Martijn Lancee has an accountant, but he still hates doing his taxes. They're complicated: He and his wife have a mortgage on a home in the Bay Area, she owns a small business, and he makes money consulting for companies on artificial intelligence. Each year his tax adviser sends him a long list of tedious requests for information on business expenses, bank statements, bills. In February, Lancee had an idea. "Hey," he typed into Claude Code, "how can you help me file my taxes?"
At the chatbot's instruction, Lancee downloaded a bunch of his tax documents as PDFs on his computer and dropped them into a folder on the desktop for Claude. He asked it to create a spreadsheet organized with several tabs the way his accountant likes. Lancee checked the work, sent the file to his adviser and then spent the rest of his night playing Mario Kart with his kids. He didn't tell his adviser the file was AI-generated. "I'm sure he noticed," Lancee says, "because the output was a lot better than last year."
Interviews with more than a dozen tax professionals offer a clear takeaway: This is the year AI has come for tax prep. Artificial intelligence has finally improved to the point where it at least seems capable of navigating the US tax system as leading AI developers release more features for financial services. About a quarter of US workers say they plan to use AI to help file their taxes this year, up from 11% last year, according to a new survey by Adobe Inc. What better way to untangle the tax code -- something so sprawling and complex that even the most experienced professionals can't master every detail -- than to use computer code?
The problem, according to tax pros, is that the chatbots keep messing up. They can give bad planning advice, sometimes informed by outdated rules, and tend to make mistakes reading digits off tax documents, especially on less standard forms such as a K-1, for so-called pass-through entities, or a 1099, for various kinds of payments. "Tax is incredibly nuanced," says April Walker, a senior manager at the American Institute of Certified Public Accountants.
Recent tax changes, if anything, have added to contradictions nested in the code. "There are so many exceptions," says Misty Erickson, content manager at the National Association of Tax Professionals. "You can even have an exception to the exception."
Benjamin Cox, a business owner in Beaverton, Oregon, has tried using AI to prepare for conversations with his accountant. He asked ChatGPT whether he's eligible for tax breaks such as the federal deductions for qualified business income or those for state and local taxes, known as SALT. The bot, according to Cox, was oblivious to the law approved last year making each provision more generous. "I kept on trying to correct it," he says. "I'm sitting here arguing with AI, which is sort of like arguing with a pig. You feel stupid, but the pig's just having fun."
One technique software engineers use to get the best code out of AI is to deploy multiple chatbots to work on a project at once. Zhiyao Pei, a credit-analytics consultant, asked three AI models to offer their assessment of how California's business-franchise tax would apply to his new business. "They actually had a fight," he says. Google Gemini derided the work of another model as a hallucination. "That's a typical AI mistake," Gemini wrote. In the end, Pei just paid California the highest amount the chatbots suggested.
Joshua Youngblood, whose business specializes in assisting people who have run-ins with the IRS, sees an opportunity in all this. "I hate to be that person," he says, but "it's definitely going to create some issues." One customer is already in trouble with the IRS for failing to report a cryptocurrency transaction after an AI tool erroneously told the client that income of less than $3,000 didn't need to be reported. "ChatGPT is very good at telling you what you want to hear," he says.
The most effective users of AI understand what it's good at -- and what it's not. Large language models on their own aren't skilled with numbers or adept at navigating a complex branching system like the tax code that's supposed to lead to a single answer. A Loyola University Chicago study in November tested LLMs on a common tax question -- how much of a home sale is exempt from the capital gains tax? -- and found the bots failed two-thirds of the time. "It makes numbers up," says Joshua Scott, whose Greensboro, North Carolina, firm specializes in the returns of health-care professionals.
But modern LLMs have capabilities that can help compensate for their shortcomings. Some of the accountants who complain most loudly about AI are the ones who use it the most -- including Scott, who says he's careful to check all its work. He also relies on AI to develop his internet marketing content. Nayo Carter-Gray, owner of 1st Step Accounting in Baltimore, uses it to walk clients through difficult concepts, such as "breaking down what depreciation is," she says. Given the weaknesses of LLMs, Michael Geller, a partner at Gursey Schneider in Los Angeles, says, "I'm not afraid of it taking my job." Still, he urges colleagues to adopt AI as "a new tool to do things quicker than before."
Will an MBA Pay Off for You?
Will an MBA Pay Off for You?
Will an MBA Pay Off for You?
Use our calculator to determine the return on your investment.
Use our calculator to determine the return on your investment.
Use our calculator to determine the return on your investment.
Crunch the numbers Crunch the numbers Crunch the numbers
Crunch the numbers
Several software startups offer AI services for tax preparers but say they put careful limits around what the AI is allowed to do. It can be "dangerous," says David Yue, chief executive officer of two-year-old Accordance, which helps preparers with research and other tasks. Tax returns contain a lot of numbers, and the IRS doesn't have much tolerance for errors, says Dave Haase, founder of Juno, another tax software startup. "If you have a 99% accuracy rate," he says, "you have a mistake on every tax return."
TurboTax, the dominant tax filing software in the US, has been using AI to answer customers' questions and help them navigate the app for years, but the implementation is deliberately limited, says Keela Robison, vice president for product management at parent company Intuit Inc. "As of right now, we do not feel that LLMs can process taxes, certainly not complicated taxes," she says.
Some of the big AI companies recognize the potential. Felix Rieseberg, an engineering lead on Claude Code and Claude Cowork at Anthropic PBC, posted on X that he'd used Claude Cowork to read his own financial documents and fill out forms on a tax-filing site. The owner of X, Elon Musk, advertises his chatbot as a tax adviser. Musk's xAI has been hiring accountants, seeking those with experience at Big Four accounting firms, to train its AI models, according to the company's website.
The hype around AI and the annual feeling of dread associated with tax season together make this combination hard to resist. Andrew Pierno, a software entrepreneur in Los Angeles whose company has acquired multiple startups, has a complicated tax picture. Tax prep usually takes an entire weekend, "a three-day epic adventure where you're unhappy and drinking a lot," he says.
This year he's been obsessed with OpenClaw, software that allows him to control his computer by sending messages to a bot through WhatsApp. Between games of pickleball in February, he texted his Claw asking it to build a tool that would pull together his personal and professional financial information into a file for his accountant. The whole thing took about three hours and was, in his words, "wonderful and fun." He couldn't immediately verify the level of fun his accountant experienced.
