Let AI assemble the revenue memo; keep the accounting judgment human

A fluent memo is not accounting evidence

ASC 606 and IFRS 15 organize revenue analysis around the same five-step model: identify the contract, identify performance obligations, determine the transaction price, allocate that price, and recognize revenue as performance obligations are satisfied. AI can make the file easier to navigate. It cannot convert missing amendments, unsupported estimates, or unclear control transfer into a defensible conclusion.

The highest-value use is document control. A contract package can include a master agreement, order form, statement of work, change order, side letter, sales approval, price exception, delivery evidence, invoices, and a prior technical memo. An approved AI tool can extract candidate facts, cross-reference promises, identify inconsistent dates, and draft questions. Every output remains subordinate to the executed source and qualified accounting judgment.

The biggest risk is premature synthesis. A model recognizes familiar language and produces a coherent five-step narrative before the team establishes whether the documents are complete. Because the prose resembles a technical memo, reviewers may spend less time on the very items that drive the answer: enforceability, combination, modification, distinctness, variable consideration, standalone selling price, and transfer of control.

Start with a closed contract population and exact citations

Assign a source ID to each executed document and supporting record. Record version, effective date, signature or approval status, owner, and relationship to other documents. Then map every material fact to a section or page. A citation should let another accountant reproduce the fact without reading the prompt history.

Use source excerpts sparingly. Store the original under records policy and quote only what the memo needs. Do not paste confidential contracts into a consumer AI account. Use an enterprise-approved service with appropriate terms and access controls, or redact the record while keeping a controlled mapping to the original.

PCAOB AS 1105 describes audit evidence in terms of sufficiency and appropriateness, including relevance and reliability. AS 1215 addresses audit documentation. These are auditor standards, not instructions that automatically govern management's memo, but they reinforce a useful design principle: another experienced person should be able to understand the work performed, evidence obtained, and conclusions reached.

Use AI to prepare questions inside the five-step model

1. Identify the contract

Ask the tool to locate approval, rights, payment terms, commercial substance, collectibility facts, termination, renewal, and related agreements. A lawyer or qualified accountant resolves enforceability and contract-combination questions. A purchase order and an email concession may change the population even when the model only received the order form.

2. Identify performance obligations

Inventory every explicit and implied promise: license, hosting, implementation, migration, training, support, updates, warranty, options, and third-party items. Have the model produce a distinctness question set, not a conclusion. The accountant evaluates whether the customer can benefit from the good or service and whether the promise is separately identifiable in the contract context.

3. Determine the transaction price

Separate fixed consideration from usage fees, bonuses, penalties, refunds, rebates, credits, price concessions, financing, noncash consideration, and payments to the customer. Estimates need method, source data, update frequency, constraint analysis, and an owner. A historical average is not self-justifying; the facts may differ by customer, product, or period.

4. Allocate the transaction price

For each approved performance obligation, identify observable standalone selling price or an approved estimation method. Use the AI to structure evidence and create a calculation specification. Execute the calculation in a visible, versioned spreadsheet or controlled accounting tool, then independently review formulas, ranges, allocation, precision, and rounding.

5. Recognize revenue when or as obligations are satisfied

Map the human-approved obligations to evidence for over-time or point-in-time recognition and a measure of progress. Tie delivery, usage, milestone, elapsed-time, or acceptance evidence to the contract asset, contract liability, receivable, revenue, and disclosure mapping. The model should surface contradictions rather than select the most convenient date.

Worked example: SaaS subscription with implementation and support

Assume an executed one-year arrangement contains a hosted software subscription, a separately sold implementation service, and premium support. The fixed transaction price is $120,000. After qualified review, the team concludes that the three items are separate performance obligations. Approved standalone selling prices are $100,000 for the subscription, $20,000 for implementation, and $10,000 for support. This example illustrates arithmetic and workflow only; real conclusions depend on the actual agreement and applicable guidance.

The formula is allocated consideration = transaction price × item SSP ÷ total SSP. The workbook should preserve full precision, round only at the stated output stage, and force the allocated total to equal $120,000. Any rounding adjustment should follow policy and remain visible.

The difficult questions are not arithmetic. Is implementation distinct from the hosted service? Is support truly separate? Does customer acceptance affect control? Is any consideration variable? Does a renewal discount create a material right? Did sales promise a concession outside the signed order form? AI can ensure each question appears in the file and link it to a clause. The preparer and reviewer must resolve it.

independent workbook checks:
  contract_price = 120000.00
  total_ssp = SUM(subscription_ssp, implementation_ssp, support_ssp)
  allocation_each = contract_price * item_ssp / total_ssp
  allocation_tie = ROUND(SUM(allocation_each) - contract_price, 2) = 0
  cumulative_revenue_tie =
    opening_contract_balance
    + billings
    - cash_and_receivable_effects
    - closing_contract_balance
    - cumulative_revenue = 0
  journal_entry = human-approved memo conclusions + reviewed schedule only

Make the memo reviewable without the AI conversation

The final file should stand on its own. A reviewer should not need chat history to understand which contract versions were used, which facts came from the customer agreement, how amounts were calculated, where judgment entered, and who approved the conclusion.

1. Purpose and scope: entity, customer alias, arrangement, period, currency, framework, policy, and exact documents.
2. Executive conclusion: obligations, transaction price, allocation, pattern, balances, entry effect, and open limitations.
3. Material facts: precise contract citations, effective dates, approval status, conflicts, and later amendments.
4. Technical analysis: five-step model, relevant authoritative citations, alternatives considered, and basis for each judgment.
5. Calculation support: controlled workbook reference, input sources, formula review, precision, rounding, and tie-outs.
6. Operational evidence: delivery, usage, milestone, acceptance, invoice, subledger, and ledger support.
7. Judgment log: preparer view, reviewer challenge, consultations, changes, and final approver.
8. Entry and disclosure: proposed posting with accounts and periods, plus disclosure or control implications.
9. Appendices: source register, contract map, obligation matrix, allocation schedule, and exception log.

If the AI drafted prose, compare each material sentence with the evidence register. Remove unsupported citations and generic claims. Record meaningful human corrections, but retain prompt data only when company privacy, security, records, legal, and audit policy supports it.

Design the workflow around segregation and recalculation

PCAOB AS 2501 addresses auditing accounting estimates, including fair value measurements. Revenue estimates are not identical to every estimate in that standard, but the emphasis on understanding methods, data, assumptions, and company processes is operationally relevant. Make variable consideration and SSP estimates traceable to the method, data population, exceptions, approval, and period update.

Technology-assisted analysis does not weaken the evidence requirement. The PCAOB's 2024 amendments clarify auditor responsibilities when technology is used in audit analysis. Management should expect auditors to ask what information was used, how it was tested, and why evidence supports the conclusion. An AI-generated summary with no source lineage creates more review work, not less.

Failure modes to test before close

Implement on one bounded contract class

1. Choose a recurring, moderate-complexity arrangement with prior approved memos and no active dispute. Define what AI will and will not do.
2. Approve the tool, data classification, redaction, access, retention, and incident path with security, privacy, legal, and records owners as needed.
3. Build the contract source register and five-step template without AI first. Add the AI only to extraction, cross-reference, question generation, and drafting.
4. Create known-error tests: missing amendment, conflicting dates, hidden rebate, bundled setup, unsupported SSP, wrong acceptance date, and rounding imbalance.
5. Require the preparer to link every material fact and mark every unresolved item. Measure unsupported-claim rate and evidence completeness.
6. Run allocation and recognition calculations independently. Compare formula errors, tie-outs, correction time, and reviewer effort with the prior process.
7. Have a qualified reviewer blind-review a sample against the complete sources and applicable authoritative guidance.
8. Adopt only if evidence completeness and review quality improve without confidentiality, unsupported judgments, or posting-authority failures.
9. Reapprove after changes in contract type, policy, reporting standard, AI model, tool connections, data class, calculation method, or control owner.

An open-source reference discovered during this research, GAJETOso/financeskills, publishes an MIT-licensed revenue-recognition skill with contract analysis, allocation calculation, memo structure, and fixtures. This guide uses it only as a workflow design reference; the wording and controls here are independently written, and authoritative standards and company policy govern the accounting.

Frequently asked questions

Can AI decide the performance obligations?

No. It can inventory promises, assemble clause evidence, and prepare distinctness questions. A qualified accountant makes and documents the conclusion using the complete contract and applicable guidance.

Should AI calculate the allocation and journal entry?

It may draft the calculation specification, but an approved spreadsheet or accounting system should reproduce every amount with visible formulas and tie-outs. The proposed entry follows only after human-approved conclusions and review.

What evidence should accompany the memo?

The complete executed agreement and amendments, exact citations, approvals, price and estimate support, delivery evidence, calculation workbook, ledger ties, judgment and consultation log, reviewer changes, final memo, and approved entry according to retention policy.

Can contracts be pasted into a public AI service?

Not by default. Customer terms, pricing, signatures, personal data, and business commitments can be confidential. Use only an approved tool and data path, minimize inputs, redact where appropriate, and follow contract, security, privacy, and records requirements.

Does this replace ASC 606, IFRS 15, or professional advice?

No. It is a preparation and review workflow. Use the authoritative literature, company accounting policy, qualified professional judgment, and legal, tax, regulatory, audit, or valuation support when required.