Practical guide

AI capability and verification risks in accounting

AI-assisted accounting work can fail even when the output is fluent and the arithmetic appears plausible. A model can use the wrong reporting period, omit a population, misread a document, invent a policy reference, or turn an uncertain classification into a definitive entry. The practical question is not whether an accountant used AI. It is whether the workflow preserved evidence, judgment,.

By Two Prune

AI-assisted accounting work can fail even when the output is fluent and the arithmetic appears plausible. A model can use the wrong reporting period, omit a population, misread a document, invent a policy reference, or turn an uncertain classification into a definitive entry. The practical question is not whether an accountant used AI. It is whether the workflow preserved evidence, judgment, review, and accountable release. This guide uses CLOSE: constrain the use case, link evidence, observe transformations, scrutinize judgments, and escalate exceptions.

Source basis and scope

The International Ethics Standards Board for Accountants' technology-related revisions to the Code strengthen expectations for professional accountants when using technology. IFAC's professional-skepticism guidance emphasizes evaluating the source, relevance, sufficiency, quality, and integrity of information and remaining alert to conflicting evidence. The IAASB's Technology Position and current technology work address appropriate technology use while safeguarding audit and assurance quality. These sources support the risk-control method below. They do not determine the accounting treatment for a transaction, establish compliance, or replace the standards and professional judgment applicable to a particular engagement and jurisdiction. IESBA technology revisions IFAC professional skepticism IAASB technology work

Sources: [1] [2] [3]

The CLOSE control method

### C - Constrain the use case

Define the task, permitted inputs, output type, accountable owner, prohibited uses, and release decision before choosing a tool.

| Control question | Evidence to retain | | --- | --- | | What accounting task is being assisted? | Workflow description and system boundary | | Which entities, ledgers, periods, and currencies apply? | Scope record | | What may the model see? | Data classification and access decision | | What may it produce? | Draft-output specification | | Who makes the accounting judgment? | Named preparer and reviewer | | What must never be automated? | Prohibited-use list |

Examples of bounded assistance include extracting candidate invoice fields, proposing reconciliation matches, summarizing an approved policy, or drafting a variance explanation. A generated journal, accounting conclusion, management representation, or assurance conclusion should not be treated as approved because the tool produced it.

### L - Link every material output to evidence

Create an evidence map before review.

| Output ID | Proposed amount or claim | Source document and location | Period/entity | Transformation | Uncertainty | | --- | --- | --- | --- | --- | --- | | | | | | | |

Open original documents. Confirm identity, date, version, entity, currency, units, tax treatment, and authorization. Detect duplicate documents and missing sequences. For narrative outputs, require passage-level links for material policy statements and factual assertions. A link to a folder or a generic policy page is not enough when the decision depends on a specific clause or transaction.

### O - Observe transformations and completeness

Trace how evidence becomes an output. Record extraction, classification, calculation, aggregation, filtering, and rounding steps. Reproduce material calculations independently of the generative model.

Use control totals and population checks:

Accuracy on a convenient sample does not establish completeness. A workflow that extracts 98 invoices correctly but silently drops two is not controlled by reviewing only the 98 visible outputs.

### S - Scrutinize judgments and contradictory evidence

Separate deterministic checks from judgments. Classification, estimates, provisions, useful lives, impairment indicators, revenue recognition, and unusual transactions can require context the model does not have.

| Judgment | Evidence for | Evidence against | Assumption | Sensitivity | Reviewer disposition | | --- | --- | --- | --- | --- | --- | | | | | | | |

Ask what would make the proposed treatment wrong. Search for amendments, side letters, subsequent events, unusual terms, related parties, and evidence that conflicts with the draft. Require the preparer to explain why the source is sufficient and why an alternative was rejected. Escalate matters outside the reviewer's competence.

### E - Escalate, approve, and monitor

Define exception thresholds before production use. Examples include missing source documents, conflicting entity data, unsupported policy citations, low-confidence extraction, unexplained reconciliation differences, changed model behavior, unauthorized data exposure, or a high-consequence judgment.

| Exception | Severity | Immediate action | Owner | Release condition | Evidence of closure | | --- | --- | --- | --- | --- | --- | | | | | | | |

Pause affected output rather than forcing it through a workflow. Record reviewer disposition and approval. Monitor error types, override reasons, unresolved exceptions, and changes to inputs or tooling. These signals diagnose the process; they do not by themselves prove accounting quality or professional competence.

  • reconcile record counts and amounts before and after transformation;
  • compare subledger, general-ledger, bank, and source totals where relevant;
  • inspect excluded, low-confidence, and failed records rather than only accepted results;
  • test duplicates, reversals, cut-off, currency conversion, signs, tax, and period mapping;
  • retain model version, configuration, prompt or instruction set, and output needed to investigate an exception;
  • rerun a stable test set after a material change.

Risk register for AI-assisted accounting work

| Risk | Example failure | Detection evidence | Minimum response | | --- | --- | --- | --- | | Scope mismatch | Wrong entity or period | Scope-to-output reconciliation | Quarantine and rerun | | Source loss | Missing invoice page | Sequence and page-count checks | Retrieve source or hold item | | Fabricated authority | Invented policy clause | Passage-level source check | Remove claim and escalate | | Transformation error | Reversed sign or currency | Independent recalculation | Correct logic and retest population | | Automation bias | Reviewer accepts plausible entry | Blind re-performance sample | Expand review and inspect overrides | | Data exposure | Restricted record sent to unapproved tool | Access and transfer log | Stop use and follow incident process | | Drift | Updated model changes classifications | Stable regression set | Requalify before release |

Worked example: fictional month-end accrual review

A finance team uses an AI assistant to identify unbilled services from purchase orders, receipt records, and supplier correspondence. The draft proposes a $420,000 accrual and describes it as complete. Most lines include a purchase-order reference, and the narrative looks ready for the close file.

CLOSE first restricts the task to proposing candidate accrual lines for one entity and period. The controller remains responsible for the accounting conclusion and posting. The evidence map links every candidate to a purchase order, service evidence, and amount basis. It reveals that the model treated three open orders as fully delivered because it interpreted scheduled dates as receipt dates.

Population controls also show that two supplier emails were excluded after an attachment parser failed. An independent recalculation removes the three unsupported lines, adds one supported service from the recovered attachments, and produces a $287,000 candidate total. The judgment table records a disputed milestone separately because acceptance evidence conflicts with the project manager's email.

The team does not post the model's $420,000 proposal. It escalates the disputed milestone, repairs the attachment-failure alert, and reruns the full population. A preparer documents the evidence links, a reviewer re-performs the control totals and high-value lines, and the controller approves the final entry. The retained pack includes the population reconciliation, exception log, calculations, judgment disposition, and evidence of approval.

The example demonstrates a controlled workflow, not a universal accounting treatment. Different facts, standards, controls, and jurisdictions can require a different conclusion.

Two Prune boundary

Two Prune can assess how a participant handles a bounded accounting-style simulation: whether they frame the task, navigate supplied records, use AI thoughtfully, verify calculations and claims, exercise judgment, and communicate exceptions. It reports observed assessment evidence. It does not provide accounting or audit services, certify compliance, approve entries, validate an organization's controls, predict workplace performance, or make employment decisions.

Sources

IESBA: Technology-related Revisions to the Code

  • IFAC: Professional Skepticism and Inquiring Mind
  • IAASB: Technology and the Future of Audit and Assurance

Sources: [1] [2] [3]

Sources

  1. 1.IESBA technology revisions · www.ethicsboard.org
  2. 2.IFAC professional skepticism · www.ifac.org
  3. 3.IAASB technology work · www.iaasb.org