Artificial intelligence can enter a small business much faster than the governance structure needed to supervise it.
A business owner can now use AI to categorize expenses, process invoices, analyze cash flow, draft correspondence or examine financial data without maintaining the technology, compliance and risk-management staffs common at larger organizations. Federal research suggests small businesses are already narrowing the AI-adoption gap with larger companies.
That creates a question extending beyond what AI can do: Who is responsible for determining when it should be allowed to do it?
Matan Bar, CEO of Xero US, said small businesses should approach AI in much the same way an employer approaches a capable but inexperienced worker.
“I think of AI in 2026 as a very good intern,” Bar told Checkpoint News. “It can do a lot, but you need to understand its limits and check what it delivers.”
The comparison captures an emerging challenge for accountants and tax professionals. Automation may remove significant amounts of routine work, but greater automation can make professional judgment more important when the work affects taxes, payroll, cash, employees, or other consequential business matters.
That theme was the focus of Accountable Intelligence: Building a Responsible AI Future for Small Business, a Nov. 10 event in New York featuring Bar, Bo Young Lee, CEO of AI4ALL, and representatives from the small business and accounting communities.
Start narrow and earn trust
For small businesses without dedicated AI governance or technical teams, Bar advocates beginning with tasks for which accuracy can be readily evaluated.
“Capturing a bill, reading the invoice details, putting them into your software and categorizing them is a task where AI can be very accurate today,” Bar said.
A considerably more consequential process calls for a different approach.
“Running a full tax cycle with no human involved isn’t there yet, maybe not this year and maybe never,” he said.
The government’s own small business guidance follows a similar incremental philosophy.
The Small Business Administration recommends that companies unfamiliar with AI “start small,” test tools and determine whether they provide value before expanding their use. The agency identifies potential applications including routine administrative tasks and business-data analysis, while also cautioning businesses to consider AI’s risks as well as its benefits.
The approach also aligns with the National Institute of Standards and Technology’s AI Risk Management Framework, although NIST provides a more formal risk-management structure than most small companies are likely to maintain.
NIST organizes AI risk management around four functions: Govern, Map, Measure and Manage. Its framework identifies characteristics including validity and reliability, security and resilience, accountability and transparency, explainability, privacy and management of harmful bias as elements associated with trustworthy AI.
For small businesses, those principles can translate into considerably simpler questions: What is the AI being allowed to do? What data is it using? How will someone know when it is wrong? And what happens when the system is uncertain?
Human oversight has to mean more than approval
Simply placing a person at the end of an automated workflow does not necessarily establish meaningful oversight.
“Oversight has to be more than someone clicking ‘approve,'” Bar said.
A reviewer, he said, should understand what the system was asked to do, the information on which it relied and what the AI is recommending or changing. The person also needs the ability to compare the result with underlying records and either correct the work or escalate the matter when appropriate.
That distinction is particularly relevant for the tax profession.
The IRS Office of Professional Responsibility issued guidelines June 24 specifically addressing responsible AI use in federal tax practice. The agency warned that generative AI can produce fabricated information, commonly described as hallucinations, along with bias and transparency problems. The technology can also create confidentiality and data-protection problems when client information enters AI systems.
The guidance reaches a particularly important conclusion for tax practitioners: using AI does not eliminate the professional’s responsibility to review the resulting work. The IRS said tax professionals using generative AI should carefully review documents created by the technology.
The agency pointed to real-world consequences as well. Courts have sanctioned attorneys over AI-generated hallucinations and false citations, with consequences including financial sanctions, professional censure, ethics training requirements and disciplinary referrals.
That history provides a useful counterweight to the idea that an AI system’s demonstrated reliability alone can justify progressively reducing supervision.
Past accuracy can support trust in a narrowly defined process. It cannot necessarily establish that the next output is correct.
Transparency becomes part of the control
Bar argues that responsibility cannot fall entirely on the small business owner or professional reviewing an AI system.
“The product has a big role to play,” he said.
Systems should make their actions understandable, flag cases in which confidence is low and provide users with the information needed to review results, he said.
“A responsible tool gives you the output and also how the decision was made and based on what information,” Bar said. “Second, it needs a confidence threshold. Below that threshold, it should say, ‘I’m not sure,’ and bring you in.”
The concept closely tracks broader AI risk-management principles.
NIST’s AI Risk Management Framework specifically identifies accountability and transparency, as well as explainability and interpretability, among the characteristics of trustworthy AI.
NIST’s separate generative AI profile is intended to help organizations identify risks particular to generative systems and select risk-management actions appropriate to their goals and circumstances. The framework is voluntary rather than a regulatory mandate.
The distinction matters for small businesses. Responsible AI does not necessarily require creating an enterprise-level AI compliance department, but somebody still must identify which uses deserve heightened scrutiny.
Tax and financial work raise the stakes
For accountants and tax professionals, the consequences become more significant when AI moves from administrative assistance to substantive professional work.
The IRS’s own approach illustrates the distinction. Its revised Internal Revenue Manual establishes governance requirements for agency AI and separately addresses high-impact AI, model and data inventories, generative AI, recordkeeping, taxpayer rights, privacy and security.
The IRS Electronic Tax Administration Advisory Committee likewise called in its 2026 annual report for the government to expand artificial intelligence using appropriate transparency and governance while keeping taxpayers and tax professionals at the center of system design.
For practitioners, another boundary is confidentiality.
IRS guidance says generative AI creates privacy and confidentiality concerns because information entered for one client potentially can be incorporated into systems in ways that affect other interactions.
Accordingly, asking whether an AI tool produces accurate answers is only part of a professional’s evaluation. Firms also need to understand what happens to client information, whether prompts or documents are retained, who can access the data and whether information provided to a model can subsequently be used for other purposes.
Advisors may gain value as routine work disappears
AI could nevertheless dramatically alter the economics of accounting and advisory work.
Bar said accountants still spend substantial amounts of time requesting receipts, locating invoices and resolving unexplained transactions. Those activities can increasingly be automated.
What remains is potentially higher-value work.
“Once the admin is handled, advisors can spend their time on what actually builds the relationship,” he said. “That means analyzing the business’s data, giving cash flow insight and helping the business grow.”
AI might analyze the books, he added, but it may not possess important context about the owner, employees and circumstances surrounding a decision.
That could make the advisor less important as an information intermediary but more important as the person responsible for context, professional skepticism and judgment.
There is a complication, however: effective oversight itself requires expertise.
If AI increasingly performs the entry-level work through which accountants traditionally learned how transactions, reconciliations and workpapers function, firms eventually could find professionals supervising activities they have performed less frequently themselves.
That issue is already surfacing within the accounting profession. A 2026 discussion hosted through an AICPA & CIMA community questioned how younger auditors will acquire sufficient judgment to supervise AI if automated systems perform the procedures through which earlier generations learned auditing fundamentals.
The challenge therefore is not simply keeping a human in the loop. It is keeping a qualified human in the loop.
AI agents raise the accountability question again
That question becomes more consequential as AI evolves from generating answers to performing actions.
An AI application that recommends how an expense should be classified presents one level of risk. An agent that enters the transaction, initiates a payment, modifies financial records or takes another action without contemporaneous review presents another.
Bar said the underlying accountability principle should remain the same.
“Control stays with the human manager,” he said.
AI agents should initially operate in what he described as a learning mode, showing how they work, requesting help when uncertain and gaining independence on particular tasks as performance establishes trust.
That model provides a potentially practical framework for small firms:
- Automate narrow, repeatable work first.
- Require review while the system establishes reliability.
- Make underlying records available to the reviewer.
- Require the system to escalate low-confidence results.
- Apply greater human review as financial, tax, employment or legal consequences increase.
- Keep responsibility for consequential decisions with a qualified person rather than treating AI output as the decision itself.
Those controls become increasingly important as systems gain authority to act rather than merely advise.
Accountability cannot be outsourced to the algorithm
There is also a broader consumer-protection reason not to confuse advanced technology with professional expertise.
The Federal Trade Commission has already pursued companies over claims about what AI can accomplish. Its 2024 “Operation AI Comply” included an action against DoNotPay over claims that its AI service could substitute for a human lawyer when, according to the FTC, the company had not adequately tested whether the service performed at that level.
Small businesses themselves have also been targets of questionable AI promises.
The FTC alleged in 2025 that Air AI used deceptive claims about business growth and earnings potential to sell services to entrepreneurs and small businesses, some of whom suffered losses as high as $250,000. The FTC announced a settlement in March 2026 under which Air AI would be banned from marketing business opportunities.
Those cases do not mean small businesses should avoid AI. They demonstrate why businesses cannot treat the word “AI” itself as evidence of reliability.
Artificial intelligence may democratize capabilities once available primarily to companies with substantial finance, analytics and technology staffs. But easier access to sophisticated analysis also creates a paradox: professional-looking output can make it harder for an inexperienced user to recognize when professional expertise is still needed.
Bar returns to the intern analogy to draw that boundary.
“Use AI where it’s accurate and you can see its work,” he said. “Where it’s unsure, or the work is as consequential as a full tax cycle, keep a person involved.”
For accountants and tax professionals, that may ultimately be the more consequential effect of AI.
The technology can increasingly perform the work.
Accountability for deciding when that work can be trusted remains considerably harder to automate.