Where Should CPA Firms Start With AI?

CPA firms should start with a recurring task that has a visible cost, clear review criteria, and a responsible owner. Client document intake is a strong candidate when incomplete files repeatedly stop preparation. Bookkeeping assistance may come first in a client accounting services practice. A specialist tax team may get more value from source-grounded research. The right starting point follows the firm's work, not a universal order of tools.

For a managing partner, the practical decision is which small workflow can improve service capacity without adding more verification work than it removes. That means comparing tasks across one service line, checking features already available in the firm's software, and budgeting staff time to review the proposed output.

Whether to use AI is no longer the open question for most firms. In Thomson Reuters' 2026 AI in Professional Services Report, organization-wide AI use across professional services almost doubled to 40% in 2026, from 22% in 2025, in a survey of more than 1,500 respondents in 27 countries spanning legal, tax and accounting, corporate and government work. Which work to point it at first is still each firm's decision.

See an illustrative client-document handoff with entity and period checks.

Use the reason work stops to choose the first use case

Start with recent engagements and record the actual reason for each interruption. Waiting on a client is different from staff searching a portal. A preparer blocked by an unclear transaction is different from a partner's review backlog. They can all appear as a late engagement while needing different remedies.

This decision table is a proposed selection method, not a ranking of returns available to every firm.

A firm can legitimately choose research first. Its value should be demonstrated on the technical work being done, including the time needed to check citations and conclusions. There is no reason to rebuild a working intake process simply to satisfy an AI roadmap.

What an intake pilot should actually deliver

Choose one repeatable engagement type, such as monthly bookkeeping onboarding. Agree the client entity, reporting period, required accounts, expected documents, and readiness owner. Use the existing portal for collection and the practice-management system for status where supported. Before building a separate pipeline, ask whether current checklist, request, and document features can resolve the problem.

The proposed output is a preparation packet: the approved checklist, received files, source-linked classifications, unresolved exceptions, and assigned preparer. It is not a certification that the records are correct or sufficient for every professional purpose. The administrator checks collection; engagement staff decide whether the material supports the work.

This fictional example shows why those responsibilities matter. A client uploads a May bank statement to a June engagement, plus a revised sales export with the same filename as the earlier version. The portal records two successful uploads. The AI-assisted view proposes May as the statement period and flags two versions of the sales export.

The June statement stays outstanding. The administrator confirms the mismatch and approves a specific follow-up. The bookkeeper identifies which sales export should be used and retains the superseded version. The engagement does not advance merely because the number of files matches the number of checklist items.

A useful handoff also distinguishes documents received from access granted. A client can send every statement while the team still lacks access to the approved accounting system. Record that as a separate dependency with an owner rather than treating the preparer's inability to begin as an AI extraction failure.

Compare capacity in the weeks when it matters

The following capacity calculation is illustrative; its inputs are assumptions, not a firm result. Suppose an administrator handles 80 document packets in a month at 18 minutes each, including current rework. That is 24 hours. In the proposed process, every packet needs 7 minutes of review, and 20% need another 10 minutes to resolve an exception.

Expected handling would be 80 × (7 + 0.20 × 10) ÷ 60 = 12 hours, releasing 12 hours of capacity before training and ongoing administration. If routine review takes 13 minutes instead, expected handling rises to 20 hours and the capacity difference falls to 4 hours. The business case is sensitive to the work that remains.

Replace these assumptions with observed times. Count correction, supervision, duplicate-system entry, and maintenance. Check whether the capacity appears during a deadline week or a quiet period, and whether the next team can use it. Available hours are not a payroll reduction, additional accepted engagements, or realized cash savings until those outcomes occur.

Approve the data use before connecting client files

In its August 2026 security reminder, the IRS says tax and accounting professionals must maintain a written information security plan, including attention to service-provider safeguards. A new AI connection belongs in that review.

The same IRS release explains why. Under the Gramm-Leach-Bliley Act, tax and accounting professionals are treated as financial institutions that must implement a data security plan; the FTC expects firms to select service providers that can maintain appropriate safeguards and to require them by contract; and covered institutions must generally report security events affecting 500 or more people to the FTC within 30 days of discovery, a requirement the FTC added to its Safeguards Rule in 2023.

Have the firm's security and professional-practice owners examine the actual service terms, approved uses, access permissions, retention, subcontractors, incident handling, and any required client communication or consent. A promise that data is not used for training answers one question; it does not settle all the others. Test early configurations with synthetic or appropriately de-identified material.

Put these questions to each provider in writing:

The CPA.com AI resource center includes provider due-diligence and build-versus-buy resources. Use that material alongside your own requirements; a vendor's presence in an industry resource does not establish suitability for your engagement.

For work that reaches a client, identify the reviewer and what they must verify. Technical conclusions need the appropriate authority, context, and professional analysis. Calculations should be checked in the relevant accounting or tax system. An AI-generated memo, reconciled-looking table, or completed checklist is an intermediate output.

Make the first decision small enough to reverse

Give the pilot one service-line owner and a defined end point. Compare total handling time, wrong-entity assignments, missed document gaps, and files returned by preparers. Review some outputs the system labels complete, because a queue of flagged exceptions will not reveal everything it missed. Stop if confidential material reaches the wrong engagement or the process silently loses a file.

Expand only when the team can operate the workflow, explain its limitations, and demonstrate an improvement worth its ongoing cost. If the trial shows that client responsiveness is the constraint, revise the collection process. If it shows that partner review is full, deal with that capacity before accelerating more preparation work.

The accounting and professional-services guide shows how collection, preparation, and professional handoffs connect. The starting point for your firm is the task whose improvement survives the full review process and helps the team deliver the engagement.