AI accountability: tying every automated call to a named owner
An internal triage system flags a contractor invoice or recommends a maintenance deferral, an operator clicks approve in a hurried morning standup, and three months later an auditor asks why the call was made. Nobody can point to the input snapshot, what threshold was active, which human delegated authority, or whether anyone was tracking drift after launch.
Author: Escherbridge. Published . Last reviewed: , America/Denver.
Two widely cited public references tackle that exact visibility gap without reducing it to compliance theater. NIST's Artificial Intelligence Risk Management Framework (AI RMF 1.0, NIST AI 100-1, January 2023) organizes ongoing risk work around four continuous functions—GOVERN, MAP, MEASURE, and MANAGE—explicitly noting that "Actions do not constitute a checklist, nor are they necessarily an ordered set of steps," and that "Risk management should be continuous, timely, and performed throughout the AI system lifecycle dimensions."
Before turning oversight into process jargon, the U.S. Government Accountability Office (GAO-21-519SP, June 2021) identifies why AI oversight breaks in the real world: "AI systems pose unique challenges to such oversight because their inputs and operations are not always visible." We pair that framework analysis with the engineering essay A recommendation still needs a name on it.
How GOVERN, MAP, MEASURE, MANAGE, and GAO principles work in practice
Directed under the National Artificial Intelligence Initiative Act of 2020 (P.L. 116-283), NIST AI RMF 1.0 is "voluntary, rights-preserving, non-sector-specific, and use-case agnostic," listing trustworthy system traits including valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy enhanced, and fair with harmful biases managed. Its generative companion, NIST AI 600-1 (July 2024), extends those same four functions specifically for generative models.
Meanwhile GAO-21-519SP frames operational accountability across governance, data, performance, and monitoring—opening with two deceptively sharp engineering practices: Practice 1.1 ("Define clear goals and objectives for the AI system to ensure intended outcomes are achieved") and Practice 1.2 ("Define clear roles, responsibilities, and delegation of authority for the AI system to ensure effective operations, timely corrections, and sustained oversight"). If an engineer cannot point to an owner when an output drifts, Practice 1.2 already failed.
Operationalizing the paper trail
Reading the frameworks only takes an afternoon; wiring them into release workflows takes disciplined logging—capturing what to write down in What to write down when software recommends alongside integrating security controls across SSDF lifecycles.
Scope and boundaries
This guide is an architecture reference, not legal advice and not a certification claim. It does not claim that PrimusNeo, NEOS, or AZOA implements either federal framework, treats PrimusNeo as a federal agency system, invents failure statistics, or does not use Standish CHAOS rates.
A situation like this
A team deploys an automated priority scorer for field repairs, hits a miscalibration after seasonal sensor changes, and discovers no owner or evaluation baseline was logged at launch—leaving nobody authorized to freeze the model.
Sources
- NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf · DOI: https://doi.org/10.6028/NIST.AI.100-1
- U.S. Government Accountability Office, Artificial Intelligence: An Accountability Framework for Federal Agencies and Other Entities, GAO-21-519SP, June 2021. https://www.gao.gov/products/gao-21-519sp
- NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, July 2024. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf · DOI: https://doi.org/10.6028/NIST.AI.600-1
Also in this section
- What is a digital twina representation you can evaluate, and who accepts a decision made with it.
- Idempotent HTTP requestsa protocol habit for an effect that should happen once.
- What is local-first softwarewhere the primary copy lives.
- SSDF: add security practices to the lifecyclepractices inside the lifecycle, not a phase at the end.
- What to write down when software recommendswhat to write down, not a second framework overview.
The essay form of this page is A recommendation still needs a name on it. The useful next step is to name the role, the data, the check, and the monitoring, or to read the two frameworks themselves. If something on this page is wrong, write to us.