AI Outperforms CPAs on Accounting Tasks but Still Needs Oversight to Close Books

What happened

Current AI models now outperform licensed certified public accountants (CPAs) on structured accounting tasks, both in speed and in accuracy, according to a Mercor study reported by The Decoder. The same reporting notes that this is a sharp change: about eighteen months earlier, the same class of models lagged far behind human practitioners.

The picture shifts on the harder test. On the more demanding APEX Benchmark, no model completes every task. The Decoder's conclusion is that, without human oversight, AI still cannot close the books on its own.

The findings separate two kinds of accounting work. Bounded, well-defined tasks are now within reach of automation, while the end-to-end close, which requires judgement across many linked steps, still depends on people to review and complete the output.

Why it matters

For leaders in finance and operations, the result signals that AI has moved from experimental to genuinely useful on defined accounting work. Organisations can reasonably expect models to take on more of the routine volume that currently absorbs skilled staff time, freeing those teams for analysis and exception handling. The capability gap that remains on the harder benchmark is the part that determines how the operating model should be redesigned.

The experience implications run in two directions. Internal finance teams become the people who supervise, validate and correct machine output, which changes the skills and workflow they need. Clients and stakeholders who rely on accurate, timely financial reporting will judge the result, not the tooling behind it, so any handover from people to models has to be designed so that errors are caught before they reach the customer.

The Renascence take

The headline invites a simple scorecard: AI versus accountants, with the machine winning on speed and accuracy. The more useful reading is about where the handoffs sit.

Most readers will focus on the win and miss the failure mode, which is the hand-off between a model that is right in isolation and a process that is only right when every step is checked. Behavioural economics would predict exactly this: people over-trust fluent, confident outputs, and oversight quietly erodes once a tool has performed well for long enough. A customer-obsessed operator should therefore design the review point deliberately, measuring where errors surface and who is accountable for catching them, rather than assuming supervision will simply happen. Start by mapping your close process into bounded tasks and end-to-end decisions, automate the first with confidence, and fund the second as a human-led service.