The Case for Verified Autonomy

01The coming capacity gap

Demand for accounting work has never been higher. The global accounting services market was worth roughly 700 billion dollars in 2024 and is projected to grow at a high single digit annual rate, approaching 1.5 trillion dollars by 2034.⁴ A more complex tax code, more business formation, and expanding compliance and reporting obligations mean every taxpayer and business owner needs more of this work done, not less.

The people who do it are disappearing. The AICPA reported that roughly 75 percent of CPAs had reached retirement eligibility by 2020, and an estimated 300,000 accountants and auditors left the profession between 2019 and 2022.¹ The pipeline behind them has thinned: accounting fell from about 4 percent of college majors a decade ago to 1.4 percent, and the profession produced only about 55,000 graduates in the 2023 to 2024 year.² The U.S. Bureau of Labor Statistics still projects on the order of 124,000 accounting and auditing openings every year through 2034.³ Roughly 124,000 seats to fill each year, and roughly 55,000 people entering to fill them.

This is not a cyclical staffing dip that better recruiting will close. It is structural. Figure 1 shows the dynamic in the language of a market. Demand shifts steadily to the right as the economy and regulation grow, while the supply of licensed professionals shifts left as it retires faster than it is replaced. At any sustainable price for the work, the quantity demanded now exceeds the quantity human supply can provide. The distance between those two points is a shortage that cannot be recruited away.

For taxpayers and business owners, the gap is already visible as higher fees, longer turnaround, and less advisory attention, with small businesses feeling it first. The level of service a dedicated, experienced professional once provided is becoming something the market cannot supply to everyone who needs it.

Our view of the trajectory is straightforward. As models and agents are trained on real expert feedback, intelligence becomes the only way the demand gets met. Over time the market will rely more and more on aligned AI to deliver a service level that human capacity alone cannot reach, not by replacing professionals but by multiplying each one’s reach. A firm running an aligned model serves a scale of clients that a firm without one simply cannot. ARV expands the effective supply of expert grade accounting, the dashed curve in Figure 1, so that demand can be met at an accessible cost while licensed professionals stay accountable for the work. The rest of this paper is how we make that expansion safe.

02A regulated discipline meets a capable machine

Modern models can read a shoebox of documents, reconcile a ledger, and draft a return in minutes. That capability is real, and it is why we built ARV on the same engine that top firms already trust. But capability is not the same as accountability. In accounting and tax, the output is not a suggestion; it is filed, relied upon, and defended. The discipline is bounded by statute, by professional standards, and by the simple fact that the numbers have to be right.

The prevailing regulatory consensus is that high stakes AI must remain under meaningful human control. The European Union’s AI Act requires that high risk systems be designed so they can be “effectively overseen by natural persons,” and that those persons retain the authority to interpret output, detect anomalies, and decide not to use the system or to disregard its output.⁵The United States’ NIST AI Risk Management Framework treats human oversight as the primary defense against automation bias, with its Govern, Map, Measure, and Manage functions driven by human teams.⁶We treat these not as compliance overhead but as a description of how serious work should be done.

03Why oversight is not optional in our field

The accounting profession has been unusually clear on this point. The AICPA’s guidance holds that AI supports but does not replace professional judgment, and that the human in the loop is a governance requirement wherever skepticism is required at a decision point, not merely a best practice.^7,11CPAs are positioned to provide exactly this oversight because the profession’s standards already demand independent thought and questioning.

There is a well documented failure mode this guards against. Decades of human factors research show that people over rely on automated aids, skipping the independent verification they would otherwise perform, and that this “automation bias” appears in novices and experts alike.⁸An AI that is right most of the time is precisely the kind of system that lulls its operator into waving through the case where it is wrong. Our answer is not to ask humans to try harder. It is to engineer verification into the workflow so the human is placed exactly where judgment is needed.

“AI supports, but does not replace, professional judgment. Where skepticism is required at each significant decision point, the human in the loop becomes a governance requirement, not just a best practice.”Journal of Accountancy, AICPA⁷

04Our method: alignment to expert behavior

The technique that made modern AI useful is reinforcement learning from human feedback. In the work that defined the approach, a model proposes, humans express a preference between outputs, and those preferences train a reward model that steers the system toward what experts actually want. Notably, a 1.3 billion parameter model aligned this way was preferred over a 175 billion parameter model that was not, and it produced far fewer factual errors.⁹ Alignment, not raw scale, is what makes output trustworthy.

We apply that method to accounting. Instead of harvesting the open web, ARV’s engine is trained on feedback from top CPA firms and accountants, together with our own Reinforcement Learning Center of licensed professionals. Real engagements become structured preference data. Every correction an expert makes is signal that sharpens the reward model. The result is a system shaped by some of the best accountants in the country, and reviewed by real ones.

05Structured data and standardization

No model is better than the data beneath it. The oldest rule in computing still governs machine learning: garbage in, garbage out. Structured, standardized data is what lets an algorithm interpret a document the same way every time, and it is increasingly understood as essential infrastructure for trustworthy AI.¹⁰ Inconsistent inputs do not merely lower accuracy; they hide error where it cannot be measured.

ARV runs the standardized workflow that top firms use, executed against a normalized data estate. Source documents are classified and extracted into a consistent schema before any judgment is applied, so that every figure traces back to a document and every output is measured on comparable ground. That discipline is what makes the next step possible: if you cannot standardize the input, you cannot honestly score the confidence of the output.

06The verification gate

This is the core of our position. ARV does not let a model act unsupervised on consequential work until its measured confidence clears a threshold calibrated against expert ground truth. Every output carries a confidence score. Anything below the gate is routed to a licensed professional for review before it moves. Anything above the gate is eligible for autonomous processing, and is still continuously sampled so the threshold itself stays honest.

The distribution below is the practical picture. The long tail of low confidence extractions, the genuinely ambiguous cases, is exactly where human expertise is spent. The high confidence mass is where autonomy is earned, not assumed. Required outputs are verified regardless. As alignment improves and the distribution shifts right, validated volume grows without ever removing the human from the decisions that matter.

07What the method produces

The loop compounds. As expert corrections train the reward model, accuracy on source document extraction and tax preparation has moved from an unaligned baseline of 34 percent to 97 percent, and continues to climb with each round of feedback. That is the difference between a capable model and an aligned one, drawn from our own training record.

Beyond accuracy, the advisory layer that this alignment enables has identified 13.2 million dollars in potential tax savings across 6,200 client files, work that went unnoticed under traditional service. These are not projections. They are the output of a system that screens every incentive against a client’s standardized data and hands the ambiguous cases to a professional.

08Position

We take AI in regulated work seriously enough to constrain it. ARV will not let a model run without a human verifying its output until confidence, measured against expert ground truth, reaches a level that supports autonomy. That is not a limitation on the technology; it is the condition under which the technology can be trusted. Licensed, experienced professionals are responsible for guiding the model, and they remain responsible for the decisions that carry consequence. Autonomy, in our field, is something a model earns.