Eighty percent is a rounding error in marketing and an exception list in accounting

Gartner expects global AI spending to rise 46% in 2026, a figure cited in a CPA Practice Advisor column arguing that finance functions should spend that money differently from everyone else ^[1]. The argument rests on one number: in a test of leading language models on accounting tasks run earlier this year by a CFO, the best performer, Claude Opus 4.7, completed 80% of the tasks, and every other model did worse ^[2].

Eighty percent is a perfectly good hit rate in a function where the cost of being wrong is a wasted impression. In a general ledger it is twenty wrong answers per hundred ^[10], each of which someone has to find, explain and sign off. The column puts the same point as a hiring question: would you turn your books over to an accountant who makes a mistake on 20% of decisions, and is an AI-generated fix worth a 20% chance of being wrong ^[6].

The compounding is worse than the headline rate. If each step of a ten-step close succeeds 80% of the time and the steps are independent, the full chain runs clean about 11% of the time ^[11]. Real workflows are not independent, and the arithmetic is a bound rather than a forecast, but the direction holds: per-task accuracy that reads respectably in a benchmark table degrades quickly once tasks are strung together and each one inherits the last one's output.

Hence the column's prescription, which it calls point, don't fix: use AI to analyse large volumes of data and identify problem areas for humans to address ^[3]. The author's example is a new client whose books were run through an AI system and produced, within minutes, a flagged $785,000 discrepancy in reported net income that had gone undetected across two separate accounting systems ^[4]. Books that looked clean needed a full account-by-account rebuild before they were reliable for tax or compliance, and the author is explicit that the system did not replace expert review, it indicated where to look ^[5].

The operationally useful part is the review design. The column argues that some AI output can be checked by junior staff while other output requires CPA-level oversight, with the split set by risk, dollar value and confidence score, which it presents as a structured alternative to hoping employees catch errors ^[7]. That converts an accuracy problem into a routing problem, which is the only version of it a firm can actually staff. The column also claims a second-order benefit: when experts define context, set parameters and refine outputs, that knowledge becomes part of how the system performs ^[9].

The pressure runs the other way on price. The same automation that saves hours lets firms cut headcount and salary cost, and pass that through as lower fees to win work ^[8]. Firms holding a human review layer will be bidding against firms that quietly removed it, and buyers cannot see the difference until an exception surfaces.

Treat the 80% figure as directional, not settled. The column describes the test only as conducted earlier this year by a CFO and does not give the task count, the test design or published results ^[12]. Worth watching: whether vendors start publishing per-task completion rates and calibrated confidence scores that a routing rule can actually consume, and whether the 46% spending increase ^[1] lands in detection tooling or in autonomous bookkeeping sold on price.