Best practices for using AI-Assisted Research

Follow these recommendations to get the most out of AI-Assisted research.

Use effective queries

  • Use Keywords or Questions
    : Enter clear, concise keywords or questions.
  • Exclude Personal Information
    : Avoid including personal or client-specific data.
  • Focus on Essentials
    : Only include terms essential to your research question. Don't include extra background facts.
  • Simplify Complex Situations
    : Avoid lengthy, detailed factual scenarios. Checkpoint content has illustrations and examples to help you better understand and apply tax concepts and rules, but isn't intended to give advice or detailed analysis of every possible situation.
  • Single Issue Focus
    : Address one issue per query.
  • Natural Language
    : Use natural language without Boolean strings or special characters.
  • Avoid prompts
    : Don't write a query in the style of a prompt, command, or instruction. It works best with tax research queries that clearly identify the information you would like to find. For example, avoid including extra words or phrases like "find code sections and cases” that bookend the query but are not substantive. Instead, construct queries with words or phrases you would expect to find or be addressed in Checkpoint content. For example:
    When is an S election timely?
    Or
    What is the credit for adopting a child with special needs?
  • Clarify your writing
    : Add punctuation (U.S. instead of US), use identifiers (IRC 179 instead of 179), and use full names versus abbreviations (Maine instead of ME).

Verify AI-generated responses

AI responses include source citations for verification. Verify statements using provided text snippets, or select the reference material to review the full document. While the LLM may use different phrasing than appears in the text snippet, you can usually verify the accuracy of statements in the AI response using only the text snippets and not needing to open and review full-text documents.

Ask follow-up questions effectively

You can ask more than 1 related question in a thread. You may need to expand your understanding of a research question or issue, or you'd like to review more than 1 related AI response on a given tax topic or issue at the same time.
AI-Assisted Research uses context from previous questions and answers when formulating responses to additional questions you ask. You won't need to repeat key terms. For example, if you begin your conversation with
What is the credit for adopting a child with special needs?
, you might also then ask:
  • How do I report it?
  • Are there any income limitations?
  • Can the credit be carried forward and for how long?

Troubleshoot Non-Responses

  • Rephrase Queries
    : If the AI can't generate a response, try rephrasing your query.
  • Focus on Federal Tax Law
    : Make sure your question relates to federal tax law.
  • Use Appropriate Tools
    : For specific tasks, use other Checkpoint Tools like State or International Create-a-Charts.

Understand restrictions and limitations

  • Response Variability
    : AI responses may vary slightly due to the nature of Large Language Models (LLMs).
  • Citation Gaps
    : Some sentences may not have citations. While you'll want to review these responses more carefully, the absence of a citation doesn’t necessarily mean the LLM hallucinated (made up information).
  • Checkpoint resources
    : For Federal queries, citations come from 5 Checkpoint resources: Federal Tax Coordinator Analysis, Federal Tax Coordinator Client Letters, Checkpoint Catalyst, PPC Deskbooks, and Federal Tax Handbook. For State queries, citations come from State Explanations and State Catalyst content. You may be subscribed to other Checkpoint resources that aren't currently cited. As we evolve the experience, we'll consider applying other resources.
  • No External Uploads
    : Don’t upload Checkpoint content into external AI tools like ChatGPT. Our agreement prohibits you from uploading content from Checkpoint into any external application, including any uses for artificial intelligence technologies such as large language models, generative AI, or training a machine learning or AI system.

Related Content