Reading the research without the panic
Accounting regularly appears near the top of task-exposure studies, including the ILO's work on generative AI and jobs, because so much of the occupation's task list is structured information processing. But the same research draws a distinction that headlines drop: high exposure mostly means high transformation potential, with tasks augmented or reorganized, not a simple forecast of job loss.
O*NET's description of accountants and auditors makes the split visible. Some listed tasks are exactly what generative AI handles well, such as preparing standard reports and examining records for accuracy. Others, such as advising management, assessing financial operations and taking responsibility for compliance with law, depend on context, authority and trust that a draft cannot carry.
What automates first, and what it leaves behind
The near-term automation frontier in accounting is the first pass: transaction classification, reconciliation matching, standard journal entries, draft variance commentary and the initial version of recurring reports. These share the pattern that makes tasks AI-friendly: digital input, structured output and a human checkpoint before anything is filed.
What the first pass leaves behind is precisely where accountants have always earned trust: deciding whether the numbers make sense, noticing the exception that does not fit the rule, and defending a treatment in front of a client, an auditor or a regulator. As drafting time shrinks, the share of the week spent on judgment, exceptions and communication grows, which changes what a strong performer looks like.
Controls become the leverage point
When AI systems draft entries and reports, the scarce skill moves one level up: designing and operating the controls around them. Someone must decide what the model is allowed to touch, how its output is sampled and tested, how errors are caught before filing, and how the whole chain stays explainable to an auditor. That is control design, and it sits naturally with accountants rather than with engineers.
This is also the safest ground to stand on professionally. Regulatory responsibility does not transfer to a tool. Firms adopting AI still need named people who sign, attest and answer questions, and those people will insist on colleagues who can validate machine-produced workpapers rather than merely accept them.
A realistic 90-day plan
First month: map one recurring close or reporting process end to end and mark which steps are first-pass drafting versus judgment. Introduce AI assistance on a single drafting step and keep a simple log of every correction you make; that log is your evidence of verification skill.
Second month: write a one-page control description for that AI-assisted step, covering inputs, review sampling, error handling and escalation. Third month: move one deliverable up the value chain, for example turning a standard variance report into a short scenario note for a decision maker. Advisory work is learned by doing small advisory pieces, not by waiting for the title.
