Course chapters
- 01Set the rules for the ReturnGuide pilot
- 02Map ReturnGuide and its data
- 03Map the rules that apply
- 04Assess harm and choose controls
- 05Control data access and copies
- 06Evaluate results before the pilot
- 07Design human review and customer recourse
- 08Review suppliers and plan an exit
- 09Set limits on generative AI and agents
- 10Approve and monitor the pilot
- 11Respond to incidents and verify repairs
- 12Build a governance evidence pack
CHAPTER 01
Set the rules for the ReturnGuide pilot
Give an AI pilot accountable owners, boundaries, required checks, and a way to stop.
Each attempt samples six questions from a bank of eight. The pass mark is 80%, so you need at least five correct answers out of six.
Feedback points you to the relevant chapter sections. You can review the chapter and retake the assessment.
Each chapter’s pass status depends on its assessment result. Course completion requires all twelve chapter assessment passes and instructor approval of your final ReturnGuide project.
Use the project workspace to build one pilot-decision pack through the chapters. Save each checkpoint, then send the final pack for review. The separate practice exercises and self-checks can help you prepare it.
No class attendance or panel is required. The instructor reviews your final project online. If changes are requested, revise your work and send a new version.
Assessments do not inspect practical work or notes, verify legal compliance, or certify professional competence. The completion record does not certify an organisation’s compliance or replace a recognised professional credential.
| Words you need | What it means here |
|---|---|
Governance | Who can make decisions about a system, what evidence they need, and who answers for the result. |
Pilot and synthetic examples | A pilot is a limited trial before wider use. Synthetic examples are made-up questions or records used for testing. |
Scope | The boundaries of a use: which people, data, tasks, and actions are included or excluded. |
Control | A rule, design feature, or working step intended to reduce a risk, such as blocking access to another account. |
Residual risk | The risk left after controls are in place. Judging it requires evidence that those controls work. |
Assurance | Checking whether evidence supports a claim, such as a claim that private records stay private. |
By the end of this chapter, you should be able to:
- Distinguish a governance decision from a policy statement.
- Assign responsibility and approval authority for a bounded pilot.
- Distinguish law, voluntary risk frameworks, and management-system standards.
- Write a usable governance charter with evidence and escalation.
Define the proposed ReturnGuide pilot
The service head asks, “Who can tell me we are ready?” An enthusiastic demo cannot answer that question. Someone must define the approved purpose and the proof needed. Name who can accept the remaining risk. State what happens if the conditions are not met. These arrangements are the beginning of governance.
Name who can approve or stop use
An AI policy might say that systems must be fair, secure, and transparent. Governance turns that direction into decisions. Name who assesses a proposal and who can approve a defined use. State what proof the approver must inspect, what they can refuse, and who can pause use later. A committee needs a decision to make. It also needs the information and permission to make it.
These activities overlap, but each has a different job:
Risk management analyses possible harm and chooses how to respond.
Governance establishes how those choices become authorised decisions and remain accountable.
Delivery puts the controls into practice.
Assurance challenges whether the claims and proof are adequate.
Keep the responsibilities clear. A control owner should not simply approve their own unsupported claim.
Permission to run a staff-only test with synthetic data covers that test. It does not authorise real customer records, sending messages, or expansion to a new market. Record the approval conditions so a later reader can tell exactly what was allowed.
Propose, assess, decide, and operate
State purpose, users, scope, and alternatives.
Collect supporting records and examine harms and constraints.
Approve, restrict, hold, or reject within named decision powers.
Monitor conditions and escalate changes or failures.
The steps repeat when results or use change. They are a course workflow, not a mandatory sequence quoted from a standard.
Name the people responsible for each decision
| Role in the case | Responsibility | Decision limit |
|---|---|---|
Dana, service head | Own the business purpose and resource the service. | Accept only the residual exposure within delegated organisational authority. |
Maya, product owner | Keep scope, inventory, controls, and decision records current. | Cannot waive privacy, security, or other mandatory requirements. |
Imani and Leon, privacy and security leads | Challenge data use and technical protections. | Escalate unmet requirements; specialist advice must reach the approver. |
Noor, evaluation lead; Ellis, operations lead | Produce test results and maintain operating readiness. | Can recommend a hold; Ellis can invoke the agreed containment plan. |
Priya, independent reviewer | Challenge the proof without owning the implementation. | Documents findings and verifies important corrective actions. |
The responsibility table must name who can stop an unsafe use. It must also explain what happens if the sponsor disagrees with a specialist. Give unresolved concerns an escalation route to someone who can make the decision. Record the disagreement and how it was handled. Attending a meeting does not imply approval.
Compare three tools: NIST, ISO, and law
NIST AI RMF 1.0 is a voluntary risk-management framework: a structure for organising this work. Its four functions are Govern, Map, Measure, and Manage. The Playbook offers suggestions to tailor to the context. It is not a checklist that must be completed in full. These references help organise the work; they do not establish that a product complies with every applicable law.
ISO/IEC 42001 concerns an organisational AI management system. These are the arrangements an organisation uses to direct and manage its AI activities. The public overview describes establishing and improving that system. It is different from a product quality score or a course certificate. This course discusses its purpose, not the text of its paid requirements. Laws and contracts create separate obligations. Check their applicability: whether they cover the actual organisation and use.
Write a one-page pilot charter
Purpose: support agents locating current return policy; no automatic commitments.
Scope: named staff, synthetic data, approved policy collection, named environment, and pilot period.
Who decides: approver, specialist reviewers, control owners, escalation contact, and pause operator.
Evidence:
- Inventory.
- Data-flow map.
- Risk record.
- Evaluation.
- Supplier review.
- Containment test.
Gate: unresolved critical issues block the proposed exposure; exceptions need an authorised decision and a record.
These changes require reassessment:
- New data.
- Users.
- Geography.
- Model.
- Permissions.
- Actions.
Record:
- Decision.
- Conditions.
- Dissent.
- Supporting-record versions.
- Expiry or review date.
- Owner.
Plan staff learning for the AI they use
The EU AI Act’s AI literacy duty first applied on 2 February 2025. Article 4 was amended in July 2026. The current text requires providers and deployers to support the development of AI literacy for staff and others using systems on their behalf. It does not require them to guarantee a particular level for every person.
Choose learning that fits the person’s experience and the system’s use. Include the people who could be affected. The Commission says an internal record of training or other guidance can be used; a special certificate is not required.
At Northstar, Maya could show support agents how to compare a ReturnGuide draft with its source. Ellis could practise the agreed pause route with operators. Imani could explain which data the approved pilot excludes.
| Training record field | Northstar example to complete |
|---|---|
Person and role | Name the agent or operator and their task. |
System and permitted use | ReturnGuide, synthetic-question staff pilot. |
Learning and practice | Source comparison, excluded data, and escalation route. |
Date and materials | Record the actual session date and material version. |
Follow-up | Record what the person found difficult and who will help. |
Add this record to the charter’s conditions. A signed attendance sheet alone does not show whether an agent can spot a wrong refund promise.
OPTIONAL PRACTICE
Self-guided practice: write the pilot charter
Optional solo practice: use ReturnGuide or a non-sensitive fictional use case. Write a decision that could be approved or refused, then inspect every proposed action for a named decision owner and clear limit. Compare your draft with the model reasoning and self-check below; keep any revision in your own notes.
One-page charter with the actual approval scope.
Responsibility and escalation table.
Decision record containing at least one condition and one reassessment trigger.
| Self-check | What your notes should show |
|---|---|
Scope | Name:
|
Authority | Identify who approves, who operates each control, and who can pause or escalate. |
Evidence | Connect at least one approval condition to a record or test that could be inspected. |
Change | State one material change that would require a new decision. |
Review the chapter and take the assessment
Can someone tell which pilot actions remain prohibited?
Does each important control and decision have a named owner?
Can a specialist concern reach somebody able to pause or refuse exposure?
Are framework alignment and legal compliance described as different claims?
Save your checkpoint for this chapter.
Sources and further reading
- NIST AI Risk Management Framework 1.0
January 2023 edition, especially the Core and lifecycle context. NIST states that this edition is being revised; check the official resource centre for updates. References checked 16 November 2026.
- NIST AI RMF Playbook: Govern
See GOVERN 1 and 2 for policy, accountability, and documented responsibilities. Voluntary suggested practices, not a universal legal checklist. Checked 16 November 2026.
- ISO/IEC 42001:2023: public scope overview
Public overview only. This course does not reproduce the paid standard, interpret each clause, or confer conformity or certification. Checked 16 November 2026: the older /standard/81230.html address redirects to this canonical page.
- EU AI Act: current Articles 4 and 113
Current consolidated text records the amended duty and application provisions. Checked 16 November 2026.
- European Commission: AI literacy questions and answers
The Commission explains the duty, training records, and that no special certificate is required. Checked 16 November 2026.
YOUR NEXT CHECKPOINT
Ready to check your understanding?
Mark your reading complete, then take the chapter assessment. Reading progress is separate from your assessment result. Save the related course project checkpoint as you work through each chapter.
View course enrollment