The Tax Attorney Who Tests AI Where the Numbers Meet the Law

By Richard Thornton, JD, LLM (Tax) — 2025-12-05

Richard Thornton has spent seventeen years in tax — first in international tax structuring at a leading Wall Street practice, then leading the technology transactions group at a major advisory firm. He sees AI entering tax practice with the same pattern he's watched other technology enter professional services: impressive on simple cases, dangerous on complex ones.

The surface layer and what's beneath it

"Tax has a surface layer that AI handles well," he says. "Standard deductions, common credits, straightforward entity classification. Below that surface is where it gets dangerous. Transfer pricing. Section 382 limitations. GILTI high-tax exclusion elections. Foreign tax credit planning. These require judgment about regulatory intent, administrative guidance interpretation, and how courts have treated ambiguous statutory language. AI that learned from tax returns and basic guidance documents has gaps here that aren't obvious."

Thornton joined IXO to work on tax AI evaluation — specifically focusing on complex transactional tax analysis and international tax contexts. His annotations address not just whether an AI's tax conclusion is correct, but whether its analysis would satisfy a reasonable IRS examiner, whether it identified and addressed the right authorities, and whether it flagged the planning alternatives a sophisticated advisor would have raised.

The compounding error problem

"The cost basis problem is a good example of where AI fails in ways that compound. A model that handles straightforward cost basis misses the election implications, the wash sale interactions, the constructive sale rules. Each gap creates downstream errors. Annotation that traces those chains — not just flags the endpoint error — is what actually improves the model."

Audit defense

He's particularly focused on AI performance in audit defense contexts. "Tax AI that helps identify positions is one thing. Tax AI that helps defend those positions under examination is significantly more valuable and significantly harder to build well. The adversarial dynamic requires someone who's sat across from examiners."

Read The Tax Attorney Who Tests AI Where the Numbers Meet the Law on the IXO blog