Accounting and Tax: The Credence Gap AI Hasn't Closed

The Problem Buyers Can't Solve Alone

Start with the economics, because everything else in this vertical follows from it. When a client hires an accountant or tax preparer, they are buying something economists call a credence good: a service whose quality they cannot fully judge before they buy it, and often still cannot judge after. Dulleck, Kerschbamer, and Sutter formalized this in a 936-participant experiment published in the American Economic Review, and the core finding holds up: in markets where buyers cannot verify quality themselves, reputation, liability rules, and third-party verifiability step in to do the job the buyer's own judgment can't.

A follow-up review by Balafoutas and Kerschbamer, covering roughly fifteen years of research since, confirms this generalizes well beyond accounting, to medical care, legal advice, and auto repair. The pattern is the same everywhere: when the buyer can't inspect the work, the market invents proxies. Reputation. Verifiability. Liability. A clean separation between who diagnoses the problem and who gets paid to fix it. Apply that directly to accounting and tax: a client receives a filed return or a set of financial statements and has essentially no independent way to know whether it was optimal, merely adequate, or quietly wrong in a way that will surface only at an audit two years later.

This is the frame the rest of this study sits inside. Every other data point, from review trust to AI adoption to firm marketing spend, is really a story about how buyers and sellers are adapting to that one structural fact.

A tax return that saves you money and one that quietly puts you at audit risk can look identical on the page you're asked to sign.

What Buyers Reach For Instead

If a buyer can't judge the work directly, what do they judge instead? The evidence points to a shrinking set of proxies, and one of the largest, online reviews, is losing credibility fast. BrightLocal's 2025 Local Consumer Review Survey found that trust in online reviews as a stand-in for a personal recommendation fell from 79 percent in 2020 to 42 percent in 2025. That is a five-year collapse in one of the market's primary substitute signals.

Buyers have not stopped looking, they have started looking harder. The same survey found 74 percent now check two or more review sites before deciding, consistent with credence-goods theory: when one signal weakens, buyers triangulate across several rather than trusting any single source. Interestingly, the data splits on AI's role in this process. Eighteen percent say they would make a decision based on a review summary they knew an AI system had produced alone, while 46 percent separately flag AI authorship itself as a red flag for a fake review. Buyers are not rejecting AI-assisted information outright, they are suspicious of it specifically when it might be manufacturing the trust signal rather than reporting it.

Credentials remain the one proxy that is externally checkable rather than reputational. The IRS's PTIN registry, with 872,363 currently active holders as of mid-2026, is the market's closest thing to a hard verification layer: a preparer either has a current PTIN or they don't, and a client can check.

The Filing Season Squeeze

The January-to-April concentration of demand in this vertical is one of the most institutionally established facts in the evidence: the IRS itself organizes its entire operating calendar, its own terminology, around "filing season," and firm-level surveys frame growth and staffing decisions around the same window. That much is not in dispute.

What is genuinely missing is a quantified curve. Despite a real search for one, no month-by-month or week-by-week demand index, the kind of dataset that would show exactly how search interest for "accountant near me" or "tax preparer" rises and falls across the year, turned up in this research pass. That gap matters more than it might first appear: a firm that assumes demand simply spikes in March and plans its visibility spend accordingly is operating on institutional folklore, not measurement. Understanding precisely when attention shifts, and how far in advance it shifts, is exactly the kind of question a demand-side attention map is built to answer, and exactly the kind this study cannot yet answer on its own.

AI-DIY Is Growing, But From Underneath a Low Ceiling

The headline number here is real: a single-vendor Adobe survey found reported AI use for personal tax filing more than doubled year over year, from 11 percent in 2024 to 26 percent in 2025 returns, with the growth concentrated among 1099 filers and Gen Z respondents. A separate survey, sponsored by IPX1031 and reported by KXAN Austin, found 46 percent of respondents say they trust AI to give accurate tax guidance, and about one in five planned to use AI to help file, most often to answer filing questions, spot deductions, or review a return before submitting it.

Both numbers deserve a caveat the evidence itself flags: these are single-vendor, marketing-adjacent surveys with self-reported panels, not corroborated by a neutral academic or government source. The IPX1031 figure in particular is tiered as contested here, not because it's implausible, but because nothing independent backs it up yet. Read together, though, the two surveys tell a consistent story: adoption is accelerating, and it is still, on every measurement available, a minority behavior. The top concerns cited in the Adobe data, misinterpreting tax law and data privacy, are exactly the concerns you would expect in a credence-goods market where the client cannot verify the AI's output any better than they could verify a human preparer's.

Adoption is real. It is also, on every survey that measured it, still a minority.

Even the Government Couldn't Make Free, Simple DIY Work

The most instructive data point in this study is not about AI at all, it is about what happened when the US government tried to remove cost and friction from DIY filing entirely. The IRS's Direct File pilot, free and government-run, registered only about 751,000 people out of roughly 32 million eligible taxpayers in the 2025 filing season, a 2.3 percent uptake rate, and 59 percent of those who did register never actually filed through it. This is not a new problem: Free File, which has existed far longer, saw only about 2.8 percent uptake among eligible tax units for tax year 2024, a pattern TIGTA has documented since 2020 and attributes largely to low awareness and a sign-up path that is confusing and hard to find.

That matters for how you read the AI-DIY threat. If free, government-backed, zero-cost filing can't clear a low single-digit adoption rate even with the friction of price removed entirely, the ceiling on DIY displacement generally, AI-powered or otherwise, is probably lower than raw AI-adoption headlines suggest. The barrier isn't price. It's trust, awareness, and the sheer discomfort of being solely responsible for getting it right.

There is one important exception worth separating out. Inside Direct File itself, a narrow-scope AI chatbot handling a specific, bounded task, eligibility questions, resolved 68 percent of roughly 16,000 interactions without escalating to a human. That is a meaningfully high resolution rate. The lesson is not that AI can't handle tax questions competently. It's that competence and adoption are two different problems, and this vertical's real bottleneck has so far been the second one.

Adoption is real. It is also, on every survey that measured it, still a minority.

The Enforcement Backdrop: PTIN and Ghost Preparers

Credence-goods theory predicts that markets like this one need external verifiability, some check the buyer can rely on that doesn't require them to judge the work itself. The PTIN registry is that check. The IRS tracks 872,363 currently active PTIN holders and 2,305,910 cumulative registrations since 2010, and it treats anyone preparing returns without a current PTIN as operating outside the legitimate system entirely.

The IRS has gone further and named the failure mode explicitly: "ghost preparers," defined as preparers who complete a return but refuse to sign it or provide a PTIN, often after promising inflated refunds or fabricated deductions and credits, leaving the taxpayer legally on the hook for whatever gets filed. There is a dedicated reporting channel and a formal warning-letter process (Letter 4733/6623) built specifically around this pattern. For a legitimate firm, this is a competitive fact worth internalizing: your PTIN status is one of the only pieces of information in this entire buyer's path that is objectively, externally verifiable, and it should be as easy to find as your phone number.

Data Stewardship Becomes a New Trust Test

The credence-goods problem in this vertical is expanding beyond "was my return done correctly" into "what happened to my information once I handed it over." In September 2023, the FTC formally warned five tax-preparation companies that using or disclosing data collected for tax preparation for other purposes, such as advertising, could expose them to civil penalties. That is a regulator explicitly treating data handling as a distinct compliance and trust risk, separate from filing accuracy.

For a client who already cannot verify the technical quality of the work, this adds a second layer of unverifiable trust: did the firm protect what I gave them. Firms that can speak plainly and specifically about how client data is handled are addressing a credence-goods gap that most competitors aren't even acknowledging exists.

How Firms Are Actually Competing

On the supply side, the evidence shows firms responding to exactly the dynamics the theory predicts. A CPA.com and BILL survey of 400 respondents, including 200 firm leaders, found technology now ties with firm reputation and brand management as the top-cited driver of new-client acquisition, ahead of both referrals and price. Fifty-four percent of firm leaders pointed to responsiveness and 52 percent to accuracy and reliability as reasons clients choose them. Separately, Thomson Reuters' annual global survey of tax, audit, and accounting firm leaders found that "grow firm and expand client base" jumped from fifth to second in firm priority rankings year over year, trailing only efficiency and automation gains.

Both figures come from single-vendor, sponsored industry surveys rather than independently audited research, which is why they are tiered as emerging here: read the specific numbers as a directional read on where firms say their attention is going, not a precise census of the whole market.

Read together, this is firms doing precisely what credence-goods theory says a seller should do when price and pure word-of-mouth referral aren't enough on their own: compete on the things a buyer actually can observe. How fast you respond. How modern your systems look. What your reputation signals about your reliability before the client has any way to judge the work itself. None of this replaces genuine competence, but it is the visible proxy for competence that the market has settled on, and it's intensifying, not fading.

When a buyer cannot judge the work, firms compete on what the buyer can judge: how fast you respond, how modern you look, and what other people say about you.

The Layer Nobody Has Measured Yet

Everything above describes a buyer's path built on referrals, credentials, reviews, and firm reputation, tested now by AI-DIY tools on one side and a struggling government DIY experiment on the other. What it does not yet describe is where AI-answer engines themselves fit into discovery: when someone asks ChatGPT, Google's AI Overviews, or Perplexity who to trust with a tax return, what gets recommended, cited, or surfaced, and how often.

That measurement simply does not exist yet for this vertical. It was searched for directly in preparing this study and not found, and the right way to handle it is to treat it as unmeasured rather than to assume it's small because the vertical is trust-heavy, or large because AI tools are growing. It is the frontier this study can name but not yet size, and it is exactly where a demand-side attention map earns its keep: not by guessing at the answer, but by actually watching where a market's attention is going, filing season after filing season, and reporting back on what it finds.