Occupation manual 004 / Accounting operations

Let AI prepare the close. Keep financial authority human.

A conservative workflow for source control, reconciliation preparation, variance questions, missing-evidence queues, entry proposals, and close review—without automating accounting judgment, posting, tax filings, signatures, payroll, or payments.

Short answer: use AI like a close-preparation clerk. It may organize approved records, calculate differences from supplied numbers, flag missing evidence, draft neutral variance questions, and prepare an entry-proposal queue. It must not determine accounting or tax treatment, post entries, lock periods, file or sign returns, release payroll, change bank details, initiate payments, or make audit and attest judgments.

Professional and financial-data boundaryThis is educational workflow design, not accounting, tax, audit, attest, legal, payroll, investment, or compliance advice. The entity's reporting framework, jurisdiction, engagement terms, professional standards, firm policies, system controls, and actual facts govern. Never put taxpayer information, bank/card data, payroll records, credentials, customer information, or confidential financial records into an unapproved AI service.

What the official sources make clear

Computers may organize records

The IRS says businesses may choose a suitable recordkeeping system that clearly shows income and expenses. Good records support financial statements, tax preparation, and substantiation. A bounded tool can help organize already-approved records and surface missing items.

Evidence and judgment do not transfer

For PCAOB-covered audits, auditors must exercise professional judgment and obtain sufficient appropriate evidence. More low-quality evidence does not cure poor quality. A generated schedule or explanation is not a substitute for source reliability, procedures, review, or a professional conclusion.

Scope warningPCAOB standards apply to PCAOB-covered audits, not every bookkeeping engagement. Circular 230 applies to covered practice before the IRS. Use these sources to define hard boundaries, then have the qualified owner determine what actually applies.

Build the seven-part Accounting Close Preparation Desk

01

Set the close boundary

Record entity, period, basis/framework, system of record, approved sources, permitted tools and data, preparer, reviewer/posting authority, and tax/audit/payment escalation owners.

Output: close boundary card
02

Name every source and period

Track exact source system, period/date, owner, completeness confirmation, and human verification. AI must not silently treat an export as complete or authoritative.

Output: source and period ledger
03

Prepare reconciliations

Place ledger and independent/source amounts beside their source IDs. AI may calculate the supplied difference. A qualified human determines reliability, accounting meaning, materiality, and resolution.

Output: reconciliation preparation table
04

Route missing evidence and variances

Turn unexplained differences into neutral questions with assigned owners and required evidence. Do not generate a convenient explanation or management representation.

Output: variance and evidence queue
05

Separate proposals from posting

Document any proposed entry with source evidence, period, accounts, supplied amounts, and reviewer decision. The preparation tool cannot silently approve or post it.

Output: proposed-entry review card
06

Stop at tax, attest, and money movement

Route tax positions and filings, audit/attest judgments, payroll, bank-detail changes, refunds, transfers, payments, fraud indicators, control failures, and legal matters to authorized owners.

Output: hard-stop and escalation record
07

Issue a close-review receipt

Record independent review, open items, entry approvals and posting evidence, data handling, specialist routes, and the qualified human who certifies the close.

Output: human close receipt

Give AI bounded jobs, never financial authority

JobApproved inputAI outputHuman gate
Source-control clerkApproved source names, periods, owners, and status fieldsSource ledger and missing-period flagsOwner confirms completeness and authority
Reconciliation clerkHuman-supplied ledger/source amounts with IDsDifference calculation and open-item tableQualified reviewer determines meaning and resolution
Variance-question clerkApproved comparison tableNeutral questions and evidence requestsResponsible owner answers; reviewer validates evidence
Entry-proposal clerkReviewer-approved facts and source evidenceDraft proposal card onlyAuthorized human approves/rejects; separate authority posts
Close-checklist clerkCurrent firm/entity checklist and verified statusesOpen-item and signoff queueController/qualified owner certifies completion

Do not delegate these decisions or actions

AI may prepare

  • source and period inventories
  • supplied-number difference calculations
  • missing-evidence lists
  • neutral variance questions
  • proposed-entry review cards
  • close checklist status views
  • public rule/update summaries with direct links

Authorized humans must control

  • accounting classification and treatment
  • materiality and source reliability
  • journal-entry approval and posting
  • period lock and financial statement issuance
  • tax positions, elections, filings, and signatures
  • audit/attest evidence, conclusions, opinions, and signatures
  • payroll, bank changes, refunds, transfers, and payments

A difference is not an explanation

AI can subtract an independent/source amount from a ledger amount when both are supplied and source-linked. It cannot establish that either input is complete, authentic, correctly timed, or appropriate. Keep the output as an open item until a qualified reviewer inspects evidence and records a disposition.

Safe output vocabularyUse: matched, difference found, evidence missing, source conflict, period unclear, owner assigned, reviewer resolved. Avoid: correct, fraudulent, deductible, immaterial, compliant, final, or approved unless the authorized human has made and documented that conclusion.

A proposed entry is not a posted entry

Drafting a proposal and posting to the ledger are different permissions. A proposal needs visible source evidence, preparer identity, period, rationale from approved notes, reviewer decision, and posting evidence. Preserve segregation of duties; never give a general AI assistant silent write authority to the accounting system.

Tax filing, signatures, and client data are hard stops

The IRS describes competency, diligence, and ethical standards for covered Circular 230 practice. IRS e-file resources also distinguish authorized providers and taxpayer signature authorizations. The IRS and FTC require covered tax practices to protect client information. A generic AI workflow should prepare questions and checklists—not file, sign, transmit, impersonate authorization, or receive unapproved taxpayer data.

Use this prompt only with approved inputs

Act as an accounting close-preparation clerk, not an accountant of record, tax professional, auditor, controller, approver, or payment operator.

From the approved source-linked tables, create: (1) a missing-evidence list, (2) supplied-number difference calculations, (3) neutral variance questions, (4) a proposed-entry review queue without posting, and (5) a close checklist.

Do not invent facts, documents, amounts, classifications, business purpose, accounting treatment, tax treatment, audit conclusions, or management representations. Do not file, sign, post, lock a period, send a message, change bank details, or initiate a payment. Mark every uncertainty for qualified human review.

Final review before the close is certified

First-party source desk

Build the close controls before adding another AI tool.

Download the desk, test only with synthetic or expressly approved data, and have qualified accounting, tax, attest, security, and finance owners review the workflow against the real engagement and systems.