The short answer: a tax return has to be exactly right, and today's AI models aren't exactly right often enough. So at Navo, a tax engine that follows the IRS's rules does every calculation, and AI never decides your numbers.
AI can already do a lot of the tedious parts of filing taxes. It can read a W-2, explain what a 1099-K is, and turn a confusing question into a plain one. So it's fair to ask: why not let it do the whole return? This post explains what the public tests show, why tax math is a hard problem for AI in particular, and how Navo divides the work.
What the benchmark found
In July 2025, the tax software company Column Tax published TaxCalcBench, a test of whether AI models can calculate a federal tax return. It gave models 51 realistic 2024 returns, each with every fact already filled in, and compared their answers with a traditional tax engine's. The models didn't have to find or ask for anything. They only had to do the math. The best model calculated 32% of the returns exactly right.
In June 2026, Column Tax released a harder version for 2025 returns. It has 50 returns with realistic W-2 and 1099 forms to read, and most include a state return for California, Illinois, New York or Virginia. Models have improved quickly. On the current leaderboard:
| Returns exactly right | |
|---|---|
| First version (2024 returns), best model | 32% |
| Current version (2025 returns), best model without web search | 44% |
| Current version, best model with web search | about 78% (38 of 49) |
(The 50th return was left out of that last score because the model found the answer key online.)
That's real progress. It also means that even the best AI got roughly one return in five wrong somewhere.
Why "close" isn't good enough
On most tasks, 78% would be a good score. A tax return isn't most tasks.
- The IRS checks the math. Its systems recompute your return. A wrong number can mean a smaller refund, a bill with interest, or a letter months later.
- One error spreads. A return is a chain. A wrong adjusted gross income changes your deductions, which changes your taxable income, which changes your tax and your credits.
- You don't know which return is the wrong one. If a tool is right four times in five, you can't tell from the outside whether yours is the fifth.
- You sign it. When you sign a return, you declare it's true and correct. You need to be able to trust where every number came from.
Why tax math is hard for AI
The original paper describes the kinds of mistakes the models made: using the wrong tax tables, making arithmetic errors, and getting wrong who qualifies for what. Its authors concluded that deterministic tax engines are still needed to get returns right.
Those mistakes come from how language models work:
Models predict; tax rules prescribe
A language model writes the most likely next words based on what it has seen. That's ideal for explaining a concept in plain English. A tax calculation needs the one right answer, computed from a specific table and a specific rule, every time.
The rules change every year
Brackets, the standard deduction, phase-out ranges and credit amounts are adjusted for inflation every year, and laws add new rules. For 2026 alone there are new deductions for tips, overtime, seniors and car loan interest, each with its own phase-out math. A model that learned from older material can apply last year's numbers with complete confidence.
Small rules matter
IRS instructions say things like "round down to the next lower multiple of $1,000" or "for each $1,000 or part of $1,000." Getting those exactly right is what separates a correct return from one that's off by a few dollars, which is still wrong.
Ask twice, get two answers
Ask a model the same tax question in two separate chats and you may get two different numbers. A tax engine gives the same answer for the same facts, every time.
How Navo splits the work
Navo uses software for what it's good at, and AI only where it helps.
The math is done by a tax engine, every time
Navo's engine follows the IRS's rules line by line, the same way on every return. It started from the engine the IRS built for Direct File, and we've extended it to cover more tax situations, including the 2026 changes.
AI helps with the conversation, not the calculation
AI can help you understand a question or find the right box on a form. It never decides your numbers. When Navo needs a fact from you, it asks for it in plain language and the engine does the rest.
We test the engine constantly
Every change to Navo's engine runs against hundreds of test returns before it reaches you. If any result changes without a reason we can explain, the change doesn't go out.
When something's out of scope, Navo says so
If your return includes something Navo doesn't handle yet, Navo tells you plainly instead of guessing. Why Navo tells you when it can't do your return.
What this means for you
When you file with Navo, the amount on your return comes from rules, not from a guess. You can see how your numbers were figured, and you review your return before you sign it.
And when you use Navo with an AI assistant, which is coming soon, the assistant can help collect your information, but Navo does the math, you enter your most sensitive details on Navo, and you sign on Navo.
Filing with Navo is free for the 2027 season.
Frequently asked questions
Does Navo use AI at all? Yes, for help and explanations. Never for the numbers on your return.
Will AI eventually be accurate enough to do the math? Scores are rising fast. But a tax engine is already right every time for the situations it covers, so there's no reason to trade that for "usually right."
Can I use ChatGPT or Claude alongside Navo? Yes. Ask an assistant to explain anything on your return. If its answer disagrees with Navo's number, Navo shows how the number was figured so you can check.
Related guides
- Can ChatGPT, Claude or Gemini do your taxes?
- How to file your taxes free with AI
- How Navo started from Direct File
Navo is a private company and isn't affiliated with or endorsed by the IRS. This post is general information, not tax advice for your situation. Sources: TaxCalcBench paper (arXiv:2507.16126); TaxCalcBench leaderboard.