Generative AI took another step forward with the arrival of Deep Research, and agentic AI is widely expected to be the theme of this year. It is worth thinking about what that actually does to an accounting function.

That the impact will be large is not in question. It will also be larger at big companies, where the scale economics work.

75% already

SMBC Group has reported automating 75% of its accounting operations, targeting 90% within the year. That is a startling penetration rate — and, read the other way, a demonstration of how far automation can actually go.

The reporting does not say which processes are automated, but at 75% it must extend across the function. Reporting, expense processing and depreciation — the recurring, rules-based work — are almost certainly in scope.

The parts I would want to ask about

Maintenance after a rule change. How far are tax filing, disclosure and audit support automated? Corporate tax and disclosure change every year, so any automation built around them needs to be re-fitted whenever the rules move. How that maintenance is resourced is the interesting question, and it is rarely the part that gets reported.

Prompt design. Where generative AI is used rather than pure RPA, everything depends on the instruction. Take a deliberately crude example: a model wired into the accounting system, told to “post the recurring month-end journals.” In February, does it post February’s entries or March’s? The system has no way to resolve that, and posting the wrong period is not a theoretical risk.

Once the technical detail behind cases like SMBC’s becomes clearer, I would expect this to spread well beyond a handful of large firms and into mid-sized and smaller companies.