Practical guide

How to measure applied AI capability in accounting

Applied AI capability in accounting is not the ability to recall tool features or produce a polished spreadsheet. It is task-bounded evidence that a person can combine accounting knowledge, professional skills, values, judgment, technology use, and verification to complete representative work to a defined standard. This guide uses LEDGER: local task, evidence design, domain criteria, governance.

By Two Prune

Applied AI capability in accounting is not the ability to recall tool features or produce a polished spreadsheet. It is task-bounded evidence that a person can combine accounting knowledge, professional skills, values, judgment, technology use, and verification to complete representative work to a defined standard. This guide uses LEDGER: local task, evidence design, domain criteria, governance and ethics, exception handling, and repeatable review.

Source basis and limits

IFAC's 2026 International Education Standards describe professional competence as the integrated application of technical competence, professional skills, and professional values, ethics, and attitudes. IES 2 includes applying accounting principles, analyzing data, using technology, and assessing controls. IES 3 includes critical thinking, judgment, communication, self-review, and review of others' work. IES 6 requires formal assessment activities to have high levels of reliability, validity, equity, transparency, and sufficiency. These standards address professional accountancy education and member organizations. They inform the method below but do not make a Two Prune simulation a professional qualification or formal accounting assessment. IFAC 2026 Handbook IES 2 IES 3 IES 6

Sources: [1] [2] [3] [4]

L - Select a local accounting task

Define the work before choosing a score. A representative task should contain a decision, a deliverable, inspectable evidence, and realistic constraints.

| Design field | Example questions | | --- | --- | | Task and context | Close review, forecast update, control analysis, reconciliation, or management report? | | Applicable basis | Which policy, accounting standard, reporting rule, and period apply? | | Decision | What must the reviewer or manager decide? | | Evidence | Which ledger extracts, contracts, schedules, policies, and messages are available? | | AI role | What may AI summarize, calculate, classify, or draft? | | Principal failure | What error would materially distort the work? | | Deliverable | What artifact and supporting record must be produced? |

Avoid a generic “accounting AI” test. Capability is meaningful only relative to the task, standard, evidence, and authority.

E - Design observable evidence

Collect the final artifact and the minimum process evidence needed to interpret it:

Do not score lack of captured activity as poor performance when the design did not create or record a fair opportunity to demonstrate it. Mark unavailable evidence and reduce confidence.

  • task framing and assumptions;
  • sources opened and passages or records used;
  • AI requests, material outputs, corrections, and rejected suggestions;
  • calculations, reconciliations, and exception logs;
  • judgments, uncertainty, and escalation decisions;
  • final review notes and submission timing.

D - Use accounting-specific criteria

Score behavior against anchored criteria tied to the task.

| Criterion | Strong evidence | Risk evidence | | --- | --- | --- | | Technical application | Applies the relevant policy or standard to the actual facts | Uses a generic rule without checking scope or facts | | Data and calculation | Reconciles populations, units, periods, formulas, and exceptions | Accepts a plausible total with a hidden denominator or mapping error | | Control awareness | Identifies source, access, change, review, and approval controls | Treats AI output as an approved accounting record | | Professional judgment | Compares interpretations and explains the chosen treatment | Presents one confident answer despite material ambiguity | | Skepticism and verification | Challenges contradictory or unsupported output and preserves evidence | Defers to fluent AI language or a citation without support | | Communication | Makes the conclusion, evidence, limit, and required decision clear | Hides a material exception in an appendix |

Use task-specific anchors rather than labels such as “good with AI.” A result can be strong for a management-accounting analysis and uninformative about tax, audit, or statutory reporting.

G - Preserve governance, values, and authority

The assessment should test whether the participant respects data restrictions, separates preparation from approval, identifies conflicts or bias, and escalates decisions beyond their authority. Use synthetic or explicitly approved non-sensitive data. Do not ask the participant to make a regulated sign-off or professional conclusion they are not authorized to make.

Record which criteria require an accounting subject-matter reviewer. An AI-fluency reviewer can assess source use and verification behavior, but should not invent the correct accounting treatment.

E - Include exception handling

Plant realistic, reviewer-keyed exceptions rather than trivia:

For each exception, document the clue, expected check, acceptable dispositions, severity, and escalation route. Credit a well-supported escalation when the evidence cannot resolve the issue.

  • a vendor invoice assigned to the wrong period;
  • a spreadsheet total that excludes a hidden filtered row;
  • an AI-generated policy citation that is real but does not support the treatment;
  • a reconciliation difference masked by netting;
  • a forecast explanation that confuses correlation with cause;
  • a request to disclose restricted data to an unapproved tool.

R - Review repeatability and interpretation

Use at least two trained reviewers for a sample of work. Compare their evidence references and reasons, not only total scores. Investigate material disagreement and refine ambiguous anchors. Repeat an equivalent task only when the evidence environment, difficulty, assistance, and scoring rules are sufficiently comparable.

Report:

Do not convert one task into a claim about every accounting role, workplace outcome, incident rate, or future performance.

  • observed task and conditions;
  • criterion-level evidence and missingness;
  • score confidence and reviewer disagreement;
  • accounting-domain and transfer limits;
  • development areas and follow-up questions.

Worked example: fictional month-end accrual review

A participant receives a ledger extract, three supplier emails, an accrual policy, a prior-month schedule, and an AI workspace. The assignment is to prepare an accrual review note for a fictional controller. The AI proposes a neat schedule and says a £48,000 invoice belongs in the next period because its invoice date is after month end.

The source emails show that the service was delivered before month end, while the policy bases recognition on service delivery. The AI schedule also nets a £7,500 reversal against an unrelated vendor, hiding a reconciliation difference. A strong response checks the policy against the facts, separates the items, reproduces the schedule total, and flags that one email lacks sufficient delivery evidence. It prepares a bounded recommendation and asks the controller to resolve the unsupported item.

LEDGER scores the technical application, reconciliation, verification, judgment, control awareness, and communication with direct evidence references. The participant is not penalized for declining to choose a treatment where the supplied evidence is insufficient. The report states that the simulation observed performance on this accrual task and does not establish professional qualification, audit competence, or performance across accounting domains.

Measurement QA checklist

[ ] Task, applicable basis, decision, evidence, AI role, failure, and deliverable are explicit.

  • [ ] Criteria integrate technical work, professional skills, values, judgment, and technology use.
  • [ ] Every exception has a reviewer evidence key and a reasonable detection route.
  • [ ] Missing evidence lowers confidence rather than becoming a negative score by default.
  • [ ] Accounting judgments receive qualified subject-matter review.
  • [ ] Reviewer disagreement and task limits remain visible.
  • [ ] Interpretation stays bounded to the observed task and conditions.

Two Prune boundary

Two Prune can run realistic accounting-style simulations and report observed evidence across its six AI-fluency pillars. It does not qualify professional accountants, conduct an audit, approve accounting treatments, provide accounting advice, predict workplace performance, or make employment decisions. Customers remain responsible for task validity, subject-matter review, accessibility, and human interpretation.

Sources

IFAC: 2026 Handbook of International Education Standards

  • IFAC IES 2: Technical Competence
  • IFAC IES 3: Professional Skills
  • IFAC IES 6: Formal Assessment of Professional Competence

Sources: [1] [2] [3] [4]

Sources

  1. 1.IFAC 2026 Handbook · www.ifac.org
  2. 2.IES 2 · education.ifac.org
  3. 3.IES 3 · education.ifac.org
  4. 4.IES 6 · education.ifac.org