Common AI Mistakes Accounting Make — and How to Avoid Them

The Real Risk Isn't Hallucination — It's Confident Misreading

When accountants worry about AI going wrong, they usually picture hallucination — the model inventing a number out of thin air. The actual evidence points somewhere more unsettling. In a study of 5,000 annual reports filed between 2014 and 2023, hallucination rates stayed under 2% when filing context was provided, but AI misread real, disclosed numbers at error rates of 9.19% on structured XBRL data and up to 18.24% on plain text, according to research reported by Thomson Reuters.

That distinction matters more than it first appears. A fabricated figure is often visibly wrong — it fails a sanity check, contradicts the source document, and gets caught in review. A misread is different: the model pulls a real number from a real filing but gets the scale wrong (dollars versus millions), grabs the wrong line item, or transposes values. The output looks plausible because it is grounded in genuine data. These are precisely the errors most likely to slip past a tired reviewer scanning for obvious nonsense.

The study found three patterns that make misreading especially dangerous in accounting workflows:

• Errors scale with complexity. Larger companies, more complex filings, and footnote information produced more mistakes — ironically affecting the companies investors care about most, where accuracy matters most.
• Input format drives error rates. The same task produced 9.19% errors with structured XBRL data, 15.75% with HTML, and 18.24% with plain text. XBRL's biggest advantage was virtually eliminating scale errors.
• Newer models aren't automatically safer. A newer version of Google's Gemini made more mistakes in some areas, particularly with HTML filings. As co-author Ariel Markelevich put it, "Newer AI models don't mean that they're better."

This reframes how teams should think about verification. If hallucination were the main risk, spot-checking for invented numbers would suffice. Because misreading dominates, review has to be systematic: every AI-extracted figure needs to be traced back to its source in the filing, not just eyeballed for plausibility. Markelevich's advice captures the balance — "Use it, it's powerful. But just be cautious a little bit."

The finding also reinforces a broader consensus in the profession. The same principle appears in guidance from Becker's accounting education team: treat every AI output as a draft rather than a final answer, and apply professional skepticism to results that "sound confident and credible" but may be factually incorrect. Confident delivery is a property of the model, not a signal of accuracy — and it's exactly what makes a plausible wrong figure harder to catch than an obvious fabrication.

For teams building AI workflows around financial data, the practical takeaway is to test rather than trust. Feed models structured inputs where possible, benchmark them on your actual documents before upgrading, and keep validation steps non-negotiable. The same test-don't-trust logic underpins how we approach agent skills at SkillzBench — every skill is benchmarked in a sandbox against the same model with and without it, and the measured lift is published separately so you can see what actually helps before it touches your workflow.

Mistake #1: Over-Trusting AI and Skipping Validation

The most dangerous AI mistake in accounting isn't using AI at all — it's believing it. When a tool sounds confident, polished, and authoritative, professionals instinctively treat its output as finished work. That instinct is exactly where errors slip through.

The scale of the concern is striking. An ICAS study of accounting professionals found that 72% fear generative AI could produce errors or incorrect decisions — making error risk the single biggest worry in the profession. These fears aren't hypothetical. Becker's guidance is blunt: AI outputs "must always be treated as drafts rather than final answers."

Here's the nuance most accountants miss. Research on 5,000 annual reports found that AI's dominant failure mode isn't hallucination — it's misreading real, disclosed numbers. Hallucination rates stayed under 2%, but plausible-looking wrong reads of actual figures ranged from 9.19% to 18.24% depending on input format. A wrong number pulled from a real filing looks exactly like a correct one.

And AI doesn't fix everything equally. A peer-reviewed study on AI and accounting errors found that while AI helps with several error types, from the CPA user's perspective it still "falls short in eliminating mathematical errors." Skipping your own arithmetic checks because "the AI did it" is precisely how mistakes survive.

The longer-term risk is subtler: over-reliance erodes professional judgment. ICAS warns of hidden cognitive biases that skew judgment, while Becker notes that heavy AI reliance can reduce critical thinking and cause skipped validation steps. Your skepticism muscle atrophies every time you accept output without review.

To keep validation practical, build these habits into every AI-assisted workflow:

• Treat every AI output as a draft — apply the same professional skepticism you'd give a first-year associate's work.
• Verify figures against source documents, especially scale (thousands vs. millions), where misreads concentrate.
• Feed AI structured, clean data — error rates dropped nearly by half when inputs moved from plain text to structured XBRL.
• Keep humans accountable for every decision; AI supports judgment, it never replaces it.

The same test-don't-trust principle applies to the AI tools and agent skills themselves — SkillzBench is being built to benchmark agent skills in a sandbox so you can see measured lift before relying on them. Confidence, whether from a model or a tool, is never a substitute for verification.

Skills, proven. Explore SkillzBench's free directory and MCP connector to discover, verify, and benchmark the agent skills that actually deliver.

Mistake #2: Garbage In, Wrong Numbers Out — and Mistake #3: Leaking Client Data

When an AI produces a wrong number, the first suspect shouldn't be the model — it should be the data you fed it. A study of 5,000 annual reports found that error rates nearly doubled depending purely on input format: 9.19% for structured XBRL data versus 18.24% for plain text, with HTML filings landing in between at 15.75%.

The single biggest win from structured data was virtually eliminating scale errors — the classic mistake of reading a figure in thousands as if it were in millions. Those are exactly the errors that look plausible, pass a quick glance, and end up in a board deck. Notably, hallucination wasn't the main problem here; it stayed under 2% when filing context was provided. The real risk was AI misreading real, disclosed numbers — wrong figure, wrong scale, or confused line items.

The lesson generalizes beyond filings. As study co-author Ariel Markelevich put it, "The better the input, the better the output." ICAS's year-long research reaches the same conclusion: weak inputs lead to weak outputs. Before blaming the model, check what you're giving it.

• Feed structured data wherever possible — XBRL, tables, clean exports — rather than pasted prose.
• Watch scale and units explicitly; ask the AI to state the reporting currency and magnitude it read.
• Expect errors on the hardest targets: the study found larger, more complex companies and footnote data produced more mistakes — ironically, the companies investors care about most.
• Treat every output as a draft requiring validation, not a final answer.

The second fixable mistake compounds the first: feeding sensitive client data into unapproved tools. 52% of accounting professionals worry about client data privacy and confidentiality with generative AI, and the concern is well-founded — improper use can trigger confidentiality breaches, data breaches, and unauthorized exposure with legal, regulatory, and reputational consequences. The rule from Becker's guidance is blunt: use only organization-approved AI tools, and never enter sensitive data into unapproved platforms.

Both mistakes share a root cause: trusting tools on reputation rather than evidence. That instinct fails in a surprising way — newer models aren't automatically better. In the filings study, a newer version of Google's Gemini actually made more mistakes in some areas, especially with HTML filings. "Newer AI models don't mean that they're better," Markelevich warns. Upgrade decisions should be driven by testing on your actual work, not release notes.

This test-don't-trust principle extends to the skills and extensions you attach to AI agents. Roughly a quarter of published agent skills carry at least one vulnerability, which is why SkillzBench is building independent benchmarking — testing each skill in a sandbox with the same model, with and without it, and publishing the measured Lift. Its free directory, security scans, and MCP connector are designed so you can evaluate what you plug in before it touches real workflows. Browse and test free agent skills on SkillzBench to extend your AI work safely.

Your Prevention Playbook: Drafts, Clean Inputs, Approved Tools, Trained Teams

Knowing where AI fails is only half the battle — the other half is building habits that stop those failures before they reach a client deliverable. The research points to a consistent five-step prevention playbook.

Step one: treat every AI output as a draft. Becker's guidance is explicit — AI outputs "must always be treated as drafts rather than final answers." This matters more than most professionals realize, because the most dangerous errors are the believable ones. A study of 5,000 annual reports found that AI's main failure in financial data is not hallucination but misreading real, disclosed numbers — a wrong figure, wrong scale, or confused line items that look perfectly plausible.

Step two: feed AI structured, clean data. Input quality directly determined error rates in that same study: 9.19% errors with structured XBRL data, 15.75% with HTML, and 18.24% with plain text. As co-author Ariel Markelevich puts it, "the better the input, the better the output." Structured formats virtually eliminated scale errors — AI misreading whether figures are in dollars, thousands, or millions.

Step three: restrict sensitive data to approved tools. With 52% of accounting professionals worried about client data privacy, Becker's rule is simple: never enter sensitive client data into unapproved platforms. The consequences — confidentiality breaches, regulatory exposure, reputational damage — far outweigh any convenience.

Step four: test models empirically on your actual work. Don't assume newer means better. The filings study found a newer version of Google's Gemini made more mistakes in some areas, and Markelevich warns that "newer AI models don't mean that they're better." This test-don't-trust principle applies to agent skills too — SkillzBench is building exactly this kind of verification, benchmarking skills in a sandbox with the same model, with and without the skill, and publishing the Lift so you can see what actually helps.

Step five: invest in training and governance. Becker's blunt warning: "the biggest risk is using AI without proper training." ICAS recommends clear frameworks for when and how AI should be used, strong oversight, and clear accountability for every decision.

A realistic mindset ties it all together. While 74% of accountants say generative AI speeds up their tasks, only 22% report a substantial increase in overall work rate — the gains are task-level, not transformative. The playbook won't make AI flawless; it makes AI use defensible.

• Validate every output before it reaches a client
• Feed AI structured data, not raw text dumps
• Keep client data inside approved tools only
• Test models on your real work, not vendor claims
• Train your team and document who is accountable

Going Further: Test, Don't Trust — Building AI Workflows You Can Verify

The test-don't-trust principle doesn't stop at chat prompts. When accounting teams extend AI into agent workflows — skills that automate reconciliations, filings analysis, or client reporting — the same skepticism applies, only now the risk is packaged inside someone else's instructions.

The scale of that risk is real. Published research found that roughly 26% of agent skills carry at least one vulnerability, meaning about one in four extensions you might install for your workflow could introduce a security or reliability problem. For a profession where 72% of accounting professionals already fear AI could produce errors or incorrect decisions, unverified add-ons compound the concern.

So what does verification look like for agent skills? The same logic that applies to a model applies to a skill: don't trust the description, test the behavior. That's the approach SkillzBench is building — a free, independent directory of agent skills where every skill is security-scanned, version-pinned, and benchmarked in a sandbox before it earns a trust score.

The pillars mirror the practices we've covered throughout this article:

• Security scanning — every skill is validated, license-checked, and scanned for vulnerabilities before listing.
• Sandbox benchmarking — skills are tested with the same model, with and without the skill, and the Lift in percentage points is published.
• Trust scores — a 0-100 score built from security, quality, license, maintenance, and community signals, with Lift always shown separately.
• Controlled access — a free MCP connector lets you toggle exactly which skills and tools each agent sees, so nothing runs that you didn't choose.

This matters because the alternative — installing skills on faith — contradicts everything the research tells us. The filings study of 5,000 annual reports showed that newer AI models don't mean they're better; co-author Ariel Markelevich's advice was to "be cautious a little bit." Skills deserve the same empirical treatment: measure the lift they actually deliver, not the lift their description promises.

For accounting teams, the practical takeaway is straightforward. Treat AI outputs as drafts, keep humans in the loop, and apply the same rigor to the skills and agents you extend your stack with. Verification shouldn't be a luxury reserved for large firms — it should be free and available to anyone building AI workflows.

SkillzBench is actively building this and is pre-launch. If you want to browse and test free agent skills — with security scans, published Lift scores, and trust scores in place — join the waitlist at skillzbench.com and be first in line when verification-ready skills go live.