Considerations for Agentic Research in Public Health

At a glance

This resource offers guidelines to help state, tribal, local, and territorial (STLT) public health agencies use agentic research tools, known as deep research, to enhance decision-making and efficiency, while emphasizing the importance of human oversight.

Best practices for using advanced AI research capabilities to support evidence-based decisions.

Purpose

This informational resource provides principles and practical suggestions for using agentic1 research tools most commonly referred to as deep research, an agentic artificial intelligence (AI) capability that autonomously plans and executes multi-step tasks, in public health settings. It is intended to help state, tribal, local, and territorial (STLT) public health agencies use deep research tools effectively to support evidence-based decision-making, improve efficiency, and accelerate early-stage research and planning. Human oversight and clear expectations are essential when using any AI tools. This informational resource complements the Considerations for GenAI in Public Health, which covers broader GenAI adoption, policy development, and governance considerations.

What is Deep Research?

Deep Research refers to a class of agentic AI capabilities and models offered by several commercial vendors. Unlike traditional chatbots, Deep Research tools are designed to autonomously conduct multi-step research across the web, analyze findings, and generate citation-based reports with transparent reasoning. They can ask clarifying questions, refine their approach, and synthesize information from multiple sources in real time. CDC's internal exploratory evaluation during 2025 found that Deep Research is particularly well-suited for some complex public health tasks, including:

  • Literature synthesis for emerging health threats or intervention strategies.
  • Policy and legal scans across jurisdictions.
  • Strategic planning and scenario analysis.
  • Public health communications and message development.
  • Comparative analysis of programs, interventions, or regulations.

Determining When Deep Research Is the Right Tool

Based on CDC's internal exploratory evaluation, Deep Research tools are most beneficial when used for well-scoped, information-rich public health tasks that require synthesis across multiple sources. It is not a replacement for expert judgment or statistical modeling. The table below outlines suggested use cases for Deep Research tools.

Clearly Defined Scope: Tasks such as analyzing trends over a specific time period, comparing policies across jurisdictions, or summarizing regional media coverage.

Rapid Information Synthesis: Using publicly available information to support grant writing, community health assessments, or legislative briefings.

Using Information-Rich, Open Data Sources: Projects using CDC publications, state statutes, news articles, or public dashboards.

Expert Can Validate Outputs: Prompts and outputs will be reviewed and validated by a qualified subject matter expert before use in decision-making or public communication.

Generating First-Draft Content: Informing downstream work such as health advisories, strategic plans, or leadership briefings.

Restricted Data Sources: Required information is behind paywalls, in proprietary systems, or requires login credentials to access (e.g., subscription-based journals, internal portals, electronic health records, surveillance databases).

Sensitive/Regulated Data: Task involves Personally Identifiable Information (PII), Protected Health Information (PHI), or other sensitive data that cannot be safely exposed to external tools.

Statistical Rigor Needed: Statistical precision or reproducible analytics are needed, such as calculating disease incidence rates or modeling outbreak scenarios.

Professional Judgment Needed: Legal, clinical, or ethical decisions such as interpreting laws or issuing clinical guidance.

Comprehensive Literature Review Needed: Does not replace the depth or rigor of a traditional literature review.

Vague Prompts: Instructions are unclear, contradictory, or too broad; increases risk of hallucinations or incomplete results.

Time Sensitivity: The timeline is too tight to complete appropriate prompt refinement, verification, and SME review.

Note: Examples provided are illustrative and not exhaustive.

Deep Research Prompting: Considerations for Public Health Agencies

CDC's experience with Deep Research tools suggests that their usefulness depends on how well tasks are framed. For public health agencies, prompts could reflect real-world needs such as policy analysis, health communication, or legal landscape reviews. The following considerations are drawn from CDC lessons learned and emerging patterns to help agencies achieve accurate and actionable outputs.

  • Define the scope and deliverables clearly: Include timeframe, geography, topic, and desired format (such as a summary, comparison table, or policy brief). Well-defined, specific prompts improve reproducibility and enable others to validate or update your work.
  • Supply examples, templates, or output formats: Provide a structure or outline to guide the model's response, such as a table comparing state laws or a draft alert format.
  • Break down complex tasks into smaller, sequential prompts: For multi-part questions, start with data retrieval, then follow up with analysis or synthesis.
  • Refine prompts through clarification: Use the model's follow-up questions to improve the prompt before initiating a full research run. Consider refining prompts in a traditional chatbot before using Deep Research tools.
  • Plan for subject matter expert review: Treat outputs as a starting point. Save prompts and track outputs for reproducibility with future review. Consider having a subject matter expert review and validate both prompts and outputs before use in decision-making or public communication.
Use Case
Prompt Intent
Example Deep Research Prompt

Policy Analysis

Analyze public health relationships and partnership frameworks

I am seeking a comprehensive analysis of the relationships between public health systems and Tribal Nations2 in the United States. The research should explore historical, legal, and practical dimensions of these relationships, with a focus on understanding the following key areas: (1) Historical and Legal Context - Overview of existing treaties between Tribal Nations and federal, state, and local governments, with a focus on how these treaties impact public health services and jurisdictional authority; (2) Current Public Health Collaborations - Examples of successful collaborations between public health agencies and Tribal Nations at the federal, state, and local levels; (3) Needs and Priorities of Tribal Nations - Key public health challenges faced by Tribal Nations, (4) Best Practices for Building Trust and Partnerships - Strategies for engaging Tribal Nations in meaningful and respectful collaboration. Present the findings in a structured research format with clear headings and subheadings.

Epidemiological Investigations

Summarize infectious disease outbreaks and sentiment

Provide a summary of recent news articles from [specific time period] highlighting outbreaks of infectious diseases, focus on [insert disease], within the United States, noting regions affected and the severity of each case. Provide an executive summary at the top. Explore the local sentiment around each of the outbreaks as much as possible.

Public Health Communications

Communicate value of rapid risk assessments

Acting as a public health analyst, write a 1-2 page summary piece on the utility of public health rapid risk assessments and how they could be used, using online sources from [specific public health agencies]. Keep the writing style formal, and tailor the use for general and U.S. government, non-technical audiences.

Scientific Planning

Support diagnostic development and research planning

Identify cross-linking reagents used in proteomics studies of norovirus-host interactions. Summarize their use in XL-MS workflows and suggest potential applications for diagnostic development.

Scientific Application

Draft an outline for validation steps and cite relevant CLIA sections and require expert quality review

Develop a method validation plan for the bioinformatic analysis of a Clinical Laboratory Improvement Amendments (CLIA) certified test. Make the analysis pipeline CLIA compliant and write out the validation steps needed to follow CLIA regulations. Use Illumina MiSeq sequencing of e. coli 0157 as an example.

Legal

Identify key legal and regulatory considerations relevant to [specific agreement type], highlighting areas that typically require legal counsel review.

Review this agreement and determine what legal risks it may create for CDC and decide whether it is impermissible for CDC to agree to certain terms. Please briefly explain your determinations and include citations to the relevant laws or policies supporting your determinations. Please ignore if the agreement says "sample" or "draft" and pretend like it is a finalized document.

CDC Effective Use Principles

Based on CDC's internal exploratory evaluation, the following principles support the effective use of Deep Research tools in public health. They align with the NIST AI Risk Management Framework, which emphasizes transparency, accountability, and oversight. Early engagement with legal, privacy, and policy offices helps ensure appropriate use and compliance.

  • Mission Alignment: Tie use to a clear public health goal such as improving efficiency, informing decisions, or enhancing communication.
  • Human Oversight: Have a qualified reviewer verify sources, facts, and conclusions before use or release. A person should be accountable for the final product.
  • Scientific Integrity: Ensure outputs are accurate, maintain scientific rigor, are free from fabrication or plagiarism, and note limits and uncertainties.
  • Disclosure and Transparency: Disclose when GenAI is used in public-facing or significant outputs. Use disclaimers or attribution statements where appropriate.
  • Security and Privacy: Follow agency's data security policy; non-public data may be used only with agency-approved enterprise environments with explicit authorization and security/privacy review.
Print
Content Source:
Office of Public Health Data, Surveillance, and Technology
  1. Agentic artificial intelligence (AI) builds upon the capabilities of generative AI to not just create content, but also to make and adjust plans when the actions required to accomplish a goal are not clearly defined by a user. Unlike generative AI, AI agents can interact with their environment to perform tasks for users. U.S. Government Accountability Office. Science & Tech Spotlight: AI Agents. GAO-25-108519. Washington, DC: U.S. Government Accountability Office; 2025 Sep 10. Available from: https://www.gao.gov/products/gao-25-108519
  2. When conducting research involving Indigenous communities, consult with Tribal partners and follow principles of Indigenous data sovereignty.