Why the ledger automates faster than the opinion
Will AI replace accountants? Not as a whole job. The work splits into producing numbers and standing behind them, and software is far better at the first. Preparing and examining accounting records, reconciling accounts and computing taxes owed all run on structured data that a machine reads cleanly. The second part is different. A tax return and an audit opinion carry a named license, a firm’s liability, and a judgment about what the evidence actually supports.
So the pressure lands on tasks rather than titles. Advising management on resource use, on tax strategy, and on the assumptions sitting under a budget forecast depends on knowing a client’s business and being willing to be wrong in public. A model can draft a position paper. It cannot be the party a regulator, a lender or a board holds responsible for it.
Scale matters here too. BLS counts 1,449,500 accountant and auditor jobs in the United States, median pay of $83,680, and projected growth of 5% from 2025 to 2035 (BLS, 2025). Steady headline growth does not mean the inside of the job stays the same. The first thing to thin out is the preparation ladder: the junior hours spent pulling data, tying out balances and formatting schedules.
That same preparation work is the core of the job done by bookkeeping, accounting and auditing clerks, which is why the two roles move at different speeds. If you want to watch the hiring end of this, our entry-level jobs tracker follows postings for starter roles.
What AI does, what it helps with, and what it leaves
Some of this work already runs with light supervision. Drafting routine financial statements from clean ledger data and computing tax figures for standard returns are the clearest examples. Share of task time in that group: 10%.
A larger part is shared work. Analyzing records for compliance gaps, inspecting account books against reporting standards and documenting recordkeeping systems all go faster with a tool that flags anomalies and drafts the write-up. The accountant still decides which exception is noise and which one is a problem. Share of task time there: 67%.
Then there is the part that stays with a person: 23%. Advising management on tax strategy and resource use sits in that group, and so does deciding whether the evidence gathered is enough to support an opinion and then signing it. Our coverage score for this job, out of 100, is 38. Coverage answers “Can AI do it?”, and the coverage method explains how task time is weighted.
What has actually been tested
No one has run a clean head-to-head test on this job. Our evidence grade for quality parity here is D. A D grade means the comparison has not been measured, so we publish no parity number at all. The quality parity method sets out what counts as a real test.
The best-known estimate is old and indirect. Frey and Osborne (2013) put the probability of computerization for accountants and auditors at 0.94, based on how routine the job’s characteristics looked on paper, not on any trial against working accountants. More than a decade later, employment in the occupation is still counted in the millions and is projected to grow (BLS, 2025). That gap is a reminder to treat task-level evidence as the serious kind.
What would settle the question is specific. Give models and licensed accountants the same files, including a messy reconciliation, a contested tax position and an audit sample, then have independent reviewers score accuracy and whether the reasoning holds up under challenge. Until that exists, the honest answer is that no one has measured it.
When the balance could shift
Most likely between 2035 and 2047 (8 in 10 of our scenarios). That window moves for reasons you can watch rather than guess at. Two things could pull it closer. Accounting data is already digital and structured, so nothing has to be invented in hardware for this job to change. And the running cost of machine review is a small fraction of licensed staff time, which pushes firms to try it on high-volume preparation first.
Two things hold it back. Sign-off rules and auditing standards expect a named, licensed person behind the report, and that requirement changes slowly. Liability also demands an evidence trail a reviewer can follow step by step, which general-purpose models do not reliably produce. Client confidentiality and data governance add another layer of friction inside firms. The replacement-year method explains what the range covers, and you can put this job next to another one on compare two jobs.
How to stay the person they call
Lean into the tasks the split leaves with people. Advising management on tax strategy, financing and resource use is the first. Designing, documenting and modifying recordkeeping and internal control systems is the second, because someone has to decide what the system should check. The third is exception work: investigating discrepancies, testing whether the evidence behind a conclusion is enough, and defending that call to a client or a regulator.
Two skills raise your floor. One is control over machine-produced numbers: sampling the output, testing it against source documents, and documenting what you checked. The other is plain client communication, because the value of advice is in the explanation, not the schedule. Our guide to AI skills employers want covers the first in more detail.
What to do: pick one recurring close or review task this quarter, automate the preparation, and spend the hours you free up on review and advice.
If you are weighing nearby work, the closest roles are Tax Preparers, Financial Examiners and Budget Analysts, each scored the same way on its own page. You can also see where this job sits among financial specialist occupations, read how employers are using these tools across accounting firms, or check how we score every job before you trust any of the numbers above.