Understand the failure mode
Generative systems can produce plausible but unsupported citations, facts, calculations, or explanations. Tax work adds changing law, jurisdiction, tax-year specificity, and client facts, so fluent language cannot serve as evidence.
NIST's generative AI profile identifies risks and suggested actions within its broader voluntary risk framework. Use it as a governance reference, not a guarantee.
Ground and constrain
Limit the system to approved client evidence and current authoritative sources for the task. Specify tax year, jurisdiction, output type, exclusions, and when the system must return an exception instead of an answer.
Do not provide unnecessary taxpayer data or allow broad tool permissions for a narrow drafting task.
Verify before reliance
Open cited authorities, confirm they exist, apply to the relevant period, and support the proposition. Recalculate material numbers independently and reconcile client facts to source documents.
Require professional approval before generated content becomes advice, a filing position, a client communication, or a return input.
Monitor and learn
Retain prompt or task context, model or workflow version, sources, output, edits, reviewer, and outcome where appropriate. Test known difficult cases after updates.
Track unsupported citations, omissions, contradictions, and overconfident responses by category. Do not hide these events simply to improve a superficial success metric.
Sources and limitations
- AI Risk Management Framework — National Institute of Standards and Technology; reviewed August 15, 2026.
This article is educational and is not tax, legal, accounting, security, or investment advice. Product capabilities and tax requirements can change. Confirm current vendor scope and authoritative guidance for the relevant facts, tax year, and jurisdiction.
How this article was prepared
We separate current sourced facts from operational recommendations, avoid invented performance claims, and show the primary sources and review date used.
Read the editorial methodology