Why Tax AI Must Show Its Sources, Assumptions and Uncertainty

For tax professionals, an AI answer is only useful if they can explain how it was reached. David Yue, co-founder and CEO of Accordance, argues that a confident tone says little about whether a tool understood the facts, used the right source or recognized uncertainty. The practical test is whether a reviewer can follow the reasoning when a client, partner or examiner asks how the team reached a judgment.

Traceability matters more than a confident answer

Yue’s concern is that language models can make an uncertain response sound settled. In tax research, that can obscure the difference between a conclusion supported by the facts and one that merely fits the wording of a question. A useful system needs to expose the steps behind its answer: which facts it relied on, what authority it applied, which alternatives it considered and where uncertainty remains.

That record serves a working purpose. Tax judgments may need to be explained to colleagues and clients, and the reasoning can matter as much as the conclusion. If a tool shows its sources and assumptions, a professional has a basis for checking the analysis, identifying an omitted fact and deciding whether the issue needs human judgment. If the tool gives only a polished answer, the reviewer has less to test.

This changes how teams should assess AI assistance. Fluency is easy to notice; sound research is harder to establish. A response can read well and still fail to address an important part of the fact pattern. Yue’s standard is therefore about evidence and process: the tool should make its work visible enough for a professional to challenge it.

Tax research depends on the weight of sources

A second distinction, in Yue’s view, is whether a tool understands that tax sources do not all carry equal weight. Binding law should take precedence over guidance, and guidance should take precedence over commentary. A system that retrieves relevant passages but treats every source as interchangeable can produce an answer that looks supported while relying on material with less authority.

That hierarchy affects how research should be built and reviewed. The reader needs to know not only which passage supports a conclusion, but what kind of source it is and how it fits into the analysis. Commentary can help explain an issue, for example, but it should not quietly displace a more authoritative source. Professionals still need to assess whether the cited material applies to the particular facts.

Citations also need to be verifiable. Yue says a reference should appear only when the tool has retrieved and read the material it cites. Without that link, a plausible citation may be difficult to distinguish from a source the system has reconstructed from its general training. A reference that cannot be checked does little to support the reasoning. For firms, the useful question is whether each citation leads back to material the researcher can inspect.

Verification can happen during the work

Checking AI-assisted research after an answer is produced can add time and leave gaps for a reviewer to catch late. Accordance’s described approach puts checks into the research process itself. Before returning an answer, the tool checks whether it has addressed every part of the question, found relevant numbers and thresholds, and read the statute or regulation itself. If it cannot confirm those points, it continues working rather than handing over an incomplete response.

Yue also says the tool checks quotations against the source text. That matters because an inaccurate quotation can change the apparent meaning of a passage, even if the surrounding analysis sounds reasonable. Comparing the wording with the underlying material gives the reviewer a concrete way to assess whether the quotation supports the point being made.

These checks do not remove the need for professional review. They can make review more focused by surfacing the facts, sources and gaps that deserve attention. Teams still need to judge whether the question was framed well, whether the supplied facts are complete and whether the cited authority fits the circumstances. A system that reports what it has not established may be more useful than one that conceals the gap behind a smooth response.

The measure is an answer a professional can defend

For firms considering tax research tools, Yue’s comments point to a practical evaluation: ask the system to show its sources, distinguish their weight, confirm that its citations and quotations match the underlying material, and flag what it could not establish. Then assess whether a professional can retrace the analysis without guessing at the model’s steps.

The payoff depends on how well those controls work in real research tasks, including questions with incomplete facts or several plausible interpretations. A tool’s confidence score or writing quality cannot settle that question on its own. What matters is whether its work leaves a clear trail for a person to inspect, correct and explain.