Key Takeaways
- When AI-assisted software makes an error on a tax return, the liability doesn't shift to the vendor — it stays with the CPA who signed it. The PTIN doesn't transfer.
- AI tools achieve roughly 85% accuracy on accounting categorizations; the profession requires 100%. That gap forces human review of every AI output and shrinks the time savings vendors advertise.
- The firms closing the AI gap aren't replacing accountants with AI — they're redesigning workflows so AI handles high-volume, low-risk tasks while practitioners own every judgment call.
- Before adopting any AI tool, map your workflows first. Identify exactly where 85% accuracy is acceptable and where the PTIN standard applies.
Every week in 2026, accounting software vendors pitch new AI tools to CPA firms. Yet most practitioners can't tell which tools actually work, which create liability and which are just marketing dressed up as product. General tech media doesn't cover this beat for accountants, and many accounting publications still lag on AI-focused coverage that speaks to how the profession actually operates.
That gap between what vendors promise and what CPAs can actually trust is what this piece addresses. Six practitioners — CPAs, fractional CFOs and AI software operators — share where the skepticism starts, where AI is already earning its place and what redesigning an accounting workflow around AI actually requires in practice.
Why Are Accountants Skeptical of AI Tools?
The skepticism isn't irrational — it traces directly to a liability structure AI vendors have never resolved and rarely acknowledge in their sales process.
George Dimov, CPA, Founder and CEO of Dimov Tax, a full-service CPA firm serving individuals and businesses across the U.S., frames it as the question every vendor should answer but doesn't. "The question that nobody who sells us these tools wants to answer is whose Preparer Tax Identification Number goes on the return when the Artificial Intelligence gets it wrong. Is it Intuit's Preparer Tax Identification Number? Is it QuickBooks' Preparer Tax Identification Number? No. It is mine. I am the one who signs under penalty of perjury, not the people who sell me the tools."
The PTIN is the licensed practitioner's legal signature of record on every return filed in the United States. No vendor has proposed taking that on. None is likely to. That means the accountability structure for AI-assisted tax preparation runs exactly as it always has — with the CPA, not the software company.
Aman Chahal, Industrial Professor of Innovation and Entrepreneurship at the University of Alberta, names the technical reality plainly: "AI is glitchy. Accounting is very important." His broader point gets at why the field resists automation. "Accounting is like money's language. And a good accountant is an excellent translator. It will be hard to automate them away." A domain where a single misclassification can trigger penalties, require amended returns or create audit exposure doesn't reward imprecision.
Who Is Liable When AI Gets a Tax Return Wrong?
When AI makes an accounting error, liability stays with the licensed CPA, not the vendor. Every AI adoption decision flows from that single fact.
Dimov puts the accuracy problem in concrete terms: AI tools achieve roughly 85% accuracy on accounting categorizations. Tax preparation requires 100% correctness. The gap between those two numbers means practitioners can't hand off a return to AI and sign it. They have to check every line. When you account for that review time, the efficiency gains vendors advertise look considerably smaller.
Lakshya Jain, Director of Finance / Mortgage Technology at Annaly Capital Management, works in a regulated environment where explainability in AI decision-making carries compliance weight. "Most companies are not struggling to access AI tools. They are struggling with trusting the tools and understanding how they work. If you can't explain the decision, it's hard to sign off on it."
That explainability constraint is why firms start AI adoption on low-risk, high-volume tasks: document data extraction, automated workflow routing and statistical outlier flagging. It's worth being precise here: AI-assisted tax preparation means using AI for data ingestion, categorization and anomaly detection, not for signing off on a return. High-risk decisions — signing a complex tax return, concluding an audit — stay with a licensed professional. The pattern isn't timidity. It's a rational allocation based on where AI accuracy is sufficient and where it isn't.
Here's how AI accuracy compares with the precision the profession actually requires across core workflows:
| Workflow Task | AI Accuracy Range | Required Standard | Current Approach |
|---|---|---|---|
| Transaction categorization | ~85% | 100% | AI-assisted; full human review required |
| Document data extraction | >95% | High — spot-check acceptable | AI-first; human handles exceptions |
| Anomaly detection and flagging | High | False positives tolerated | AI flags; human investigates |
| Accounts payable automation | High (structured data) | High — approval controls apply | AI routes; human approves |
| Tax return sign-off | Not applicable | 100% — PTIN at stake | Human only |
Where Is AI Actually Helping in Accounting Right Now?
Firms that have moved past skepticism let AI own the back office so practitioners can spend time on the decisions clients actually pay for.
Laura O'Neill, CEO of Breakaway Bookkeeping & Advising, leads a team of more than 50 financial advisors delivering fractional CFO services to small and mid-size businesses. Her team uses AI to streamline routine workflow so advisors can spend more time on strategy with clients. In practice, AI handles accounts payable automation, reconciliations and routine transaction entries. The fractional CFOs and controllers direct their time toward forecasting, client conversations and risk-based judgment — the work that actually justifies the engagement.
Rohit Gupta, Founder and CEO of Auditoria.AI, which runs AI agents inside enterprise ERP systems for AP and AR workflows, names the structural mismatch AI is built to fix: "Business is continuous. Finance is periodic. AI is the bridge."
Transactions don't stop for the month-end close. Vendor payments, payroll, customer invoices and expense reports accumulate daily. Finance teams still operate on monthly closes, quarterly reports and scheduled reconciliations. That gap creates blind spots between reporting cycles — and it's where AI-assisted bookkeeping earns its place. Gupta's platform eliminates the email chaos and manual triage that buries finance teams between closes, not by replacing them but by handling the work that was never a good use of professional time. His summary: "AI can support more continuous visibility across financial activity without disrupting established processes."
How Do You Redesign Accounting Workflows Around AI?
AI workflow redesign starts with mapping your existing processes, not selecting a product. Workflow clarity comes first; tool selection follows from it.
Jain's framing applies here: "Companies are using AI to help people make decisions, not to replace people." Gupta makes the same point from the software side. Both converge on the same design principle — AI belongs in the spaces between decision points, handling document ingestion, workflow routing and exception flagging. Professional judgment stays at the steps that carry legal and fiduciary weight.
John Frank, Founder and CEO of Third Road Management, a fractional CFO and controller firm, has built this model into his practice. AI handles the volume work cleanly enough to trust, at a cost that makes fractional services economically viable for smaller clients. That frees his team to spend engagement time on the strategic relationships and financial analysis those clients actually need. The economics only hold because the AI-assisted work is genuinely handled — not just passed through.
Drawing from Jain's and Gupta's framing, here are three questions to apply to every step in your accounting workflow before bringing in an AI tool:
- Where does accuracy actually matter? Can you accept 85% on this step or does professional liability demand 100%?
- How does the AI explain its output? If the decision gets questioned in an audit, can you reconstruct how the tool arrived at it?
- Where does the PTIN standard apply? Any step that ends with a licensed signature stays human-owned regardless of what the AI produced upstream.
What the Profession Still Needs
The AI gap in accounting isn't a technology problem. It's a standards and liability problem that looks like a technology problem from the outside. The profession needs formal guidance on what documentation of AI-driven decisions should look like for audit and regulatory review — and on whether vendor contracts can or should shift any liability for AI-generated errors away from the practitioner.
Dimov's PTIN framing is already a de facto best practice. Most experienced CPAs carry that rule in their heads. It hasn't been codified and vendors certainly haven't written it into their agreements. CPAs who can articulate a clear AI policy to clients and auditors — knowing exactly what the tools touch, what controls are in place and who owns the decisions — are already better positioned than firms that haven't thought it through.
The firms moving forward aren't waiting on formal guidance. They've drawn their own line: AI handles categorization, ingestion and triage; humans own the decisions that carry professional and legal weight. That line will shift as AI accuracy improves through 2026 and beyond. But vendor-agnostic, plain-language guidance — not vendor marketing — is what CPAs and CFOs need to draw that line confidently in their own practices. That's the gap Nexairi is here to close.
