What AI Can Take Off an Accounting Firm’s Plate, and What Still Needs an Accountant

A client’s books still need an accountant’s judgment. Yet the path to that judgment can involve hours of chasing invoices, sorting documents and entering familiar transactions. For US accounting firms short of staff, those hours are becoming harder to spare. AI can help recover capacity when a firm identifies the steps that follow clear rules and keeps a person responsible for exceptions, advice and final approval.

The staffing shortage makes routine hours harder to absorb

According to an analysis by Sequoia Capital, the US has lost roughly 340,000 accountants over the past five years while demand has continued to grow, and around three in four CPAs are approaching retirement. Bloomberg Tax has also reported a shortage of roughly 340,000 CPAs. The figures point to a problem that reaches beyond recruitment: firms have fewer experienced people available to handle a growing volume of work.

For a small or mid-sized firm, the shortage shows up in everyday scheduling. A qualified accountant who spends the morning locating documents has less time to review a difficult entry or discuss a client’s tax position. Hiring seasonal staff or sending work offshore can add capacity during busy periods. Neither approach, on its own, reduces the volume of routine work arriving with each client file. That is where a change in the workflow can make a difference.

Where the hours go in a typical workflow

Consider what happens before an accountant can review a month’s activity. Someone requests missing documents from clients, follows up when they do not arrive, opens the files that do arrive and checks which client and period they belong to. Recurring transactions are posted and categorized. Staff then reconcile records against statements, investigate mismatches and tell clients what is still outstanding. These tasks differ in difficulty, but many consume time because information must move between an inbox, a document folder and an accounting system.

A client may send an invoice without the context needed to process it. An employee finds the right record, checks whether the invoice has already been received and enters details that follow a familiar pattern. The same employee may repeat that sequence for the next client. Mapping that path matters more than simply adding an AI tool to the inbox: it reveals which decisions can be specified in advance and where professional review belongs.

Where rules end and judgment begins

A missing supplier invoice offers a useful test. A system can request the document, read the fields it needs, match the invoice to the right client and period, propose a category and post an entry when defined checks pass. Some steps may use AI to read varied document formats; others are ordinary rules and software integrations. The firm needs to decide what counts as a valid match, when a posting is allowed and what should happen if the invoice conflicts with an existing record.

Uncertainty should stop the automatic path. A questionable amount, an unfamiliar supplier or a possible duplicate goes to a person instead of being quietly accepted. The system should log what it read, which checks it applied, what it did and who resolved an exception. That record lets staff review decisions without reconstructing them from scattered messages.

Unusual transactions still call for an accountant’s judgment. So do tax advice and the review and signing of returns. Automating document handling can give accountants more time for those responsibilities, but the handoff must be designed into the process. Without a clear handoff, a fast system can leave the firm with more entries to untangle rather than less work.

Two workflows, rebuilt before they were automated

At an accounting firm in Portugal, collecting supplier invoices from clients used to take around 30 hours of staff time each month. Klevie built a system that requests the invoices, reads and classifies them, and posts them. Accountants now review the exceptions, while the collection work takes under two hours a month. The build took six weeks. Those figures describe that firm’s workflow, rather than a promised result for every practice.

A second example comes from an insurance brokerage. Handling documents for each new client took around eight hours. After the process was redesigned and document reading and checks were automated, the work took about one hour per client. The industries differ, but both teams faced the same operational problem: staff had to collect information, check it and move it through a sequence of systems before a specialist could act on it.

In both cases, the process was mapped before anything was automated. The teams identified steps that existed only because people were doing the work manually and removed them rather than asking software to imitate them. For an accounting firm, that might mean examining how many times a document is copied or rechecked as it moves between systems. A shorter, clearly defined process gives both the technology and the people reviewing its output a better starting point.

Two things worth getting right before relying on it

Client documents demand careful handling. The system should run in an environment the firm controls, with role-based access so staff see only what their work requires. Client data should not be used to train outside AI models. Every automatic decision needs a log, including the information used and any subsequent correction. Those requirements belong in the design before invoices begin moving through the system, not in a policy drafted after launch.

Trust takes a separate test. Run the new workflow alongside the current one for a few weeks and compare its requests, classifications and proposed postings with what staff would have done. Review the exceptions as well as the straightforward cases: a system that handles familiar invoices quickly but repeatedly misroutes unusual ones still creates work. The parallel run gives accountants a chance to see how decisions are made and lets the firm adjust its rules before relying on the output in daily operations.

Two ways to run the system

A firm can have a system built around its own process and train its team to operate it. That is how the Portuguese accounting firm works: its accountants run their own platform and review the cases that need attention. This approach suits a practice that wants to keep day-to-day control of the workflow and has people who can take responsibility for operating it.

The alternative is to hand routine processing to a provider that runs the system each month and delivers the results. The firm’s accountants still review the work and sign off. Klevie, a business and technology consultancy based in Porto and New York, offers both arrangements: firms can work with it to design and run AI systems within their own operations or have the consultancy carry out the recurring work. The choice depends on who should manage the daily process, not on whether the firm retains professional responsibility.

A five-minute exercise to find the hours

Start with the work already happening. Write down five tasks the team repeats every week. For each one, note who does it, how long it takes and what information must be checked before the task is complete. Include work that may be easy to overlook, such as following up on a missing document or sending a client an update. Then look at the time recorded and ask how many of those hours required a qualified accountant. The answer will not make every remaining step suitable for AI. It will show where the firm should examine its process first, which routine decisions need clear rules and where an accountant’s attention is being used because the workflow gives them no alternative.