The AI vs offshoring argument in accounting is usually run by people selling one of them. A useful starting point is a number that both sides quote: in a peer reviewed 2024 study of accounting certification exams, ChatGPT eventually scored an average of 85.1 percent and passed every section of the multiple-choice practice questions tested (Review of Accounting Studies, 2024).
The same authors then wrote that they never tested whether it could do accounting work.
That gap, between answering questions and doing the job, is where a firm's decision sits. Below is what each option removes from your week, what the law says about both, and how to sort your own task list instead of picking a side.
The Short Answer: They Fail in Different Places
AI removes keystrokes from work that is already defined. An offshore accountant removes hours of qualified work from people who are already at capacity, including the undefined parts that nobody has written down yet.
Neither removes the signature, and neither removes the liability. A tool cannot hold a preparer tax identification number, and an offshore preparer working under your firm does not take the partner's name off the return. So the question is never which one replaces your review. It is which one shortens the queue in front of it.
What the Research Measured, and What It Did Not
Eulerich, Sanatizadeh, Vakilzadeh and Wood tested ChatGPT against practice questions for the CPA, CMA, CIA and EA exams, and published the results in Review of Accounting Studies in June 2024. The early ChatGPT 3.5 model averaged 53.1 percent and passed nothing, and the model only cleared every section after three separate upgrades (Review of Accounting Studies, 2024).
Those upgrades matter more than the headline. Moving to ChatGPT 4 added an average of 16.5 percent, prompting it with 10 examples added another 6.6 percent, and giving it tools to reason with, including a calculator, added a further 8.9 percent to reach 85.1 percent (Review of Accounting Studies, 2024). The jump came from training and tooling, not from the model waking up one morning smarter.
Then the authors listed their own limits, and they are the part vendors skip. They omitted questions "that require greater cognitive ability, such as interpreting situations and contexts," they used practice exams rather than real ones, and they wrote plainly that "we do not test whether it can perform actual accounting tasks, such as bank reconciliations, tax preparation, closing the books." Their closing line asks for research that can show "whether AI can move from knowing to doing."
Read that as a buyer, not as a headline. A high exam score is evidence that a model has absorbed the rules. It is not evidence that it can pull a trial balance apart, notice the accrual that was reversed twice, and tell you which client to call.
Where AI Genuinely Wins
AI earns its place on work that is high volume, rule bound, and checkable in seconds. Coding recurring bank transactions, extracting fields from documents, matching one ledger against another, drafting a first version of a memo, and scoring a population to say which items look odd are all tasks where a wrong answer is cheap to spot and cheap to fix.
The clearest public example is the tax authority itself. In September 2023 the IRS said it was expanding its Large Partnership Compliance program with the help of AI, describing machine learning applied to identify potential compliance risk, and said it would open examinations of 75 of the largest partnerships in the country (IRS, IR-2023-166). Note what the machine was doing there. It was ranking a population so that people knew where to look, not writing the audit conclusion.
The Bureau of Labor Statistics splits the work the same way from the other side. It says some routine accounting tasks may be automated as cloud computing, AI and blockchain spread, and that this will make accountants' "advisory and analytical duties more prominent" (BLS Occupational Outlook Handbook).
One honest caveat. Everything above is a snapshot, and model capabilities move faster than the reviews of them. Whatever a tool could not do last season is worth retesting on your own files before you decide it still cannot, and whatever a vendor claims it can do is worth testing on a return you already know the answer to.
Where an Offshore Team Wins
Offshore staff earn their place on work that is judgment heavy, badly documented, or dependent on chasing someone. A missing K-1, a client who sends a shoebox of PDFs in the wrong order, a fixed asset schedule that has drifted from the depreciation software, a set of workpapers that has to be rebuilt so a reviewer can follow it: these are hours of qualified attention, not keystrokes.
They also win on the thing a firm sells, which is accountable capacity. Circular 230 is the Treasury rule set governing practice before the IRS (IRS, Office of Professional Responsibility). Section 10.35 states that a practitioner "must possess the necessary competence to engage in practice before the Internal Revenue Service" (eCFR, section 10.35), and section 10.22(b) gives a due diligence presumption where a practitioner "relies on the work product of another person" and "used reasonable care in engaging, supervising, training, and evaluating the person" (eCFR, section 10.22).
The wording there is worth sitting with. The reliance presumption is written around a person you engage, supervise, train and evaluate. Software is not a person you train, and on its face the rule offers no equivalent shelter for output you accepted because a tool produced it. Competence stays with the practitioner either way.
The Preparer of Record Is Always a Human Being
A filed return prepared by a tax return preparer must include the identifying number of the preparer required to sign it, and for returns filed after December 31, 2010 that number is the individual's preparer tax identification number (eCFR, section 1.6109-2).
That number belongs to a person. The IRS requires it of anyone who prepares or assists in preparing federal tax returns for compensation, credentialed or not, and its own page speaks directly to the "non-credentialed return preparer" (IRS, PTIN requirements for tax return preparers).
You will still read that the number is open only to attorneys, CPAs, enrolled agents and registered tax return preparers. That requirement does sit in section 1.6109-2(d), but the paragraph opens "Except as provided in paragraph (h) of this section," and paragraph (h) lets the IRS prescribe exceptions "including the requirement that an individual be authorized to practice before the Internal Revenue Service before receiving a preparer tax identification number" (eCFR, section 1.6109-2). Quote the gate without the exception and you have described a rule that does not operate the way it reads.
None of that weakens the point, because the point is who holds the number. It is issued to an individual, not to a product. Whichever route you pick, a person signs, and the firm carries what that signature means. Both AI and an offshore preparer sit upstream of that line, which is why comparing them as if one of them ends your review is a category error.
Your AI Vendor May Be a Tax Return Preparer Under Section 7216
This is the part most comparisons miss entirely. Section 301.7216-1(b)(2)(i)(B) defines a tax return preparer to include "any person who is engaged in the business of providing auxiliary services in connection with the preparation of tax returns, including a person who develops software that is used to prepare or file a tax return and any Authorized IRS e-file Provider" (eCFR, section 301.7216-1).
Section 301.7216-1(b)(2)(iii) sets the limit worth reading before you assume the definition does or does not reach your tool. A person is engaged in that business if, in the course of the person's business, the person "holds himself out to tax return preparers or to taxpayers as a person who performs auxiliary services" (eCFR, section 301.7216-1). Software your team points at a return is squarely inside the question. A general assistant nobody sells as tax software may not be. That is a question to answer in the contract, not to assume in either direction.
So the AI product processing your clients' data is not automatically outside the regime that firms think of as an offshoring problem. Where it is inside, the same section's definition of tax return information reaches what the tool generates, since the term "also includes information the tax return preparer derives or generates from tax return information in connection with the preparation of a taxpayer's return."
The routing rule turns on location. Section 301.7216-2(d)(1) permits preparer to preparer disclosure to another tax return preparer "located in the United States (including any territory or possession of the United States)" for preparation or auxiliary services, "so long as the services provided are not substantive determinations or advice affecting the tax liability reported by taxpayers," and the same paragraph defines a substantive determination as one that "involves an analysis, interpretation, or application of the law" (eCFR, section 301.7216-2). Where the recipient sits outside the United States, consent under section 301.7216-3 comes first (eCFR, section 301.7216-3).
Read that paragraph closely, because it governs a different firm. It applies to "another tax return preparer (other than an officer, employee, or member of the same tax return preparer)" (eCFR, section 301.7216-2).
Where the person sits inside your own firm, section 301.7216-2(c)(2) governs instead, and it says that if the officer, employee or member receiving the information is located outside the United States, "the taxpayer's consent under § 301.7216-3 prior to any disclosure is required" (eCFR, section 301.7216-2). Same answer on location, different paragraph, and the distinction decides which one your engagement letter should cite.
Two practical questions follow, and they apply to a software contract exactly as they apply to a staffing one. Where is the processing physically done, and does the service stop short of substantive determinations. A vendor who cannot answer the first question in writing has answered it.
The penalty side is the same for both routes too. A section 7216 violation is a misdemeanor carrying a maximum penalty of up to one year imprisonment or a fine of not more than $1,000, or both, together with the costs of prosecution, and section 6713 adds a civil penalty of $250 for each prohibited disclosure or use, not to exceed a total of $10,000 for a calendar year (eCFR, section 301.7216-1).
Both Options Are Service Providers Under the Safeguards Rule
The FTC Safeguards Rule requires a covered firm to oversee its service providers, and it defines a service provider as any person or entity that "receives, maintains, processes, or otherwise is permitted access to customer information through its provision of services directly to a financial institution that is subject to this part" (eCFR, section 314.2). An AI tool that touches client records is inside that definition on the same terms an offshore provider is.
Section 314.4(f) sets three duties: taking "reasonable steps to select and retain service providers that are capable of maintaining appropriate safeguards for the customer information at issue," "requiring your service providers by contract to implement and maintain such safeguards," and "periodically assessing your service providers based on the risk they present and the continued adequacy of their safeguards" (eCFR, section 314.4).
Read that against a typical AI purchase. A tool bought on a card by one manager, with terms nobody read and no periodic reassessment, fails the same test a badly chosen offshore vendor fails. Diligence is not a penalty applied to offshoring. It is a duty attached to sending client data anywhere.
What the Labor Numbers Say About "AI Will Replace Them Anyway"
The official projection does not describe a shrinking profession. BLS counted 1,579,800 accountants and auditors in 2024 and projects employment to grow 5 percent from 2024 to 2034, against 3 percent for all occupations, an increase of 72,800 jobs (BLS Occupational Outlook Handbook).
BLS also projects about 124,200 openings for accountants and auditors each year on average over that decade, and says many of those openings are expected to result from the need to replace workers who transfer to different occupations or exit the labor force (BLS Occupational Outlook Handbook). Read that carefully, because it is widely misquoted. It counts openings expected to appear, not accountants standing available to hire.
On automation specifically, BLS states that although the change will increase accountants' efficiency, it "is not expected to reduce overall demand" (BLS Occupational Outlook Handbook). For scale on what a domestic seat costs, the median annual wage was $81,680 in May 2024, or $39.27 an hour (BLS Occupational Outlook Handbook).
A forecast is not a guarantee, and one federal projection cannot settle an argument about technology. But if you are betting that the hiring market fixes itself because software will empty the seats, the government's own numbers are betting the other way.
AI vs Offshoring in Accounting: A Task by Task Sort
Stop comparing the two categories and start sorting the work. Run last busy season's task list through five questions, in this order.
- Is the task defined well enough that a competent stranger could do it from the written instructions? If not, no tool and no new hire helps until it is written down.
- Is the output checkable in seconds by someone who already knows the answer shape? If yes, it is a candidate for automation.
- Does the task require chasing a human being, interpreting an incomplete fact pattern, or deciding what the law means? If yes, it needs a trained person: section 301.7216-2(d)(1) treats an analysis, interpretation, or application of the law as a substantive determination (eCFR, section 301.7216-2).
- Does the task touch tax return information, and if so, where will it physically be processed? Answer that before signing anything, for software and for staffing alike.
- What does a mistake on this task cost? Expensive mistakes stay with people and stay under review, whatever the demo looked like.
Two things usually fall out of that sort. Some of the list is mechanical and should be automated. The rest only clears when someone qualified sits down and does it. The firms that get this wrong buy a tool for the second kind of work, then conclude that AI is overhyped, when what they did was point it at work it was never going to do.
The table below turns that sort into a routing rule you can apply line by line.
| Signal in the task | Route it to |
|---|---|
| High volume, defined rules, output checkable at a glance | Automation |
| Exceptions, chasing, incomplete records, judgment calls | A trained preparer |
| Reviewer time, final judgment, the signature | Your own firm, always |
When Neither One Is the Answer
If review is your bottleneck, adding either one makes the queue worse. More prepared returns arriving at the same reviewer is not capacity, it is a longer line. Fix the review layer first, then feed it.
If the work is undocumented, automation has nothing to act on, and an offshore hire ends up writing your procedures for you. That is real work, and it is slower than any sales page suggests. If the volume is small and truly seasonal, the setup cost of either route can outrun the benefit in year one. Saying so out loud is cheaper than finding out in March.
One more thing is worth deciding on purpose rather than by accident. If every mechanical task leaves the building, whether to a tool or to another country, the work your juniors used to learn on leaves with it. That is a training decision rather than a sourcing decision, and it is cheaper to make deliberately than to discover three years from now when nobody in the room can review.
Questions Firms Ask
Is AI Replacing Offshore Accounting Teams?
Not on the evidence available. The authors of the certification exam study wrote that their own discussions with accounting professionals suggest partners are considering whether AI can replace offshoring work, and in the same paper noted that ChatGPT deployment "is still very modest in accounting, especially at smaller accounting firms" (Review of Accounting Studies, 2024). Interest is not adoption, and adoption is not substitution.
Can AI Prepare and Sign a Tax Return?
It can help prepare one. It cannot be the preparer of record, because the identifying number a filed return must carry belongs to the individual required to sign it (eCFR, section 1.6109-2), and the IRS issues that number to people who prepare returns for compensation, not to products (IRS, PTIN requirements for tax return preparers).
Do We Need Client Consent to Run Returns Through an AI Tool?
It depends on who the vendor is and where the processing happens. A software developer whose product prepares or files returns is a tax return preparer under section 301.7216-1(b)(2)(i)(B) (eCFR, section 301.7216-1), and disclosure to a preparer outside the United States needs consent under section 301.7216-3 (eCFR, section 301.7216-3). The trigger is location, not technology, so ask where the processing physically happens and get the answer written into the agreement.
Should a Firm Pick One or Run Both?
They solve different halves of the same overload, so this is a sequencing question more than a choice. Document the process, automate the mechanical slice, and staff the judgment slice with people you can supervise, train and evaluate. If the process is not written down yet, neither half starts.
Start by Grading Real Work, Not Claims
Neither an AI demo nor a staffing pitch should be bought on description. Both should be bought on output you have graded yourself, on your own files, against the standard your reviewer applies on a Tuesday in March.
That is the whole test, and it is the one thing both sides of this argument tend to avoid. Don't trust us. Test us. Our Free 40-Hour Proof Pilot puts a fixed block of your own representative work through the full review chain, so your reviewer grades real output before a single client file is committed. See how the pilot works.