Abstract

NIST frames AI risk management around governing, mapping, measuring, and managing risks across the lifecycle. [1] Its generative AI profile adds risks such as confabulation, data privacy, information integrity, information security, intellectual property, and human overreliance. [2] Executive support teams can apply those categories without assuming that an AI tool is safe or accurate because it is widely available.

Methodology

This synthesis uses the NIST AI Risk Management Framework and generative AI profile as its control backbone. [1][2][3] Privacy, consumer-protection, international principles, secure-development, application-security, national cybersecurity, and labor sources supply distinct boundaries. [4][5][6][7][8][9][10] It does not estimate time savings, repeat product marketing, or infer a universal risk score.

Map the workflow before approving the tool

Executive support tasks differ sharply in consequence. Reformatting a public agenda is unlike summarizing a confidential board packet, drafting an employment message, or changing a traveler record. NIST recommends mapping context, intended use, affected parties, and foreseeable impacts before measurement and treatment. [1][3]

For each proposed use, name the data source, tool and account, permitted input class, expected output, human reviewer, authoritative record, and prohibited action. The tool must not receive a new category of sensitive information simply because it shortens a draft.

Inputs remain governed data

The NIST Privacy Framework supports identifying data processing, purposes, roles, and privacy risk. [4] An assistant should know whether prompts are retained, used for model improvement, accessible to administrators, transferred to subprocessors, or recoverable after deletion. Vendor settings and contracts require qualified review.

Apply the confidentiality map before a pilot. Public, internal, restricted, privileged, personal, credential, and regulated information may need different rules. Redaction can reduce exposure, but names are not the only identifiers in a distinctive itinerary or board matter.

Outputs require consequence-based review

The generative AI profile treats confabulation and information integrity as specific risks. [2] Review should test facts, cited sources, dates, names, arithmetic, authority, tone, and whether the draft exposes information to the wrong audience. A fluent answer is not evidence that the underlying claim exists.

OWASP identifies risks including prompt injection, sensitive-information disclosure, improper output handling, excessive agency, and overreliance in large-language-model applications. [8] Keep consequential actions outside autonomous execution unless designated security and business owners have approved and tested the complete path.

Vendor claims and security evidence

The FTC Workado analysis alleges that a 98 percent AI detection claim was false, misleading, or unsubstantiated because the company lacked competent and reliable supporting evidence. [5] Procurement should request evidence relevant to the proposed workflow rather than accepting broad accuracy or productivity language.

Secure-development guidance from the NCSC and partner agencies covers design, development, deployment, and operation. [7] CISA places secure and trustworthy AI within a wider cybersecurity roadmap. [9] For executive support, review identity, access, logging, data flow, incident reporting, service changes, and exit procedures before production use.

A bounded pilot and its limits

Choose a low-consequence task with approved data. Record a baseline, error categories, review time, corrected outputs, rejected outputs, incidents, and user feedback. The draft approval queue can keep human authorization visible. Stop criteria should include sensitive disclosure, unsupported claims, or unreviewable changes.

OECD principles emphasize human rights, transparency, robustness, security, and accountability. [6] ILO exposure research concerns how tasks may be affected, not proof that a specific tool is adopted, safe, or suitable for an executive office. [10] Local evidence can support a bounded decision, but it cannot establish universal safety or productivity.

Key findings

  • AI governance begins with a defined use, data class, reviewer, and prohibited action.
  • Confabulation, privacy, information integrity, security, and overreliance are documented generative AI risks. [2]
  • Sensitive inputs remain governed even when placed in a prompt.
  • Consequential outputs need human verification and bounded system authority.
  • Product claims and local pilot results do not establish universal safety or productivity.

Sources

  1. Artificial Intelligence Risk Management Framework, National Institute of Standards and Technology
  2. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, National Institute of Standards and Technology
  3. AI Risk Management Framework Playbook, National Institute of Standards and Technology
  4. Privacy Framework, National Institute of Standards and Technology
  5. Workado, LLC: Analysis of Proposed Consent Order, Federal Trade Commission and Federal Register
  6. AI Principles, OECD
  7. Guidelines for Secure AI System Development, UK National Cyber Security Centre
  8. OWASP Top 10 for Large Language Model Applications, OWASP Foundation
  9. Roadmap for Artificial Intelligence, Cybersecurity and Infrastructure Security Agency
  10. Generative AI and Jobs: A Refined Global Index of Occupational Exposure, International Labour Organization

Return to the Research Content library