NVIDIANVIDIA-Certified Professional: Agentic AI2026-109.5 Ensure compliance with licensing and regulatory standards.
Connect agent risks to accountable controls
Completed by the knowledge check
Learning objective
Map agent risks to regulatory frameworks and maintain compliance evidence.
Core idea
For an enterprise agent, start compliance assessment with its intended use, affected people, processing locations, data, and your organization’s role. “Uses an LLM” does not determine its obligations. An assistant that finds a vacation policy and a system that ranks job applicants can share a model while creating different consequences and regulatory duties.
How it works
Establish scope before choosing controls
Produce an intended-use assessment that records the supported task, prohibited uses, affected population, decision authority, data flows, and provider/deployer roles. Include downstream tools: an agent that retrieves an application and one that rejects it have different effects.
The EU AI Act’s Article 6 and Annex III classify specified uses as high-risk, including recruitment systems that analyze or filter applications and evaluate candidates. Article 6(3) provides a qualified exception for some Annex III systems, such as narrow procedural tasks, where its conditions are met. Annex III systems performing profiling of natural persons remain high-risk. Article 6(4) requires a provider claiming the exception to document its assessment. A human review button alone does not establish the exception.
An Annex III system remains high-risk unless it meets the Article 6(3) conditions without profiling of natural persons, and a provider claiming that exception documents its assessment.
Resolve the role in the same record. The Act’s provider is the body that develops an AI system, or has it developed, and places it on the market or puts it into service under its own name, where putting into service includes supplying it for its own use; the deployer uses it under its authority. An organization that builds its own agent and runs it is both, and the obligations of both apply.
The Commission’s AI Act Service Desk reproduces the consolidated Article 113 schedule: Chapter III, Sections 1–3, except Article 6(5), apply from 2 December 2027 for Article 6(2)/Annex III systems and from 2 August 2028 for Article 6(1)/Annex I systems. Record the provision and applicable date together. Sector rules stack on top of these horizontal acts: a medical-device agent, for example, also answers to the device regulator.
Distinguish risk guidance from legal duties
NIST AI RMF 1.0 is voluntary. Its GOVERN, MAP, MEASURE, and MANAGE functions organize the work; they do not grant regulatory approval. Use GOVERN 1.1 to record legal requirements, MAP 1.1 to document intended purposes and context, and MEASURE 2.1 to retain test sets, metrics, and evaluation-tool details. MANAGE 4.1 connects monitoring with appeal, override, incident response, recovery, and change management.
The Generative AI Profile, NIST AI 600-1, adds risk categories including confabulation, data privacy, harmful bias, and intellectual property. A plausible but false candidate summary is a confabulation risk; unauthorized disclosure of an applicant’s record is a privacy risk. The categories tell you what to test without replacing the law’s own classifications.
Keep a risk register keyed to the intended use and affected people. For each risk, record its source, potential harm, control, measurement method, residual risk, and accepting owner. For a high-risk system, that register is also how a provider meets AI Act Article 9: “A risk management system shall be established, implemented, documented and maintained”, run as “a continuous iterative process” across the system’s lifecycle. NIST supplies the structure; Article 9 makes the process a legal duty.
Inherit the model card’s disclosures
Seed the register from the model card. NVIDIA model cards carry Model Card++ Bias, Explainability, Safety & Security, and Privacy subcards. The NVIDIA-Nemotron-Nano-9B-v2 Bias subcard answers “Which feature(s) have the worst performance overall?” with “Age” and “Measures taken to mitigate against unwanted bias:” with “None”. Its Explainability subcard lists, under “Potential Known Risks”, greater susceptibility to prompt injection and jailbreaking, and says the model “should be paired with additional rails or system filtering”.
Enter each disclosure as an inherited risk with its card revision, then test the gap in your own workflow: age is a protected characteristic in hiring, so an age disparity the provider left unmitigated becomes your measured risk. See 9.3.1 for locating a disparity’s origin and 9.3.2 for building the tests.
Maintain the required assessment records
GDPR Article 5 establishes principles including data minimization and storage limitation; Article 6 supplies the lawful-basis framework. Article 30’s records of processing activities include purposes, data and recipient categories, transfers where applicable, and envisaged erasure periods where possible, subject to its scope and exceptions. Record these facts across retrieval, memory, tool calls, and telemetry, not only the chat interface.
Article 35 requires a data protection impact assessment before processing likely to result in high risk to people’s rights and freedoms. Article 22 addresses decisions based solely on automated processing with legal or similarly significant effects, with specified exceptions and safeguards. Route those applicability questions to the privacy owner; consent, a redaction rail, and a human reviewer answer different questions.
For high-risk AI systems, distinguish provider documentation from deployer operations. AI Act Article 11 addresses technical documentation, Article 12 logging capability, and Article 14 human oversight. Article 26 requires deployers to assign competent, trained, authorized oversight and monitor use. Its paragraph 6 keeps automatically generated logs under the deployer’s control for at least six months, unless applicable Union or national law, including data-protection law, provides otherwise. That qualified rule is not an instruction to store all raw prompts.
Link each control to its verification
Create a control register with fields such as obligation, scope, control implementation, responsible owner, release identifier, test result, retained record, and next review. Keep it versioned beside the release it describes; the compliance owner maintains the obligations and engineering the results. Link the exact system configuration and evaluation dataset revision that produced each result; a passing report from another configuration proves nothing about the deployed one.
See 9.2.1 for PII redaction, 9.2.2 for retention, and 9.1.3 for action-record integrity; here you connect those controls to a requirement. For deletion, retain the policy and test result showing which stores were exercised, rather than making the compliance record another unrestricted copy of the deleted personal information.
Key distinctions and trade-offs
Choose the record by the question it answers:
| Framework or record | Deciding question | Mechanism | Cost and owner |
|---|---|---|---|
| EU AI Act assessment | What intended use, operator role, and obligations apply? | Article 6, Annex III, and provision-specific application schedule | Legal classification and engineering scope review |
| GDPR assessment | What personal-data processing is lawful and proportionate? | Articles 5, 6, 30, and 35; Article 22 where relevant | Privacy review and data-flow maintenance |
| NIST risk register | What risks need measurement and treatment? | AI RMF functions and AI 600-1 risk categories | Risk-owner analysis and recurring evaluation |
| Control register | What proves the approved control operated? | Requirement-to-configuration-to-result linkage | Engineering verification and controlled retention |
Architecture / decision scenario
A European recruiting firm processes 8,000 applications monthly. It builds its own assistant on NVIDIA-Nemotron-Nano-9B-v2 to summarize applications and propose a ranked shortlist, and it uses the assistant itself, so it is the provider and the deployer. The release is planned for March 2027. Recruiters can review 400 applications weekly, while the proposed automated rejection path would remove 70% before review. The product owner suggests calling the system “advisory” and retaining only its aggregate accuracy score.
The compliance owner requires an intended-use assessment covering ranking and rejection, not just summarization. Engineering disables the rejection tool for the first release. Ranking still evaluates candidates, the use Annex III point 4(a) lists; removing autonomous rejection does not by itself remove high-risk classification. GDPR applies from the first applicant processed in March, so the privacy owner completes the lawful-basis and impact-assessment work before go-live and examines the Article 22 implications of any later rejection workflow. The Chapter III duties, Articles 9, 11, 12, 14, and 26, apply from 2 December 2027, nine months later; the control register lists them with that date and an owner, so the records exist before they bind.
The risk owner enters the card’s disclosed age weakness and confabulation as risks with versioned tests. A later request to restore rejection reopens the assessments before the tool is enabled. The trade-off is reduced automation and more recruiter work in return for a release whose authority and supporting records match its assessed purpose.
Common exam traps
- Treating a voluntary NIST assessment as legal certification.
- Classifying the system by model name instead of intended use.
- Assuming human review automatically removes high-risk status.
- Treating a firm that builds and runs its own agent as only a deployer.
- Using a PII detector as proof of lawful processing.
- Retaining raw personal information indefinitely “for compliance.”
- Reusing an assessment after enabling a consequential downstream tool.
Where the exam guide differs
The study guide’s suggested readings call the AI Act a “Proposal.” Regulation (EU) 2024/1689 is enacted legislation, and the Commission’s Service Desk incorporates amendments in its consolidated text. Read an exam item’s reference to the proposal as a pointer to risk-based AI regulation; assess a deployment against the applicable enacted provisions and schedule rather than the proposal’s status or the original application dates.
What to remember
Document intended use and legal applicability, organize risks with a named framework, and connect each obligation to a tested control and owner. Maintain assessments when purpose, data, tools, or authority change. Records support the organization’s compliance decision; a guardrail score alone cannot make it.
Knowledge check
An EU employer’s agent summarizes job applications and proposes a shortlist ranked by scoring each candidate’s skills and experience. The automated rejection tool is disabled, and a recruiter reviews every shortlist before anyone is contacted. The product owner wants to record the system as outside the EU AI Act’s high-risk category. Which classification does the intended-use assessment support?
A company builds its own agent for screening internal transfer candidates and runs it only for its own HR team, under its own name. The compliance lead argues that, because nothing is sold or supplied to another organization, only deployer obligations apply. How should the AI Act role be recorded?
A provider’s recruiting agent is high-risk under the EU AI Act. The risk team has completed a NIST AI RMF assessment covering GOVERN, MAP, MEASURE and MANAGE, with NIST AI 600-1 categories such as confabulation and data privacy. How should the team use that assessment?
A deployer runs a high-risk recruiting agent whose automatically generated logs are under its control. No applicable Union or national law changes Article 26’s default retention rule. An engineer proposes keeping every raw prompt, response and tool payload indefinitely. Which retention rule should the control register record?
A bias test for a candidate-ranking agent passed in staging against an earlier model revision and prompt configuration. The release now in production uses a newer revision and a changed prompt. An auditor asks the control register for evidence that the bias control operated in this release. Which record answers the request?
References
- European Union, Regulation (EU) 2024/1689 (Artificial Intelligence Act)
- European Commission, AI Act Service Desk: Article 113
- European Commission, AI Act Service Desk: Article 26
- European Union, Regulation (EU) 2016/679 (GDPR)
- NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0)
- NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
- NVIDIA, NVIDIA-Nemotron-Nano-9B-v2 model card
- NVIDIA, NVIDIA-Certified Professional: Agentic AI Exam Study Guide