Did I Leave Money on the Table?

What AI Tax Accuracy Really Means for People Who Need Their Taxes Done Right

AI Tax Preparation Accuracy: Evaluation based on TaxCalcBench. Chart comparing Margen (81.5%) to other AI models and benchmarks.

Key Takeaways

  • 1

    Generic LLMs struggle with tax accuracy. An open benchmark shows that leading language models correctly compute only 23-42% of full federal returns, even under simplified conditions.

  • 2

    Margen significantly outperforms standalone LLMs. Using our multi-agent architecture, validation layers, and document-aware workflow, Margen achieved 81.5% strict accuracy on the same benchmark.

  • 3

    A high reasoning baseline opens the door for optimization. By building on top of a hardcoded skeleton of a standard tax preparation software, Margen is capable of exploring optimal paths to savings.

  • 4

    The goal is not just compliance. It's optimization. Margen is designed to evaluate whether there was a better way to file.

81.5%

Margen's Accuracy on Complete Returns

+49.2%

Margen's Accuracy vs. Average of LLMs

For years, the question around AI and taxes has been framed the wrong way.

Can AI calculate a tax return? Can it get the math right? Can it match or beat ChatGPT, Claude, or Gemini on benchmarks?

Those questions matter; but they are not the ones taxpayers are actually asking.

What people really want to know is much simpler:

  • 1.Was everything that mattered included?
  • 2.If there was a better way to file this, would it have been found?
  • 3.Was this handled with the care and judgment of an experienced professional?

At Margen, those are the questions we design for. Accuracy is the starting point, not the finish line.

Measuring Accuracy Is Necessary, but Not Sufficient

Last year, an open benchmark called TaxCalcBench was released to evaluate how well modern AI systems can calculate federal tax returns. It tests models on complete returns under strict conditions, where a return is considered correct only if every evaluated line matches the IRS-expected value exactly.

The results across the industry were eye-opening.

Even the strongest general-purpose AI models struggled. When asked to compute complete returns accurately, most performed in the 20-40% range. These systems often produced answers that were close, but not compliant. And in tax filing, "close" is not good enough.

The IRS does not accept approximate returns. Neither should you.

Where Margen Starts: A Stronger Foundation

When Margen's core reasoning model is evaluated on its own, without any additional safeguards, it achieves 81.5% accuracy on full return calculations.

That alone places it roughly twice as accurate as leading general-purpose AI systems when asked the same task.

But here is the critical point:

We do not ship raw model output to clients.

That number represents capability, not the service.

Where the Real Advantage Comes From

Accurate tax calculation is table stakes. What actually separates a routine return from an optimized one is knowing which paths to consider in the first place.

Tax preparation is not just math. It is a system of elections, thresholds, timing decisions, and structural choices; many of which are invisible unless you know where to look. Experienced tax partners do not just "get the numbers right." They know which options matter, which do not, and which decisions compound over time.

Margen is built to reflect that reality.

Instead of relying on free-form AI reasoning alone, Margen layers a reasoning model on top of a rules-driven foundation modeled after professional tax software and IRS guidance. That foundation does not just enforce compliance; it defines the decision space within which optimization can occur.

This architecture enables three outcomes that matter to taxpayers:

  1. Preserve valid options: Hard-coded tax rules ensure the system never takes shortcuts that eliminate legitimate strategies. Required methods are followed precisely so alternative elections, timing choices, and structural paths remain available.
  2. Evaluate paths, not guesses: Where multiple compliant approaches exist, the system does not default to the first answer it finds. It evaluates those paths against the full return; considering downstream effects, interactions, and longer-term consequences.
  3. Surface experience-driven opportunities: Many tax savings come from patterns learned over decades: how income types interact, when certain elections matter, which thresholds quietly change outcomes. Margen's intelligence layer is designed to recognize and test those patterns systematically, not rely on intuition or luck.

When this optimization-first validation layer is applied, Margen delivers more than an IRS-compliant return.

It delivers a return that reflects the same kinds of choices a seasoned tax partner would explore; with the confidence that nothing material was missed, and no defensible opportunity was left on the table.

Accuracy vs. Optimization: Why This Matters to You

Most tax software, and most AI tools, answer a narrow question:

"Given this input, what is the result?"

Margen answers a broader one:

"Given this situation, what is the best way to file?"

That distinction is everything.

Two returns can both be "accurate" and still produce very different outcomes depending on:

  • Elections made or deferred
  • Timing decisions
  • State exposure handling
  • Structural assumptions
  • Credit and deduction interaction

A correct return is not necessarily an optimized one.

Margen's system is designed to evaluate those paths, not just compute one of them.

Tax Preparation Is a Process, Not a Calculation

Tax preparation is a process, not a calculation. It involves gathering documents, making elections, evaluating tradeoffs, and producing a return that your team can sign off on. Margen is built to support that process from start to finish.

What This Means for People Filing Their Taxes

If you are using Margen, the questions you should walk away confident about are:

  • Nothing material was overlooked
  • If there was a smarter filing path, it was evaluated
  • The return reflects deliberate decisions, not defaults
  • The work meets a professional, audit-aware standard

That is the promise.

Why This Is a Premium Service

Margen is designed for people who care about outcomes.

People whose taxes are complex enough, or important enough, that "probably fine" is not acceptable.

People who do not just want to file, but want to know:

"Was this the most I reasonably could have saved?"

That question deserves a real answer.

The Bottom Line

AI tax accuracy is not about beating other models on a chart.

It is about delivering confidence:

  • Confidence that the return is right
  • Confidence that it is complete
  • Confidence that no obvious opportunity was missed

Margen exists to provide that confidence: consistently, deliberately, and at a professional standard.

From preparation to e-filing and IRS acknowledgment, the entire process is handled in one platform.

It is peace of mind.

Sources


  • Column Tax. (2025). TaxCalcBench: Evaluating Frontier Models on the Tax Calculation Task. (M.R Bock, Molisee, Ozer & Shah, 2025)
  • Column Tax. (2025). TaxCalcBench GitHub Repository. GitHub.
  • Column Tax. (2025). TaxCalcBench: A first-ever benchmark for evaluating AI's ability to calculate tax returns.
  • Griffin, A. (2025). Measuring AI Tax Accuracy: Comparing Filed to ChatGPT, Claude, and Gemini on an Open Benchmark. Filed.

Ready to see what Margen can do?

Evaluate last year's return for free. See if there was a better way to save.

Schedule a demo