Practical guide

AI-enabled knowledge-work workflows in accounting

Accounting teams can use AI to assist with document handling, analysis, drafting, and exception discovery, but the useful unit of design is a controlled workflow, not a prompt. Each workflow needs a defined accounting purpose, source boundary, transformation record, human judgment point, and evidence of review. This guide uses TRACE: task boundary, records, AI action, control checks, and.

By Two Prune

Accounting teams can use AI to assist with document handling, analysis, drafting, and exception discovery, but the useful unit of design is a controlled workflow, not a prompt. Each workflow needs a defined accounting purpose, source boundary, transformation record, human judgment point, and evidence of review. This guide uses TRACE: task boundary, records, AI action, control checks, and escalation.

Source basis and limits

IFAC's 2026 IES 2 describes professional competence as integrated application of technical knowledge, professional skills, and values, ethics, and attitudes. Its technology outcomes include using ICT for analysis and decision support while evaluating processes and controls. IES 3 covers intellectual, interpersonal, communication, personal, and organizational skills. IESBA's technology-related revisions strengthen ethical expectations when professional accountants use technology. These sources support the workflow design below. They do not prescribe accounting treatment or establish compliance for any entity, engagement, standard, or jurisdiction. IFAC IES 2 IFAC IES 3 IESBA technology revisions

Sources: [1] [2] [3]

The TRACE workflow card

| Element | Design question | Evidence artifact | | --- | --- | --- | | Task boundary | What accounting purpose, entity, period, and output are in scope? | Approved workflow specification | | Records | Which authoritative records may be used, and how is completeness checked? | Source inventory and control totals | | AI action | What may the system extract, classify, compare, calculate, or draft? | Versioned instruction and output log | | Control checks | Which checks must a preparer and reviewer perform? | Test results and review evidence | | Escalation | Which exceptions stop release or require specialist judgment? | Exception and disposition log |

Five representative workflows

### 1. Document intake and coding proposals

AI can extract supplier, date, amount, tax, purchase-order, and line-item fields, then propose a coding classification. Keep the source image or file linked to every extracted field. Reconcile document counts and amounts before and after processing. Route duplicates, missing pages, uncertain entities, unusual terms, and low-confidence fields to review.

The system proposes; an accountable person decides whether the record is complete, authorized, and coded appropriately.

### 2. Reconciliation and exception triage

AI can suggest matches across bank, subledger, and general-ledger records or group unmatched items by likely cause. Preserve the matching rule, input populations, excluded records, and tolerance. Review unmatched and many-to-one cases, not only successful matches. Reperform material matches independently and investigate unexplained differences.

Do not equate a high match rate with a controlled reconciliation. Completeness and exception treatment matter.

### 3. Close analysis and estimate support

AI can organize supporting documents, identify missing evidence, compare periods, or draft a schedule of candidate issues. Estimates and accounting conclusions require facts, applicable standards, assumptions, sensitivity, and professional judgment beyond a generated narrative.

Use a judgment record:

| Matter | Evidence for | Evidence against | Assumption | Sensitivity | Owner and disposition | | --- | --- | --- | --- | --- | --- | | | | | | | |

### 4. Variance and management-report drafting

AI can draft explanations from an approved data set, but every material amount, comparison, and causal statement needs support. Reproduce calculations, align periods and definitions, identify one-off items, and separate observation from inference. The executive summary must preserve material uncertainty and counter-evidence.

### 5. Policy and research assistance

AI can search an approved policy library, compare versions, or draft a research outline. Require passage-level citations, effective dates, jurisdiction, and source hierarchy. Escalate ambiguous facts, conflicting authority, and matters outside the user's competence. A plausible citation is not evidence until the original source is opened.

Worked artifact: fictional revenue-variance workflow

A management accountant receives a request to explain why quarterly service revenue is 8% below plan. The available files include the approved budget, general-ledger extract, customer schedule, pipeline report, and a note from sales stating that delayed renewals caused the gap.

The TRACE card limits the task to preparing a variance analysis for one entity and quarter. The source inventory reconciles the ledger revenue total to the reporting pack and records the budget version. AI may classify account movements and propose explanation candidates, but it may not decide the accounting treatment or publish the commentary.

The first draft attributes the full gap to delayed renewals. Control checks show that the sales note covers only two accounts. A period comparison finds a billing-timing change, and the customer schedule reveals one cancelled contract. The accountant creates a bridge separating volume, timing, cancellation, and unexplained residual. Each component links to a source and calculation.

The remaining residual exceeds the agreed threshold, so the workflow stops release. Finance operations discovers that a new account mapping omitted one revenue code from the management extract. After the mapping is repaired, the ledger is reconciled again and the commentary is redrafted. A reviewer reproduces the bridge, checks the source links, and records approval.

The retained artifact contains the task specification, source inventory, control totals, AI draft, corrected bridge, exception disposition, and review record. It demonstrates a controlled fictional workflow, not a universal accounting method.

Workflow qualification checklist

Is the accounting purpose and accountable owner clear?

  • Are authorized inputs, data classifications, and retention rules defined?
  • Can every material output be traced to original records?
  • Are population completeness and control totals tested?
  • Are deterministic checks separated from professional judgments?
  • Are failure thresholds and stop conditions defined before use?
  • Is the workflow retested after source, model, instruction, or system changes?
  • Can a reviewer reconstruct what happened without relying on the model's explanation?

What workflow evidence can and cannot show

Workflow evidence can show that defined checks occurred, exceptions were handled, and outputs were reviewed under stated conditions. It does not by itself prove that the accounting conclusion is correct, that controls operate consistently over time, or that a person has general professional competence. Those claims require additional evidence and appropriate professional review.

Two Prune boundary

Two Prune can assess how a participant handles a bounded accounting-style simulation, including source navigation, AI use, verification, judgment, and communication. It reports observed assessment evidence. It does not provide accounting or assurance services, approve entries, validate internal controls, certify compliance, or make employment decisions.

Sources

IFAC: IES 2, Technical Competence

  • IFAC: IES 3, Professional Skills
  • IESBA: Technology-related Revisions to the Code

Sources: [1] [2] [3]

Sources

  1. 1.IFAC IES 2 · education.ifac.org
  2. 2.IFAC IES 3 · education.ifac.org
  3. 3.IESBA technology revisions · www.ethicsboard.org