Short answer: no, not as a profession, but the volume-return part of the business is being compressed hard and the advisory part is growing. Simple returns were already largely automated before generative AI arrived. What is changing now is the middle tier: document intake, classification, reconciliation and first-draft preparation. What is not changing is who signs the return, who is liable for it, and who talks to a client about a decision that will affect them for years. Official projections reflect that split. Tax examiners, collectors and revenue agents are projected to decline 1 percent between 2025 and 2035, while accountants and auditors are projected to grow 5 percent over the same period and add 79,400 jobs.

Reading the Numbers Properly

Two occupations bracket this question, and the gap between them tells the story.

Tax examiners, collectors and revenue agents number 60,100 with a 2025 median pay of $62,370, and the projected change through 2035 is a loss of about 600 positions. That is a slow decline, not a collapse, in a role heavily weighted toward rules-based review inside government agencies.

Accountants and auditors number 1,595,200 with a 2025 median pay of $83,680 and 79,400 projected additional jobs. For context, BLS projects total US employment to grow 3.5 percent from 2025 to 2035, so accounting is projected to grow faster than the economy as a whole.

The pattern that emerges across both: work defined by applying known rules to structured inputs is under pressure, and work defined by judgement, liability and client relationships is not.

BLS also released AI exposure categories alongside the 2025-35 projections, grouping every detailed occupation into Low, Moderate, High or Very high relative exposure using five external datasets, including observed usage measures built from Claude and Copilot data. The agency states directly that exposure “does not imply job loss, productivity gains, automation probability, or wage effects.” Financial and clerical occupations sit high on theoretical exposure measures, which is a statement about task overlap, not about employment outcomes.

A Task-Level Audit of Tax Preparation

Arguing about the profession in the abstract is unproductive. Here is the work broken into its actual components.

TaskAutomation statusWhat still needs a person
Document intake and OCRLargely automatedChasing the documents the client never sends
Categorising expensesMostly automatedDeciding what is genuinely deductible in context
Data entry into formsAutomatedConfirming the source data was right
Basic return calculationFully automated for decadesNothing
Multi-state or multi-entity allocationPartly automatedPositions where guidance is ambiguous
Research on an unusual positionAccelerated, not solvedJudging whether the authority applies to these facts
Choosing an entity structureNot automatedWeighing tax against liability, exit plans and family goals
Signing and representing the returnNot automatableThe licensed preparer, by law
Explaining a bad outcome to a clientNot automatableThe person the client trusts

The rows at the top are where firms make thin margins on volume. The rows at the bottom are where firms make real money. AI is compressing the top and leaving the bottom intact, which is why the effect on individual careers depends almost entirely on which rows you spend your time in.

The liability line is the real boundary

Paid preparers operate inside an accountability structure. The IRS maintains requirements and resources for the profession through its tax professionals area, covering credentials, continuing education and representation rights. A model can draft a return. It cannot hold a preparer tax identification number, it cannot represent a taxpayer in an examination, and it cannot be sanctioned.

That is not a technicality. It means every AI-produced output in a professional context needs a licensed human to adopt it, and adoption is a review task that carries risk. Firms that treated review as a formality during the first wave of automation discovered the cost of that quickly.

A worked example of where the boundary sits

A client sells a rental property held for nine years, having refinanced twice and used part of the home as an office for three of those years. A model handles the mechanical parts well: it can lay out the depreciation recapture calculation, identify which adjustments to basis are in play, and draft the disclosure language. What it handles badly is the fact-finding that determines whether those mechanics apply. Was the office use exclusive and regular? Did the second refinance produce proceeds used for something that changes the interest treatment? Did the client ever rent to a relative below market?

None of those answers exist in the documents. They exist in a conversation, and asking the right follow-up question depends on recognising which detail in the file looks slightly wrong. That recognition is the professional skill, and it is exactly the part the software does not supply. The preparer who lets the model do the arithmetic and spends the recovered time on the conversation produces a better return than either the software alone or the preparer who does everything manually.

What to Know Before You Draw Conclusions

Consumer software already did the easy part. Simple returns moved to self-service software years ago. Preparers who still serve that segment have been under pressure for a decade, and generative AI is an acceleration rather than a new threat.

Tax law changes faster than training data. Provisions change annually, and models trained on older material confidently apply superseded rules. Every position taken from a model output has to be verified against current authority.

Errors are asymmetric. A model that is right 95 percent of the time is impressive in most domains and unacceptable here, because the 5 percent generates penalties, interest and professional liability. This asymmetry is the main reason adoption in tax practice runs behind adoption in marketing or software.

Firm economics change before headcount does. The first visible effect is not layoffs. It is fee compression on compliance work, which pushes firms toward advisory revenue. Careers change through that shift rather than through sudden redundancy.

Client trust is not transferable to software. Most people do not choose a preparer on price. They choose one because someone told them this person is reliable. That relationship is the asset the technology does not touch.

Where the Work Is Moving

The growth is in four areas, all of which existed before and are now more valuable because compliance is cheaper to produce.

  • Planning rather than filing. Entity choice, timing of income and deductions, retirement structuring, equity compensation. Work done before the year closes, not after.
  • Complex and cross-border situations. Multi-state residency, foreign income, trusts, estates. Ambiguity is precisely where models are least reliable.
  • Representation and controversy. Examinations, appeals and collections require credentialed representation and negotiation.
  • Advisory for small business owners. Cash flow, structure and the tax consequences of decisions the owner has not made yet.

Firms that repositioned toward these areas describe the same sequence: automate intake and preparation, redeploy the time saved into planning conversations, and reprice from per-return to per-relationship. The technology is the enabler, but the revenue comes from the conversation.

A Decision Framework for Your Own Position

Work out which of these describes you, because the right response differs.

  1. High-volume individual returns, seasonal practice. Most exposed. Your defensible move is to convert a subset of clients into year-round planning relationships this year, not next.
  2. Small firm generalist. Moderately exposed on compliance, well positioned on advisory. Automate intake first, then use the reclaimed hours for planning conversations rather than taking on more returns.
  3. Specialist in complex areas. Least exposed. Your risk is complacency about tooling, not about demand. Learn to use research assistants well so your speed matches your judgement.
  4. Early career, deciding whether to enter. The entry-level tasks you would have learned on are shrinking, which makes deliberate exposure to advisory work important early. Choose an employer by what you will be allowed to do in year two.

A useful test in all four cases: what fraction of your billable hours goes to work where the answer is determinable from the documents alone? If it is above half, that fraction is your exposure, and the number to move over the next two years.

Common mistakes practitioners make right now

  • Banning AI tools outright, which pushes usage underground with no review process.
  • Adopting them without a verification step, which converts a speed gain into a liability.
  • Feeding client data into consumer tools without checking data-handling terms.
  • Assuming the technology reduces the need for continuing education, when it raises it.

Skills Worth Building This Year

The preparers coming out of this well share a specific capability: they can direct these tools precisely, verify the output efficiently, and explain to a client where the machine helped and where a person decided. That is a learnable skill set, and it is more about workflow design than about technology.

Concretely, that means being able to build a repeatable intake-to-draft pipeline, write instructions that produce checkable output rather than confident prose, design a review step that catches the failure modes these models actually have, and document the process well enough to defend it. Structured learning gets you there faster than experimenting between deadlines, because it gives you a sequence instead of a pile of tips, and a certificate alongside applied practice makes the capability visible to clients and employers. If you want a structured starting point, explore Coursiv AI lessons and check current plan details on the official site.

FAQ

Will AI take my job as a tax preparer?
Unlikely as a whole job. Likely for a portion of your tasks. Official projections show decline in rules-based tax roles and growth in accounting overall, which points to redistribution rather than elimination.
Can AI legally prepare and sign a tax return?
No. Paid preparation and representation require credentialed humans under IRS rules. Software can draft; a person adopts, signs and carries the liability.
What are the real limitations in tax work?
Currency of law, handling of ambiguous authority, multi-entity fact patterns, and confident errors that read as correct. Each of these requires verification against primary sources.
Will fees fall?
Compliance fees are already under pressure and that is likely to continue, because the cost of producing a straightforward return keeps falling. Planning and representation fees are not under the same pressure, which is why the revenue mix matters more than the fee level.
What should I learn first?
Document-to-draft automation for your intake process, then a disciplined verification workflow. Those two changes free the most time with the least risk.

Your Next Move This Quarter

Pick your ten most repetitive returns and time how long intake, categorisation and first-draft preparation actually take. Automate the intake step only, keep every review gate in place, and reinvest the hours you recover into one planning conversation per client. That single change tests the technology on low-risk work, produces a number you can act on, and moves revenue toward the part of the profession that is growing rather than the part that is being compressed.