AI-Generated Tax Research: Why We Use It and Where We Don't

We use AI in our practice. We also do not use AI in our practice. Both statements are true depending on what part of the work you are asking about.

Where we use it: Initial research. If a client question lands on a topic we want to re-verify our understanding on (a specific code section, a recent case, a state rule), the first pass is a query to Claude with a narrow, well-scoped prompt. The output gives us a starting map of the terrain: what the rule says, what the relevant cases have held, what the standard arguments are. From there, we read the primary sources ourselves. We confirm the citations. We read the cases the AI summarized. We verify that what the AI claimed is what the statute says.

This workflow cuts research time by roughly half on technical questions. For judgment questions, the speedup is smaller but still real, because the AI’s first pass usually surfaces the specific sub-questions we need to dig into.

Where we do not use it: Final client deliverables. Our client memos, tax returns, and advisory work product pass through human review at every step. AI output is not the final word on anything that leaves our firm. A return that gets filed has been reviewed by a licensed preparer. A memo that gets delivered has been written and reviewed by a named person. The AI’s role is to accelerate the work, not to sign it.

Why the distinction matters: AI models generate text that sounds authoritative whether or not the underlying reasoning is correct. The failure mode is not random errors. It is confident-sounding errors on edge cases, where the AI’s training data does not perfectly cover the specific situation and the output plausibly looks like it does. A human who notices the edge case catches the error. A workflow that trusts the AI output without review does not.

For a tax practice, the cost of this failure mode is not a bad draft. It is a filed return with a wrong position, or a memo delivered to a client that gets relied on, or an advisory recommendation that produces a liability the client did not anticipate. None of those are acceptable failure modes for our clients, and we will not build a workflow that makes them more likely in exchange for faster production.

The view on firms that are not doing this: We are noticing a growing split in the profession. Some firms have adopted AI aggressively and layered in the human-review guardrails. Other firms have adopted AI aggressively without the guardrails, and the output is reaching clients without the check that would catch the errors. A third group has not adopted AI at all and is running on the cost structure of 2022, which is a different problem.

Our bet is that the first group, the ones adopting AI with discipline, will take market share from both of the others over the next five years. The firms without AI will be too expensive. The firms without guardrails will accumulate errors that eventually show up as problems.

This is the workflow we are running. We think it is the right one.

---

This briefing is for informational purposes only and does not constitute tax advice. The topics discussed depend on specific facts and current law, both of which change. A proper analysis of your situation requires professional review. Contact us to discuss whether this applies to your business.