SkarpChapter 6 of 13
Manage: Acting on AI Risk—Controls, Treatment, and Continuous Monitoring
Measurement without action leaves risks on paper. This module explores the Manage function, showing how to translate findings into risk treatments, control implementations, incident response, and ongoing monitoring for AI and GenAI systems.
From Measure to Manage: What This Function Does
Where Manage Fits
In the NIST AI RMF, the Manage function turns evaluation results into action. After you Map systems and Measure risks, Manage is where you decide what to do and actually implement changes.
What Manage Covers
Manage includes: deciding how to treat risks, implementing or adjusting controls, preparing for and responding to incidents, and continuously monitoring AI and GenAI systems over time.
Current Practice (2023–2026)
Today, organizations align Manage activities with the NIST AI RMF, ISO/IEC 42001:2023, and regulations like the EU AI Act (adopted 2024, applied in phases from 2024–2026).
Your Learning Goals
You will learn to turn findings into risk treatment decisions, choose controls, design change management and incident response, and set up continuous monitoring loops for AI and GenAI.
Step 1: Summarize and Prioritize AI Risks
Start With a Risk Register
Turn your Measure results into a risk register: list each risk with description, likelihood, impact, affected stakeholders, and evidence from tests or metrics.
Use a Simple Risk Matrix
Use a 3x3 matrix (Low/Medium/High for likelihood and impact). Place each risk in the grid. High–high cells are your top priorities for action.
Account for Law and Ethics
Give extra priority to risks involving safety, fundamental rights, or regulatory non-compliance, even if their probability is uncertain or moderate.
Make It Practical
Use a spreadsheet or tool so you can sort and filter risks by priority, system, or stakeholder group. This keeps the Manage work organized and transparent.
Step 2: Choose a Risk Treatment Strategy
Four Treatment Options
For each AI risk, choose: mitigate (reduce it), avoid (stop the activity), transfer/share (contract, insurance), or accept (live with it).
When to Mitigate vs Avoid
Mitigate when the use case is valuable and risks can be reduced. Avoid when risks are severe, rights-related, or legally prohibited and cannot be reduced enough.
Transfer With Care
Transfer or share risk by using vendors or insurance, but remember: ethical and legal responsibility toward affected people cannot be fully outsourced.
Make It Explicit
For each risk, record the chosen treatment, the reason, who owns it, and a timeline. This is your bridge from analysis to real-world action.
Activity: Decide on a Treatment Strategy
Work through this thought exercise to practice choosing treatment strategies.
You are part of a university team deploying a GenAI teaching assistant that answers students' questions about course material.
From the Measure phase, you have found:
- Hallucination risk: 8% of answers contain incorrect or fabricated citations.
- Bias risk: For questions about careers, the assistant more often suggests STEM roles to male-sounding names than female-sounding names.
- Privacy risk: Some students paste personal health information into the chat.
For each risk, decide which treatment option makes most sense and why:
- Mitigate
- Avoid
- Transfer/share
- Accept
Your task
- Write down (mentally or on paper) one treatment option for each of the three risks.
- Add one concrete action you would take for each.
Sample reasoning (do this yourself first)
After you decide, compare with this reasoning:
- Hallucination risk: Mitigate by adding retrieval-augmented generation (RAG) from course materials and warning labels, plus automatic citation checks on common questions.
- Bias risk: Mitigate by re-prompting, adding fairness constraints, and periodically auditing career advice outputs across demographic proxies.
- Privacy risk: Mitigate and partially avoid by adding input warnings, client-side redaction of sensitive data, and policies that forbid using the system for medical advice.
Think: Are any of these so severe that you would avoid the feature entirely (for example, disable free-text inputs or career advice)?
Step 3: Implement Controls for AI and GenAI
What Are Controls?
Controls are concrete measures to reduce risk. For AI and GenAI you need layers of controls: on data and models, on applications, in governance, and for end users.
Data and Model Controls
Use data quality checks, privacy techniques, robustness testing, and clear model documentation (like model cards) to control risks at the model layer.
Application and Interaction Controls
Add input limits and filters, output safety checks, and human review for high-impact decisions to keep harmful behavior from reaching users.
Governance and User Controls
Define who owns the system, set approval gates, manage vendor APIs, and provide transparency, warnings, and feedback channels to users.
Worked Example: Controls for a GenAI Customer Support Bot
The Scenario
A company deploys a GenAI support chatbot. Main risks: misleading financial advice, personal data exposure, and toxic replies to customers.
Data and Model Controls
Use a safety-tuned base model, fine-tune on cleaned support transcripts, and rely on a vetted billing knowledge base instead of the model’s general memory.
App-Level Guardrails
Block sensitive inputs like card numbers, filter outputs for toxicity and PII, and restrict the bot to known support intents with safe fallbacks.
Governance and UX Controls
Assign clear owners, require legal sign-off, log and review chats, show an AI notice, and let users switch to a human at any time.
Step 4: Change Management for AI Systems
Why Change Management?
AI and GenAI systems change often. Without structured change management, new versions or prompts can reintroduce safety, bias, or security risks.
Classify Changes
Classify changes as minor, moderate, or major. Major changes (new model, new use case) should trigger fresh risk assessment and testing.
Approve, Version, Roll Back
Set approval workflows, keep clear versioning, and ensure you can roll back to a previous model or configuration if a change goes wrong.
Document Everything
Record what changed, why, who approved it, and what tests were run. For high-risk AI, up-to-date documentation is often a legal requirement.
Step 5: Incident Response for AI and GenAI
What Is an AI Incident?
AI incidents include harmful outputs, unfair decisions, privacy leaks, or security issues where an AI system played a role. Regulators now expect AI-specific response plans.
Detect and Triage
Provide easy ways for users and staff to report issues, monitor for spikes in complaints, and classify severity to decide whether to pause or limit the system.
Contain and Fix
Contain incidents by disabling risky features, tightening filters, or rolling back changes. Notify affected users when necessary and comply with legal reporting duties.
Learn and Improve
Investigate root causes, update controls and documentation, and feed lessons back into your Map and Measure activities so the same incident is less likely to recur.
Step 6: Continuous Monitoring and Feedback Loops
Why Continuous Monitoring?
AI systems and their environments change. Continuous monitoring tracks performance, safety, fairness, user complaints, and compliance over time.
What to Track
Monitor technical metrics, harmful output rates, fairness metrics, user reports, and whether governance rules (like approvals and logging) are being followed.
Thresholds and Reviews
Define thresholds that trigger review, use dashboards, and schedule regular check-ins. High-risk systems need more frequent monitoring.
Make It a Loop
When monitoring shows problems, re-run evaluations, update risk registers, and adjust controls. This keeps the Map–Measure–Manage cycle active.
Quick Check: Managing AI Risk
Answer this question to check your understanding of the Manage function.
Your team discovers that a GenAI writing assistant occasionally reveals fragments of other users' data in its outputs. What is the MOST appropriate immediate Manage action?
- Accept the risk because it only happens occasionally and users benefit from the tool
- Temporarily restrict or pause the assistant while you investigate and strengthen privacy and safety controls
- Transfer the risk by updating the terms of service to say users are responsible for any data leaks
- Only add a warning label telling users not to paste sensitive information
Show Answer
Answer: B) Temporarily restrict or pause the assistant while you investigate and strengthen privacy and safety controls
Revealing other users' data is a serious privacy incident. The appropriate Manage response is to contain the issue by restricting or pausing the system, investigate root causes, and implement stronger privacy and safety controls. Simply accepting, shifting responsibility via terms, or only adding a warning is not adequate for this level of harm.
Key Terms Review
Flip through these cards to reinforce core Manage concepts.
- Risk treatment options
- The four standard choices for handling a risk: mitigate (reduce), avoid (stop the activity), transfer/share (shift some risk to others), or accept (tolerate it with justification).
- Control (in AI risk management)
- A specific technical, procedural, or organizational measure put in place to reduce the likelihood or impact of an AI-related risk (for example, content filters, human review, access control).
- Change management (for AI)
- A structured process for planning, approving, documenting, and monitoring changes to AI systems (models, prompts, data, features) so that risks stay within acceptable bounds.
- AI incident
- An event where an AI system causes or significantly contributes to harm, rights violations, security breaches, or major malfunctions, requiring investigation and response.
- Continuous monitoring
- Ongoing collection and review of metrics, logs, and feedback about an AI system's behavior, risks, and performance, used to trigger re-evaluation and adjustments over time.
Key Terms
- EU AI Act
- A comprehensive EU regulation on artificial intelligence adopted in 2024, with phased application from 2024–2026, that classifies AI systems by risk and sets obligations, especially for high-risk and certain GenAI systems.
- Model drift
- A change over time in the statistical properties of the model's inputs or outputs, which can degrade performance and increase risk if not detected and addressed.
- Risk appetite
- The amount and type of risk that an organization is willing to pursue or tolerate in order to achieve its objectives.
- Human-in-the-loop
- A design pattern where human oversight is built into AI workflows, so that people can review, override, or approve AI outputs before high-impact decisions are finalized.
- High-risk AI system
- Under regulations like the EU AI Act, an AI system used in sensitive domains (for example, hiring, credit scoring, critical infrastructure) that faces stricter requirements for risk management, documentation, and monitoring.
- Red-teaming (for AI)
- Structured adversarial testing of AI systems to probe for vulnerabilities, such as prompt injection, jailbreaks, or ways to elicit harmful or policy-violating outputs.
- AI RMF (NIST AI Risk Management Framework)
- A voluntary framework published by NIST in 2023 that helps organizations manage risks of AI systems through four core functions: Map, Measure, Manage, and Govern.