Key findings

• Use AI to prepare and organize evidence, not replace professional judgment.

• Protect taxpayer and financial data with written, tested safeguards.

• Link every extracted value or anomaly to its source and review status.

• Integrate AI controls into the firm's quality management process.

• Do not automate sign-off, filing, release, or unsupported accounting conclusions.

Use cases for tax, bookkeeping, and assurance operations

Accounting firms can use AI to classify client uploads, extract proposed fields, check organizer completeness, summarize meetings, index workpapers, explain variances for review, draft information requests, route notices, and prepare anomaly queues. In bookkeeping, AI can suggest transaction categories or matches while a controlled workflow applies rules and review. In tax, it can organize source documents and identify missing items. In assurance, it can help navigate evidence but should not determine sufficiency or the conclusion.

Choose a workflow with authoritative source documents and a clear reviewer. Extracting a form field with a page reference is safer than proposing a tax position. Preparing a list of unreconciled items is safer than posting a journal entry. Drafting a client request is safer than submitting a return. The pilot should reduce document handling and search while preserving evidence, review, and segregation of duties.

• Good pilots: upload classification, extraction with citations, completeness checks, workpaper indexing, and request-list drafts.

• Higher control: coding suggestions, reconciliation support, anomaly detection, variance narratives, and tax research leads.

• Keep under professional authority: entries, estimates, positions, attest conclusions, filings, signatures, and releases.

A source-to-workpaper reference architecture

An intake service scans and stores the original document, assigns client and engagement identity, and calls an orchestrator. Retrieval supplies only the required pages and current firm guidance. The model returns typed fields, source coordinates, confidence, and exceptions. Deterministic validators check totals, formats, cross-footing, period, entity, and approved thresholds. Proposed values enter a review queue and are posted to the ledger, tax, or audit system only through approved roles and workflows.

The document repository, ledger, tax suite, audit platform, practice system, and client portal remain authoritative. Preserve the original evidence, extracted value, model and prompt version, reviewer correction, approval, and posting identifier. Enforce segregation so the same automation cannot prepare, approve, and release a material action. Use idempotency and reconciliation to prevent duplicate entries. Maintain manual processing for close, filing, payroll, and deadlines if the AI or integration layer is unavailable.

Financial and taxpayer data safeguards

Client files may contain Social Security numbers, bank details, payroll, tax returns, credentials, ownership, medical deductions, and confidential business records. Use field-level minimization, encryption, strong authentication, role and engagement access, logging, retention, secure deletion, and controlled exports. Development should use synthetic or properly de-identified data where feasible. Prompts, embeddings, outputs, and evaluation sets require the same classification as the underlying client data.

IRS Publication 4557 explains that protecting taxpayer data is required and points tax professionals to security plans and safeguards. The FTC Safeguards Rule specifically includes tax preparation firms among covered examples and requires covered financial institutions to maintain safeguards for customer information; its current guidance also addresses breach reporting obligations. Firms should determine exact coverage and responsibilities, maintain a written information security program where required, review service providers, and test incident response rather than relying on a vendor's marketing label.

Professional review and quality management

Qualified professionals remain responsible for applicable standards, evidence, estimates, materiality, skepticism, tax law, communication, and conclusions. Review screens should display the original source and proposed field together, with differences and missing evidence highlighted. Require approval before posting entries, changing a return, accepting audit evidence, issuing a report, releasing a filing, or sending a material client communication. Staff must understand that fluent explanations are not evidence.

AICPA describes its quality management standards as a proactive, risk-based approach tailored to the firm and its engagements. AI should enter that system as a technology resource and quality risk: define objectives, identify risks, design responses, monitor operation, remediate failures, and document evaluation. AICPA's AI resource center also emphasizes ethics, governance, risk evaluation, confidential data, and professional judgment. Firms should apply binding standards and licensing requirements relevant to their work and jurisdiction.

KPIs for accuracy, throughput, and quality

Track document turnaround, field-level precision and recall, exception rate, reviewer corrections, time per verified document, completeness, unmatched transactions, reconciliation aging, and time to a reviewed workpaper. For anomaly queues, measure useful findings, false positives, missed known issues, and investigation time. For client communication, measure response cycle and material revisions rather than drafts generated.

Pair operations with quality and security: reopened workpapers, post-release corrections, filing rejects, unsupported entries, missed deadlines, review-note volume, access exceptions, data incidents, failed restore tests, and time to contain a vendor issue. Commercial measures can include cost per completed workflow, realization, and capacity shifted to advisory work, but avoid fabricated savings or claims that AI improved audit quality without evidence. Segment by form, client type, engagement, document quality, and confidence band.

A 90-day accounting rollout

Days 1-30: choose one document type and engagement workflow, appoint professional and security owners, map client data, review vendor and service-provider controls, baseline time and errors, and build an evaluation set with poor scans, multiple entities, conflicting periods, handwritten changes, blank fields, unusual values, and malicious document text. Build a read-only extraction that cannot post, file, or send.

Days 31-60: run in shadow mode through normal preparer and reviewer roles. Compare each field to evidence, test totals and reconciliation, record corrections, and validate client boundaries, logs, deletion, backup, and incident procedures. Days 61-90: release a narrow verified workflow with mandatory sign-off and no autonomous posting. Train staff and include the process in monitoring. Expansion requires sustained field performance, acceptable review time, no unresolved safeguarding gap, and approval under the firm's quality management process.

• Day 30: approved source type, data map, owners, baseline, controls, and evaluation set.

• Day 60: shadow mode passes evidence, accuracy, access, segregation, and recovery checks.

• Day 90: limited production has preparer-reviewer control, monitoring, and no autonomous release.

When not to automate

Do not automate when source documents are unreliable, entity or period identity is ambiguous, engagement access cannot be enforced, or vendor safeguards are inadequate. Avoid AI output that cannot point back to evidence. Do not deploy during a filing deadline as a substitute for testing, training, and fallback. If review takes longer than the original task, narrow the workflow rather than pressuring staff to accept output.

Keep professional judgment, estimates, materiality, tax positions, audit evidence evaluation, independence, attest conclusions, signatures, filings, and report release with qualified people. Do not let one automation create and approve a journal entry or change bank details without established verification. Deterministic accounting rules, validation, and reconciliation are preferable for exact calculations. A manual process is appropriate for rare, novel, high-risk transactions where a representative evaluation set does not exist.

Sources and methodology

This article synthesizes the primary sources below as of the publication date. Forecasts and recommendations are directional scenarios, not guarantees; they should be tested against your workflow, data, risk tolerance, and current vendor documentation.

AICPA & CIMA: AI resources for accounting and finance (accessed 2026-07-30)

AICPA & CIMA: Quality Management (accessed 2026-07-30)

Internal Revenue Service: Publication 4557: Safeguarding Taxpayer Data (accessed 2026-07-30)

Federal Trade Commission: FTC Safeguards Rule compliance guide (accessed 2026-07-30)