What does NIST recommend for deploying generative AI in admin workflows?

NIST AI RMF 1.0 (Jan 2023) defines four functions — Govern, Map, Measure, Manage — for AI risk management. The Generative AI Profile (NIST AI 600-1, July 2024) adds GenAI-specific risks (e.g., confabulation, data leakage, IP exposure) and recommends documentation, evaluation, and human-accountability practices. Both are voluntary frameworks, not regulations.

Are productivity-gain figures from AI studies reliable?

Individual studies (Brynjolfsson/Li/Raymond on call-centre agents, GitHub on Copilot, Microsoft on knowledge work, McKinsey on functional AI use) report meaningful gains on specific tasks under specific conditions. They vary widely by methodology, baseline, and task. I cannot confirm a single universal productivity-lift number for 'admin work' generally — treat published figures as directional and design for measurement on your own workflows.

Should AI ever auto-execute decisions in an admin workflow?

Only for low-stakes, reversible, well-bounded actions where misclassification cost is low and an appeal path exists. For approvals, payments, customer commitments, or anything with regulatory exposure, NIST AI RMF's accountability principle implies a human in the loop. The economically attractive design is usually 'AI proposes, human disposes' for material decisions, with auto-execution reserved for routine, low-risk actions.

What AI Can Remove from Repetitive Admin Work Right Now

An insight on what ai can remove from repetitive admin work right now from CodeCraft Enterprise.