AI in accounting: what early field research says about where the time actually goes

Most pitches for AI in accounting promise something close to "the software closes your books for you." What field research on teams that have actually adopted generative AI shows is narrower and more useful: a redistribution of effort, not a disappearance of the function. A study published in the Journal of Accounting Research, Human + AI in Accounting: Early Evidence from the Field, documents that GenAI adoption is associated with productivity gains alongside a systematic shift of effort away from routine data entry and toward business communication and quality assurance.

Where the effort actually moves

The finding worth acting on is the direction of the shift, not just its existence: less time on manual entry and matching, more time on review, exception-handling and communicating with clients or internal stakeholders about what the numbers mean. That is a different outcome from "fewer bookkeepers" — it is closer to the same headcount doing a different mix of work, with the highest-judgment tasks getting more attention because the lowest-judgment ones take less.

Where to start if you run finance for a growing company

The highest-volume, lowest-judgment task in most finance functions is bank reconciliation and invoice matching — exactly the kind of repetitive comparison an agent handles well, with clear pass/fail criteria and a natural audit trail. Starting there, rather than with judgment-heavy work like tax strategy or financial forecasting, mirrors what the research documents: automation concentrates where the decision space is narrow, and the freed time goes to the parts of the job that actually need a person.

• →Start with bank reconciliation and invoice matching — high volume, narrow decision space, easy to audit.
• →Keep a human validating any entry above a defined amount or exception threshold, rather than full auto-posting from day one.
• →Measure the change as a reallocation of time across tasks, not as a headcount target — that is what the research actually found, and it is the more honest ROI framing for a client or a board.