Dentsu AI clears 99％ on 20万件超 approvals

Digitizing a purchase request does not by itself remove the human work behind it. Someone still has to match the order amount against the quotation, confirm it follows internal rules, and decide how to treat exceptional transactions. Once paper and seals became an approval button on screen, that checking work remained — and the more requests came in, the more of it there was to do.

Dentsu Group's answer was to put generative AI into the approval step of its accounting system. The starting point was not smooth: in the early stage of testing, human and AI verdicts agreed only 15.4％ of the time. As the team changed models and adjusted prompts, it also found that the human approval side had points worth revisiting, not just the AI.

The larger question is where the line between AI and human judgment should sit. Dentsu says it reached 99％ accuracy on the targeted work, yet did not leave the decision to AI alone — a sign that for this kind of checking, the value of the AI may lie as much in how it reshapes the human review around it as in the accuracy number itself.