from masked transaction identifiers and controlled summaries. The NIST Generative AI Profile recommends mapping context and risks, documenting knowledge limits, evaluating outputs, and maintaining human oversight. Applied here, that means recording which fields the model saw, what transformation it performed, and which conclusions were independently checked.

• Ledger file: GLCash10102026-09.csv, extracted 2 October 2026 by the financial controller.
• Source file: September bank statement downloaded directly from the bank portal by the treasury analyst.
• Boundary: Entity SG01, account 1010, SGD, activity posted through 30 September at 23:59 SGT.
• Materiality for investigation: every difference is listed; items above SGD 5,000 require controller review.
• AI permission: classify descriptions and draft questions from redacted rows only; no portal access, posting, approval, or evidence creation.
• Human authorities: preparer owns the schedule; treasury confirms bank-side facts; controller approves disposition.

Sources: [2]

Build a reconciliation table that can be reperformed

Use one row per reconciling item. Minimum columns are item ID, source-side amount, ledger-side amount, difference, date, category, evidence reference, proposed treatment, owner, due date, status, AI contribution, and reviewer conclusion. Keep formulas visible. A reviewer should be able to recompute source balance + source-only items - ledger-only items = adjusted ledger balance without reading a narrative.

AI can propose categories such as timing difference, duplicate, fee, posting error, or unexplained. It can also group similar descriptions and draft follow-up questions. The preparer must verify each category against the source record. A model-generated category is a hypothesis until a person links it to evidence. Do not let a summary replace the row-level trail, and do not delete zeroed or corrected rows; retain their disposition so the review history remains intelligible.

Use three statuses only. Cleared means evidence explains the item and no further action remains. Repair means the difference is understood but a posting, document, or confirmation is outstanding. Hold means the difference is unexplained, the evidence is insufficient, or the required authority has not signed off. This keeps an attractive zero balance from concealing weak evidence.

Worked example: September bank reconciliation

This synthetic example is for method demonstration. The ledger closes at SGD 248,400 and the bank statement closes at SGD 252,950. The first pass finds four differences: a SGD 6,000 deposit recorded in the ledger after bank cut-off, a SGD 1,200 bank fee absent from the ledger, a SGD 3,500 supplier payment duplicated in the ledger, and a SGD 150 unexplained statement credit.

The preparer asks the approved AI tool to classify redacted descriptions and draft investigation questions. It correctly groups the deposit as a timing candidate and the fee as a posting candidate, but labels the duplicated payment as a bank timing item. The preparer rejects that label after matching two ledger entries to one approved payment instruction. The schedule records the rejected suggestion rather than erasing it.

The arithmetic is then shown explicitly: SGD 252,950 bank balance minus SGD 6,000 deposit in transit plus SGD 1,200 bank fee correction plus SGD 3,500 duplicate-payment correction equals SGD 251,650. The remaining SGD 3,250 gap reveals that the initial four-item story is incomplete. The artifact is therefore marked hold, not forced to zero. The controller requests a fresh ledger extract and source trace for the unexplained difference. The example demonstrates the main control: arithmetic and evidence can overturn a fluent AI narrative.

Verify the artifact and state its limits

Run a final three-way check. First, reperform every formula from the preserved source balances. Second, sample the evidence links and confirm they open the exact statement page, ledger row, approval, or correction referenced. Third, compare the reviewer conclusion with the status of every material and aged item. Record who reviewed, when, which version, and what remains open.

Do not describe the artifact as proof that financial statements are correct, that a control is effective across periods, or that legal or accounting requirements have been met. Those conclusions require the applicable framework, qualified judgment, and evidence beyond one schedule. AI can improve organization and surface anomalies, but it can also misclassify transactions, omit inconvenient rows, or invent explanations.

Two Prune can assess how a participant frames the reconciliation task, navigates the supplied evidence, uses AI, verifies calculations, controls risk, and communicates a bounded conclusion in a realistic simulation. It does not post entries, approve balances, perform an audit, certify control effectiveness, rank people, or make employment decisions.