AI for small accounting and CPA firms: what to automate first
The Journal of Accountancy asked small firms what they actually do with AI, and the answers were not chatbots. They were intake forms, invoice approvals, cited tax research, and a client app built in an afternoon. This is that list in the order a small firm should take it, with what stays with the CPA.
A small accounting firm automates in this order: client intake and document collection, transaction coding and reconciliation inside the software it already pays for, approval workflows that route invoices and signatures without a person forwarding email, and tax research that returns cited sources for the CPA to check. The Journal of Accountancy documented firms doing each of these in August 2026 (source). What stays with the CPA is the judgment: the position, the sign-off, and anything on a return.
Where the hours go in a small firm
A firm with five people loses its week in the same places every year: chasing clients for documents, coding transactions that follow the same rule every month, forwarding invoices for approval, rebuilding the same workpaper, and researching a question the partner half-remembers answering last spring. None of that is the work clients pay for. All of it is what a system can take, and the surveys agree on the shape. The Federal Reserve's 2026 Report on Employer Firms found the most common AI uses among small firms are writing, individual productivity, and planning or analysis, and Intuit's 2026 AI Impact Report found adoption highest in routine work and lowest where judgment is essential.
The Journal of Accountancy's August 2026 survey of small firms is more useful than either, because it names the firm and the task. Four of its examples are the spine of this article.
The order to take it in
1. Intake and document collection
Start where the client meets the firm. Agate CPA built AI-powered web forms with Lovable and Claude to gather prospect details before the first call, and reported conversion up by about 25 percent (Journal of Accountancy). The same shape works for document collection: a form that asks for exactly what this client's return needs, chases the gaps on a schedule, and lands the files where the preparer works. The hours it returns are the ones spent on "did you get my 1099" emails.
2. Coding and reconciliation, inside what you already pay for
Transaction coding and bank reconciliation follow rules a bookkeeper could write down, which is the definition of automatable work. Before buying anything, turn on what the ledger software already does: rule-based categorization, matching, and recurring entries. The gain is not glamorous and it is weekly. What the software cannot do is decide the ambiguous transaction, and the setup should route those to a person instead of guessing.
3. Approval workflows
Public Trust CPA built an approval process in Power Automate: vendors email invoices, the system triggers an Adobe Sign request, and an approved invoice becomes a QuickBooks entry with an audit trail (Journal of Accountancy). That is the pattern for every hand-off in a firm: a trigger, one action, an exception path, and a log. Nobody forwards an email and nobody wonders whether it was approved.
4. Tax research with citations
One Stop CPA uses BlueJ for AI-assisted tax research that returns cited sources, then applies CPA judgment, and reports advisory work that took hours now taking about two (Journal of Accountancy). The word that matters is cited. Research without a source the CPA can open is a draft of an opinion, and the firm's name goes on the opinion.
5. The client-facing layer
High Rock Accounting used Claude Code to build a client satisfaction app integrated with its Karbon practice management system in four to five hours, replacing a third-party tool (Journal of Accountancy). Small custom software is now within reach of a firm that has one technically curious person. The rule is the same as for every build: one process, one gap, an owner.
What stays with the CPA
Brian Davis of One Stop CPA gave the line to the Journal: AI is a starting point, not the final answer, and the approach works because it combines AI speed with CPA judgment. In practice that means the system prepares and the CPA decides, on every item that carries the firm's signature.
- The tax position. Research returns sources. A CPA reads them and takes the position.
- Anything on a return. Drafted by the system, reviewed line by line by the preparer, signed by a person.
- The ambiguous transaction. Routed to a person with the reason attached, never coded by best guess.
- Client advice. A first draft is fine. The advice is the CPA's.
Find the process that eats the week.
The ten questions ask what gets retyped, what gets chased, and what crosses systems. For a small firm the answer is usually intake, coding, or approvals, and the result says which comes first.
Two rules before any client data moves
First, decide what leaves the building. Client financial data is confidential under professional standards and often under the engagement letter, and a consumer chatbot is a third party. Use tools whose terms exclude your content from training and that state where data is stored, and write a policy that names the permitted tools and the prohibited ones. Second, decide what the system may not choose. In a firm that is short: nothing that goes on a return, and nothing a client will act on, without a CPA's review.
What it costs
Most of the list above is a setting or a subscription. Published guides put software at $50 to $300 a month per system and API usage at $100 to $500 a month when a tool calls a model on your data (AIessentials). A custom workflow, like the approval chain or the client app, runs $5,000 to $8,000 when it connects tools with APIs (Layer3 Labs), or an afternoon if the firm has someone who can build it. Set against the hours in the first section, most of it pays back inside a season.
What we have and have not built
We have not built a system for an accounting firm, and this article does not claim otherwise. It is written from what the firms in the Journal's survey reported and from the pattern in our own work: a repeated process, a trigger, one action, an exception path to a named person, and a log. Our dental payment integration is that pattern applied to a ledger, and the approval workflow above is the same pattern applied to a firm.
How to start
- Pick one of the five. Whichever cost the most hours last quarter.
- Turn on what you already pay for. Categorization rules and matching in the ledger software come first and cost nothing.
- Write the two rules. What leaves the building, and what the system may not choose.
- Buy before you build. Intake forms and approvals have products. Build only where the gap is specific to the firm.
- Measure the hours. The count you made in step one, again, a month later.
The firms that get somewhere with this are not the ones with the most tools. They are the ones that named one process and kept the CPA's name on the judgment.
What to hold on to
- Five jobs, in order. Intake and documents, coding and reconciliation, approvals, cited tax research, and the client-facing layer.
- Named firms did each one. The Journal of Accountancy documented intake forms, an approval chain into QuickBooks, cited research, and a client app built in an afternoon.
- Turn on what you already pay for first. Categorization rules and matching in the ledger software are the cheapest hours you will ever get back.
- The CPA keeps the judgment. The position, anything on a return, the ambiguous transaction, and the advice.
- Two rules before client data moves. What leaves the building, and what the system may not choose.
Questions owners ask us
What should a small accounting firm automate with AI first?
Client intake and document collection, then transaction coding and reconciliation using the rules in the ledger software the firm already pays for, then approval workflows, then tax research that returns cited sources. Each one repeats every week, follows a writable rule, and returns hours the firm does not bill.
How are small CPA firms actually using AI?
The Journal of Accountancy's August 2026 survey named four: AI-powered intake forms that lifted conversion about 25 percent, an automated invoice approval chain that lands entries in QuickBooks with an audit trail, cited tax research that cut advisory work to about two hours, and a client satisfaction app built with Claude Code in four to five hours.
Can AI prepare tax returns?
It can draft and it can research with citations. It cannot sign, and it should not take a position. The pattern the surveyed firms describe is AI speed plus CPA judgment: the system prepares, the preparer reviews every line, and a person signs. Ambiguous items route to a person rather than being coded by best guess.
Is it safe to put client financial data into AI tools?
Only into tools whose terms exclude your content from training, state where data is stored, and are named in a firm policy. A consumer chatbot is a third party under the engagement letter and professional standards. Decide what leaves the building before the first tool, and write the permitted and prohibited lists down.
How much does AI automation cost for a small accounting firm?
Most of it is a setting or a subscription: software at $50 to $300 a month per system, plus $100 to $500 a month in usage when a tool calls a model on your data (AIessentials). A custom approval chain or client app runs $5,000 to $8,000 to build across tools with APIs (Layer3 Labs), or an afternoon if someone in the firm can build it.
Has Prometheus built AI for an accounting firm?
No, and this article says so. It is written from what named firms reported to the Journal of Accountancy and from the pattern in our own shipped work: one repeated process, a trigger, one action, an exception path to a person, and a log. Our dental payment integration is that pattern applied to a ledger.
Sources
The firm examples are quoted from the Journal of Accountancy; survey figures from the reports; cost bands from published 2026 guides. Checked September 5, 2026.
- Journal of Accountancy, "Real-life ways small firms use AI". August 1, 2026. One Stop CPA, Agate CPA, Public Trust CPA, and High Rock Accounting.
- Federal Reserve Banks, "2026 Report on Employer Firms". Most common AI uses among small employer firms.
- Intuit, "2026 AI Impact Report". May 12, 2026. Adoption by type of work.
- AIessentials, "AI Consultant Cost (2026)". Updated August 21, 2026. Software and usage costs.
- Layer3 Labs, "AI Consulting Rates and Pricing in 2026". Updated June 17, 2026. Custom workflow bands.
- Dental payment integration, the pattern applied to a ledger.