Generative artificial intelligence is being used for almost everything now. It drafts your emails. It summarizes your meetings. It helps doctors write clinical notes and helps engineers write code. In a few short years, it went from novelty to infrastructure across nearly every professional domain. 

One area of adoption, however, is showing signs of strain: tax and accounting.

According to the Thomson Reuters Institute's 2026 AI in Professional Services Report, a survey of 1,514 professionals across 26 countries, only 34% of tax firms are currently using GenAI at an organizational level, and only 14% say agentic AI is currently part of their workflow. That's not necessarily resistance to change. To a large extent, it reflects professional judgment applied to tools that, until recently, weren't built for the job. 

Why LLMs struggle with tax work

Large language models are extraordinarily good at understanding and generating language. They are not, by design, reliable calculators. Ask one to summarize a complex tax code provision and it will do so with impressive fluency. Ask it to compute your client's effective tax rate across multiple jurisdictions and you are taking a risk that no CPA should be willing to take. 

The gap runs deeper than math. General-purpose AI tools were not built for tax. Their training is broad rather than deep. Their knowledge has cutoff dates that don't account for constantly evolving tax law. And they present data security and liability risks that no responsible firm should ignore. 

What makes this particularly dangerous is the nature of AI errors. Hallucinations are not obvious mistakes. They are polished, confident, and often sound exactly like authoritative guidance, while being subtly wrong or entirely fabricated. That makes them especially treacherous in accounting, where errors live in footnotes, scope exceptions, and judgment calls rather than obvious miscalculations. 

Questions to ask before adopting tax AI

Those in the accounting profession that are resisting AI aren’t necessarily doing so out of stubbornness. In many cases, it’s because the AI on offer was genuinely wrong for the job. That calculus is shifting, but not uniformly.

Some workflows are ready for AI today. Others require more careful evaluation before any firm should commit. The difference comes down to whether the technology was built for the realities of tax work. These four questions can help firms tell the difference:

1. How is the AI's content sourced and maintained? General-purpose tools are trained on broad data with knowledge cutoffs. Tax law changes constantly. The AI you use for professional work should be fiduciary-grade and grounded in authoritative, continuously updated sources — not web-scraped data with an expiration date.
2. Does the AI calculate, or does it use a calculator? This distinction matters more than it sounds. An AI that computes tax figures using its own language model is fundamentally unreliable for professional work. An AI that integrates with a validated tax engine and uses that engine to compute is a different architecture entirely. Ask specifically how the tool handles multi-step calculations.
3. Can you audit the output? For any work product that carries your name, you need to be able to trace how a conclusion was reached — what sources were consulted, what assumptions were made, and what the AI reasoned from. If the output is a black box, it cannot meet the professional standard the IRS or a client challenge will require.
4. How does the system handle tax law changes? New guidance, revised rulings, and amended code land constantly. Ask how quickly the AI's knowledge base is updated when tax law changes, who is responsible for that update process, and what happens in the window between a change and the system reflecting it.

The accounting profession is entering a genuinely new phase. The AI that will actually serve it isn't the AI that promises the most — it's the AI that was built for what the profession actually requires.