Will AI Replace Tax Jobs? The Roles Gone by 2030, and the New Ones

Preparation is going and accountability is staying. The tax roles unlikely to survive to 2030, the ones being created, and what people working in tax today should do about it.

Someone who joins a tax team as a graduate this year will be close to manager by 2030. I cannot make their four years look like the four I spent when I started my career. Most of what I did then, a machine does now.

But before I start, I would like to add a disclosure. I work in consulting where I implement tax technology, and I build tax software of my own. I have an interest in the future I am describing, so the sources are linked.

The Pace

The research group METR measures how long a task an AI model can complete on its own. For recent models that length has been doubling about every 131 days. The work is moving from answering a question to carrying out several hours of work and bringing it back for review. That is a what a junior used to do.

The employment data presents a more nuanced picture.

• The World Economic Forum's Future of Jobs Report has employers expecting 92 million jobs displaced and 170 million created by 2030. Accounting clerks, accountants and bookkeepers are all on its list of fastest-declining roles.
• Anthropic's study of exposed occupations finds no rise in unemployment among experienced workers.
• Stanford finds employment of 22 to 25 year olds in AI-exposed occupations 19% below where it would otherwise be, through reduced hiring and not layoffs.

So the people are not being walked out. But the seats are not being refilled. For a profession built as a pyramid, that is the bigger change.

Tax Technology and the Tax Authority

Inside tax, tools that answer have become tools that act. EY has put 150 AI agents alongside 80,000 tax professionals. Thomson Reuters reports 34% of tax firms using generative AI firm-wide, up from 21% a year earlier, though only 19% of them measure the return.

The authorities are moving as fast. The OECD found 69% of tax administrations using AI, up from 9% in 2016, mostly to detect evasion and select cases. HMRC plans to prepopulate returns from third-party data and is hiring 5,500 more compliance staff. The UAE makes structured e-invoicing mandatory for larger businesses from January 2027.

When the authority sees each transaction as it happens and picks cases by algorithm, the valuable work moves from filing accurately afterwards to controlling data before it leaves.

What Is Slowing It Down

If capability were the only variable, the transition would be done by 2028. Three well-founded concerns buy time for the existing jobs.

• Data security. In a Sovos survey of finance leaders, 86% were very concerned about the security of data when using AI. Tax data is a complete picture of a business.
• Accountability. A model that is right 98% of the time is excellent at drafting and unusable as a signatory. The return is still signed by a person, because the organizations need a human to be accountable, at least for now.
• Cost. Prices per unit have collapsed, but agents that work for hours consume far more. Remember the news that Uber spent its entire 2026 AI budget in the first 4 months.

Gartner expects over 40% of agentic AI projects to be cancelled by 2027 on cost, value or risk.

Each of these concerns is somebody's future job.

The Roles I Do Not Expect to Survive

By "gone" I mean a sensible employer will no longer hire for the role as defined today. There are exceptions in the world even today where the accounts are still being prepared on the paper ledgers.

• The compliance preparer. Trial balance in, computation out, tied and referenced. The largest block of junior tax work will be fully automated. It’s already being automated at large scale today.
• The VAT return compiler. E-invoicing and prepopulated returns remove the need, with or without AI.
• The data chaser. Information requests, follow-ups and reformatting. There are tons of AI agents today (paid and open source) doing this work.
• The offshore processing seat. Labour arbitrage now competes with the cost of computation, and loses.
• The tax alert writer. Monitoring and summarising changes in the law. Our own AI Agent tracks the changes on a daily basis and updates the rules.
• The benchmarking and documentation drafter. The production layer of transfer pricing. A detailed article on this topic is written.
• The first-line responder. Routine notices and the internal tax helpdesk.

In each, the knowledge required is knowledge of a process, and someone else checks the output. Where both are true, the role is exposed.

The people who sign, negotiate with authorities, structure transactions and tell a CFO no are staying.

The Roles Being Created

• The agent supervisor. A qualified professional who runs AI agents the way a manager runs a team, and knows where the machine is likely to be confidently wrong.
• The tax data owner. Accountable for the chart of accounts, the entity master and the mapping from ledger to return. Someone who knows the rules, and the mappings. Half lawyer, half engineer.
• The tax AI controller. Owns the evidence: which model did what, checked by whom, against which source, and where the data went and why.
• The real-time compliance lead. Responsible for what reaches (and what doesn’t) the authority at the moment of transaction.
• The algorithmic audit specialist. Works out why the authority's model flagged a case, and answers with data and evidence.

There will be fewer of these roles than the roles they replace. A compliance team of thirty becomes eight or ten people doing more valuable work. I would rather say that plainly than imply an even trade.

There is a second problem. Judgement in tax was never taught. It was a by-product of preparing thousands of computations. If the junior work goes, the apprenticeship goes with it, and firms cutting graduate intake to protect this year's margin are spending the seed corn. The US tax authority showed the same mistake at scale: it cut a quarter of its staff, and its auditor then found it lacked the people to run its AI.

What Today's Workforce Needs to Do

The job is changing from producing to standing behind. That is better work, and it was always the meaningful part: judgements about what a business owes, what it can defend and what it should disclose. Getting there takes deliberate effort. This is what I would do, by career stage.

If you are in your first five years:

• Learn to review before you are asked to. Take AI-prepared work and find the errors. That is the job you will have.
• Rebuild one thing by hand. One computation, one return, end to end, unaided. Not for the output. So you know what the machine is doing.
• Pick up a second skill. Managing big data, a rules platform or basic code. The hybrid is who gets hired.
• Get into the room. Client calls and authority meetings are where judgement forms.

If you are mid-career:

• Move from preparing to owning. Take a process, its data and its controls, and put your name on it.
• Use the tools every day on real work. You cannot supervise what you have never operated.
• Go deep where machines are weakest. Disputes, transactions, uncertain positions and anything that depends on knowing the business.
• Write down what you know. Why a position was taken, not only what it was. That is what makes you hard to replace and your team easy to train. Show your work.

If you lead a function or a firm:

• Keep hiring juniors, and redesign how they learn. Train judgement on purpose, because it will no longer happen by accident. Machines will do the grunt work, and juniors may never get the exposure that you got during your career start.
• Fix the data before buying the agent. A bad data will make the agent say wrong things more confidently.
• Measure the return. Four in five firms do not. Don’t implement AI just because it is trending and you have a mandate to apply it.
• Keep the numbers deterministic. Let AI read, propose and draft. Let rules compute. Let a person sign.

By 2030 a machine will prepare the return. It will not sign it, defend it in front of an authority, or tell a CFO what the business should not do. Those were always the parts of the job that mattered, and they are about to be the whole of it. The people who do well will be the ones who start doing that work now, before the rest is taken off their desk.