Practice guide · AI and professional judgment
AI for accounting firms: it prepares, you decide
An AI assistant can help a CPA firm organize a close review and draft an explanation. The useful question is how its output becomes supported work, with the professional still responsible for the conclusion.
The short answer
AI for accounting firms can prepare analysis, suggest questions and draft review commentary. A useful control separates source data and reproducible calculations from generated interpretation, then requires professional review before approval. AI output is not evidence that an account is correct, a transaction complies with a standard or a proposed adjustment should be posted.
Where AI helps a close review
A reviewer often starts with balances, movements and supporting documents that do not yet form a clear explanation. An assistant can organize the available facts, identify an unanswered question and propose a first draft. Its value is easier to assess against a defined task, such as explaining a receivables movement, than against an open instruction to “review everything”.
A draft can distinguish the opening balance, new invoices, cash received and closing balance, or suggest asking whether an unusual expense is recurring. That proposed structure helps the professional investigate. It does not establish the cause of the movement or replace an invoice, bank record or contract.
NIST’s AI Risk Management Framework provides a voluntary way to consider AI risks in their context. Its generative AI profile identifies confidently expressed false content and excessive reliance as relevant concerns. These are reasons to build a review process around the assistant, rather than treating fluent language as assurance.
Automation and AI answer different questions
A calculation should be reproducible without asking the model for the same number again. Keep its inputs, date convention, formula and rounding available. Account classification also needs an identifiable rule or mapping. Describing every software action as AI hides the controls a reviewer needs to inspect.
Generated commentary belongs to a different layer: it may summarize, interpret or ask a question. A convincing paragraph with a wrong figure still fails. A correct figure with an unsupported explanation also fails. Review both the arithmetic and the proposed interpretation.
| Work | Appropriate basis | Review question |
|---|---|---|
| Allocate a service over a documented period | Reproducible dates and arithmetic | Are the period and convention supported? |
| Map accounts into review cycles | Explicit accounting rules and configuration | Does the mapping suit this chart? |
| Explain a movement | Generated draft using supplied facts | Does the evidence support the proposed cause? |
| Approve a conclusion or adjustment | Professional judgment and engagement policy | Who reviewed the evidence and authorized the decision? |
Scroll the table if needed.
Prepare, examine evidence, then approve
Human oversight works when the professional can reject a proposed cause and reopen the question. A compulsory “approve” click after an opaque answer is a weak control. The reviewer should know what was available to the assistant and what evidence was obtained afterward.
An exchange history or draft can help explain how a note was developed. It is not the underlying accounting evidence. Preserve the relevant records and the professional’s final rationale through the firm’s documentation policy, rather than relying on the wording of a conversation.
Define the task
Select the cycle, period, reporting basis and question. Use the records actually relevant to that work.
Inspect the proposal
Trace each figure to the workpaper or source. Identify missing information and distinguish observations from assumptions.
Resolve the conclusion
Correct the draft, obtain the missing document or leave the issue open. Record the conclusion in the engagement’s normal review process.
Authorize the final work
Use the designated reviewer and approval rules. A prepared narrative is not a completed approval or an instruction to post.
Fictitious example: a $12,000 revenue movement
A fictional company reports monthly revenue $12,000 higher than the previous month. The workpaper confirms both totals. An assistant proposes that a new annual contract explains the increase. No contract has yet been supplied, so that causal statement is an assumption, even though the subtraction is correct.
The reviewer obtains the invoice and service dates. The invoice relates to a service covering several months, and the existing revenue schedule needs to be reconciled. The conclusion becomes a question about timing and the schedule, rather than “new recurring revenue”. The professional corrects the commentary and checks whether an adjustment is appropriate.
A useful evaluation asks whether the assistant can acknowledge the missing contract, preserve the verified $12,000 movement and frame the open issue clearly. This example is a test scenario, not an observed customer result or a promise that an assistant finds every cut-off issue.
Client confidentiality starts with the actual data sent
Inventory both the automatic context and the user’s own question. A carefully limited balance extract can be accompanied by a free-text question containing a client name or a confidential transaction. Attachments and support messages need the same attention. A general claim that information is anonymous does not describe those routes.
Ask which provider processes the request, where that processing takes place, what contract applies to retention and training, and how those terms are reflected in the service configuration. Separate that review from the website’s hosting and the application’s other service providers.
The firm assesses the engagement terms, client instructions and applicable professional or legal requirements. When those conditions do not permit an external AI service for a particular engagement, do not use it for that file. An assistant’s availability is not permission to transmit every record.
Ten questions to evaluate an AI review assistant
| Criterion | Question to ask | Warning sign |
|---|---|---|
| Scope and trigger | What starts the work and which cycles are included? | Undefined scope |
| Write permissions | What can be changed, proposed or approved? | Preparation confused with final approval |
| Data sent | Can we inspect context and free-text inputs? | Anonymity claimed without a field inventory |
| Provider terms | Which processor, region and contractual terms apply? | A consumer policy substituted for service terms |
| Numbers | Where does each figure come from? | An untraceable generated amount |
| Documentation | What proposal and final conclusion can we retain? | A conversation presented as evidence |
| Approval process | Who can approve or reopen the work? | The model controls final approval |
| Language | Can the draft use our terminology and register? | Confident wording hiding uncertainty |
| Rules versus AI | Which results are deterministic calculations? | Every action described as intelligent |
| Engagement choice | Can the firm choose when to use the assistant? | Use assumed for every client |
Scroll the table if needed.
AI preparation in Close Review
Close Review offers autonomous preparation and assisted drafting. Autonomous preparation works through the review cycles, prepares supporting analysis and commentary, and brings the work to the reviewer. In assisted mode, the professional uses the assistant while examining a cycle. In both modes, final review approval remains human.
The published security information describes US-only AI processing and the provider’s written commitment to zero retention and no use for model training. Assess that processing arrangement against your engagement requirements. A demonstration using fictional data can connect the draft to its source figures and show the approval sequence.
Common questions
Can AI replace the accountant responsible for a close review?
Preparation and drafting support the work. They do not transfer responsibility for evidence, accounting judgment or final approval to a model.
Can an AI explanation include a wrong number?
Yes. Trace every number to a source or reproducible calculation. Correct arithmetic also does not prove that the explanation of a movement is true.
Is an automatic prorata calculation AI?
Not necessarily. A dated allocation can follow explicit rules and arithmetic. Generated commentary about that allocation is a separate activity.
Is the assistant’s conversation accounting evidence?
It documents a proposal or exchange. The conclusion still needs the underlying contracts, records, calculations and professional assessment.
Does human approval mean nothing is prepared beforehand?
No. Work can be prepared before review. The control is that a professional examines and approves the final conclusion, rather than treating preparation as approval.
When should a firm avoid using an AI service for a client?
When engagement instructions or applicable requirements do not allow the proposed data processing. Resolve that question before transmitting the file’s information.
Sources for further reading
Dated source checks do not replace review by the professional responsible for your engagement.
Let’s talk about your firm
A conversation about the way you work.
Tell us who you are, what interests you and what you want to organize. A few details are enough to start.
WhatsApp opens an external service. Use your business contact details only, without client information.
Talk on WhatsApp