CPA AI Workflows: Research, Review, and Client Preparation

CPA AI workflows use an assistant to prepare research, organize review work, and draft communications while a qualified professional verifies the result. The useful boundary is preparation: AI can help you identify questions, but it cannot take responsibility for a tax position, accounting conclusion, or signed report. fenbs keeps approved process rules and reviewed assumptions available across assistant sessions without becoming the home for private client records.

Start with a recurring problem you can measure, such as turning a public technical update into a list of issues for staff training. Define the permitted material, the expected output, and the review evidence. A shorter first draft is not a saving if the reviewer must reconstruct every assumption before using it.

Which parts of a CPA’s day are good starting points?

Choose work where you already know what a good answer contains. Drafting a general document request, organizing a research outline, and preparing a meeting agenda are easier to check than asking an assistant to solve an unfamiliar technical issue. Test with public or synthetic information before considering permitted live engagement data.

Keep the engagement type visible. A bookkeeping process, tax research assignment, and audit procedure do not share identical evidence requirements. The AI approval workflow guide helps separate drafting from authorization; the AI by role hub covers preparation in other professions.

How do you research an accounting or tax question with AI?

Write the question with its factual and legal boundaries before asking for an answer. For a tax question, include the jurisdiction, tax period, entity type, known facts, and what remains unknown. For an accounting question, name the reporting framework and transaction features. Missing context should become a question, not an invented fact.

Ask the assistant for an issue list and a source trail. Open every authority used for a material conclusion and read the relevant text in context. A citation can exist yet discuss a different period, taxpayer, or transaction. Check effective dates and exceptions rather than treating a familiar publication title as proof.

• State the proposed conclusion as provisional until a professional reviewer accepts it.
• Separate facts received, assumptions made, missing facts, and questions for the client.
• Record the authority, relevant passage, jurisdiction, effective period, and reason it applies.
• Identify alternative treatments or contrary authority that need investigation.
• Document what was checked and who reviewed it in the approved engagement record.

For federal tax practice, the IRS guidelines on responsible AI use↗ (opens in a new tab) discuss fabricated outputs, confidentiality, and professional oversight. Use the guidance as a reason to design a reviewable workflow, not as approval for a particular vendor or a conclusion that an AI answer is reliable.

For preparation workflows outside accounting, see AI research, AI in education, AI in advertising, AI for SEO, and AI for sales. For a CPA, the additional check is whether the particular authority answers this engagement’s question. Keep the source-check procedure in fenbs Decisions and rules, and leave a task open whenever the available evidence does not support a conclusion.

Keep the reusable research instructions in the shared work context, while engagement-specific analysis stays in the firm’s approved files. The distinction between project memory and an assistant’s temporary conversation is covered in what an AI brain is. In fenbs, people can record the research and review rules once for the next assistant to read.

What makes an AI document review useful rather than misleading?

Give the assistant a defined comparison, not permission to certify the file. For example, ask whether a synthetic close checklist contains an owner, supporting record, reviewer, and completion criterion for each step. The output should identify gaps and cite where it found them. “Looks complete” provides no review evidence.

Where your firm permits source-document processing, test whether the tool reads tables, negative signs, dates, and scanned text correctly. A misplaced decimal or a dropped page can make the entire comparison wrong. Reconcile extracted information to the original before using a summary, and preserve the distinction between a missing field and an actual exception.

Use deterministic accounting tools for calculations and reconciliation. An assistant can explain a formula or suggest why balances differ, but rerun the calculation in the approved system. Check the population, period, currency, sign convention, rounding, and exclusions. A plausible explanation for a difference is a hypothesis until supporting work establishes it.

Say a hypothetical monthly close package shows an expense category increasing. The assistant may propose seasonality, timing, or classification as questions to investigate. It should not select one explanation and draft it as management’s confirmed reason. Ask the owner, inspect the support, and revise the explanation only when the evidence warrants it.

Which US professional rules still apply?

An AI tool does not change the professional responsibility attached to your work. The AICPA Code of Professional Conduct↗ (opens in a new tab) addresses matters including competence, due care, confidentiality, and use of third-party service providers for its members. Apply the requirements relevant to your role, together with state board rules and the standards governing the engagement.

In a tax practice, information handling needs a separate decision before anyone uploads a document. The IRS client-data protection guidance↗ (opens in a new tab) explains that professional tax preparers must have security plans under FTC regulations. Approval to use an assistant for public research does not automatically authorize taxpayer data, employee records, or client financial statements.

Have the appropriate people review confidentiality, access, retention, provider terms, and any disclosure or consent requirements for the intended use. Redacting a name alone may leave a client identifiable through the remaining facts. Use synthetic examples for training and testing unless the firm has approved another approach.

Set a clear stop rule for technical uncertainty. If the assistant cannot locate support, if the facts conflict, or if the reviewer lacks the relevant expertise, escalate the question. An unanswered issue is a valid outcome. In fenbs, record the standing instruction to escalate, rather than saving a speculative answer as a settled process decision.

How should client communications and handoffs be reviewed?

Draft client explanations only after the underlying conclusion has been reviewed. Give the assistant an approved summary and ask it to explain the result in everyday language, list what the client must provide, and avoid promising an outcome. Then check that simplification has not removed a condition, exception, deadline, or qualification that matters.

Keep requests precise. “Please provide support for this amount” leaves the client guessing; a reviewed request should say which period and document are needed and how to send them through the approved channel. The assistant should not invent a filing date or payment instruction to make the message look complete.

A useful handoff distinguishes drafted, checked, accepted, and sent. Record who still needs to act and what evidence remains outstanding. History can help identify who changed the work record, but it is not the same as an engagement documentation system or proof that a professional review occurred.

How to do this with fenbs?

1. Create a free account and add a project for a non-sensitive internal procedure, such as public-source research training.
2. Add tasks for preparing the source outline, checking the authorities, and reviewing the instructions. Write the expected evidence in each task’s plan.
3. Record the firm’s approved process decisions in Decisions and rules: permitted information, required source checks, the reviewer, and when to stop and ask.
4. Connect a supported AI app using its guide on the Integrations page. MCP is a standard connection for an assistant to use an outside tool; ask it to read the project context before drafting.
5. Save Where we left off when you stop, distinguishing work drafted from work reviewed. Keep supporting client records and formal review documentation in their approved systems.

fenbs does not file returns, maintain a general ledger, perform an audit, or replace your engagement software. It is not the right fit for storing confidential client material without the firm’s approval. Its useful job here is keeping the permitted work instructions, decisions, tasks, and next steps consistent.

Review the first exercise with another professional. Could they identify the source, the assumption, the required reviewer, and the unfinished work without reopening a chat? If the answer is yes, try a second procedure. If not, make the record clearer before expanding the workflow or introducing more sensitive information.