IAPP AIGP Course · Module 6 · Summary Sheets p.430–431

Governing AI Development — The Full Lifecycle

The complete Module 6 curriculum: Plan → Design → Data Collection & Prep → Model Development → Model Test & Evaluate → Monitor & Maintain → Decommission. Every topic from the course sheets, with plain-words examples so it sticks.

7 curriculum blocks L1 · L2 · L3 lessons 3 LoD oversight 5 V's of data prep 6 testing types
PLAN DESIGN DATA COLLECTION & PREP MODEL DEV MODEL TEST & EVALUATE DEPLOY MONITOR & MAINTAIN DECOMMISSION
Curriculum map: Lesson 1 (L1) = governing the planning & designing of AI models. Lesson 2 (L2) = governing data collection & use in AI design and development. Lesson 3 (L3) = governing the development of the AI model. This module follows the sheets exactly — nothing added, nothing dropped.
Scroll to begin ↓
1
Curriculum block 1 · Lesson 1 (L1)

Plan — Define, Scope & Govern

The planning stage decides whether AI is even the right answer, what problem it solves, and who is accountable. Governance failures here are baked in before a single line of code is written.

1.1 Define the business problem (L1)

In plain words: a bank wants to cut loan losses. The business problem is "predict which applicants will default" — that's a classification problem. The alternative question: could a simpler credit-score threshold do the job with less risk? If yes, AI may not be warranted.

1.2 Identify the use case (L1)

Identify the use case, mission, or gaps — what specific decision or task will the AI system perform, and what gap in the current process does it fill?

In plain words: "reduce customer churn" is a mission; the use case is "predict which subscribers will cancel within 30 days so retention staff can call them first."

1.3 Determine scope (L1)

In plain words: a hospital's AI triage tool affects every patient (high impact), needs 18 months of clinical validation (high effort), and must satisfy medical-device regulation (compliance). Scope = the tool is justified only if impact × fit outweighs effort × compliance cost.

1.4 Evaluate data availability (L1)

In plain words: the churn model needs 12 months of subscription history. The team finds only 6 months exist, and billing data is full of gaps — so the plan must either source more data or narrow the use case. Data availability gates the project.

1.5 Governance existence structure (L1)

Exam anchor — the five operational-control questionsThe sheet lists exactly five: operational responsibility, audits & reviews, feedback & appeals, issue elevation, and the kill switch owner. The exam tests that each control has a named person — not a committee, not "the team".[Module 6 sheet, p.430]

1.6 Human oversight of AI development — the 3 Lines of Defence

LineWhoRolePlain-words example
1st lineManagement & product ownersOwn the system day-to-day: design, build, operate, and first-line risk controlsThe product owner of the fraud model owns its daily performance and escalates anomalies
2nd lineRisk & complianceOversee risk frameworks, set policies, challenge the first line, monitor complianceCompliance reviews the model's PDPA basis and risk signs off the go-live
3rd lineInternal auditIndependent assurance — audits the whole governance systemInternal audit tests whether the model's sign-off records actually exist and were followed
In plain words — the 3 LoD: the players (1st line) run the game, the referees (2nd line) set and enforce the rules, and the match reviewers (3rd line) independently check that the referees and players both did their jobs.
2
Curriculum block 2 · Lesson 1 (L1)

Design — Assess Risk Before You Build

Design is where risk assessment happens before the build: algorithmic impact assessment, risk strategies, and governance across the whole data lifecycle.

2.1 Perform an algorithmic impact assessment (L1)

Algorithmic Impact Assessment (AIA) — a structured assessment of the potential impacts of an algorithmic system on individuals and groups, before deployment. The sheet flags the comparison: AIA vs PIA / DPIA.

AIA — Algorithmic Impact AssessmentPIA / DPIA — Privacy / Data Protection Impact Assessment
FocusImpacts of the algorithm: fairness, bias, rights, societal effectsImpacts on personal data: privacy risks, lawful basis, data flows
OriginCanada's Directive on Automated Decision-Making; emerging AI regimesGDPR Art. 35; PDPC PIA practice
TriggerAutomated decision systems with significant impactsProcessing likely to result in high risk to individuals
OutputImpact level, safeguards, mitigation planRisk register, mitigations, sign-off
Exam tell"Algorithm", "automated decision", "fairness""Personal data", "privacy", "Art. 35"
In plain words: the DPIA asks "what could go wrong with people's data?" The AIA asks "what could go wrong with people's lives?" A hiring model needs both — one for the résumé data, one for the career impact of the decisions.

2.2 Risk assessment strategies (L1)

Use case evaluation

Assess the specific use case's risk profile before building.

Example: a hiring tool is higher-risk than a product-recommendation tool — different scrutiny from day one.

Stakeholder mapping

Identify everyone affected, including non-users who are still impacted.

Example: a credit model affects applicants, guarantors, and even people whose data appears in training records.

Probability / severity harms matrix

Plot each harm by likelihood and impact to prioritise controls.

Example: "wrong loan rejection" = high severity, medium probability → top of the matrix → mandatory human review.

Risk mitigation hierarchy

Prefer eliminating risk over reducing it, and reducing it over transferring it.

Example: eliminate the risk by not using a sensitive attribute; reduce it with oversight; transfer residual risk with insurance.

Benchmarking

Compare against industry baselines and prior systems.

Example: the new fraud model's false-positive rate is benchmarked against the incumbent rules engine before go-live.

Pre-deployment pilots

Test in a controlled real-world setting before full rollout.

Example: the chatbot runs with 5% of customers for two weeks, with escalation metrics armed.

2.3 Governance across the AI data lifecycle

The sheet maps governance to every stage of the data lifecycle — not just training:

Lifecycle stageGovernance focusPlain-words example
Training data governanceLawful basis, quality, bias, lineage, documentationEvery training record can be traced to its source and consent basis
Evaluation and testingHeld-out test sets, subgroup metrics, adversarial testingThe model is scored per demographic group, not just overall
DeploymentSign-off gates, shadow/canary rollout, human oversightNo go-live without the accountable executive's signature
Monitoring and drift detectionPerformance thresholds, drift alerts, retraining triggersRecall below 80% for 3 days auto-flags the model for review
De-commissioningResidual risk, data disposal, archiving, documentationRetired model's training data is deleted per retention policy; the model card is archived
Exam anchor — governance is lifecycle-wideGovernance does not stop at training. The sheet explicitly spans training → evaluation → deployment → monitoring → de-commissioning. A question that treats governance as "just the training data" is wrong.[Module 6 sheet, p.430]
3
Curriculum block 3 · Lesson 1 (L1)

Data Collection & Preparation

Models are frozen data decisions. This block covers gathering the right data, lineage, quality, formats, the 5 V's of data prep, and the privacy toolkit.

3.1 Gather the right data (L1)

In plain words: the churn model needs 12 months of subscription history (what), at least 100k records (how much), collected under the PDPA consent notice (how), stored in the EU region with access logging (where).

3.2 Data lineage and provenance

Data lineage and provenance track the flow of data — its history and source. For any record you can answer: where did it come from, under what basis, transformed how, by whom, when?

In plain words: lineage is the data's biography. When a regulator asks "where did this training record come from?", the lineage log answers in one screen — not after a three-week investigation.

3.3 Data quality considerations

Assess data for training algorithms: accurate, representative, non-biased, and collected with proper statistical sampling.

Quality factorWhat it meansPlain-words example
AccurateData reflects reality; labels are correctA mislabelled "not fraud" transaction teaches the model to miss fraud
RepresentativeData covers the deployment populationA face-recognition model trained only on light skin fails on darker skin
Non-biasedNo systematic skew against groupsHistorical hiring data that under-represents women bakes the skew into the model
Statistical samplingData collected via sound sampling methodsA convenience sample of power users misrepresents the whole customer base

3.4 Data formats

DimensionOptionsPlain-words example
StructureStructured (tables, rows, columns) vs unstructured (text, images, audio, video)Structured: a spreadsheet of transactions. Unstructured: customer emails, support tickets, photos
RecencyStatic (snapshot, doesn't change) vs streaming (continuous updates)Static: a one-time census extract. Streaming: live sensor feeds from factory machines

3.5 Wrangling & preparing data — the 5 V's

Volume

How much data — scale of records and storage.

Example: 50 million transactions vs 5,000 — different pipelines, different governance.

Velocity

Update rate and speed of arrival.

Example: real-time fraud scoring needs streaming data; a quarterly churn model can use batch files.

Variety

Different types and formats of data.

Example: combining structured CRM rows with unstructured call transcripts and images.

Veracity

Accuracy and trustworthiness of the data.

Example: self-reported income is less trustworthy than bank-verified income.

Value

The right data for the outcome — relevance over volume.

Example: 1,000 relevant labelled fraud cases beat 10 million unlabelled transactions.
Exam memory hook — the 5 V'sVolume (how much), Velocity (how fast), Variety (how many kinds), Veracity (how true), Value (how relevant). The sheet's exact definitions: Volume = how much; Velocity = update rate and rate; Variety = different types/formats; Veracity = accuracy/trustworthiness; Value = right data for outcome.[Module 6 sheet, p.430]

3.5b The 5 V's — match the scenario

Five scenarios, five V's. Pick the right one — instant feedback, best score persists.

3.6 Clean, label, anonymise, minimise — and PETs

In plain words: the team cleans the transaction data (removes 4,000 duplicate rows), labels 12,000 fraud cases, pseudonymises customer IDs, drops 30 fields the model doesn't need (minimisation), and trains with differential privacy so no single record can be inferred.
4
Curriculum block 4 · Lesson 3 (L3)

Model Development

The build stage: feature engineering, model training, system architecture, and the train/validate/test discipline. This is Lesson 3 (L3) territory — governing the development of the AI model itself.

4.1 Define model features (L3)

In plain words: for the churn model, the SME says "customers who log in less than twice a week churn" — so the team engineers a "login frequency" feature. The same feature must exist identically in training and test data, or the test results are meaningless.

4.2 Feature engineering — why it matters

Improve model performance

Better features beat bigger models — the right signal in the right shape.

Example: "days since last login" predicts churn better than raw login counts.

Reduce computational costs

Fewer, better features mean cheaper training and inference.

Example: dropping 200 redundant columns cuts training time by 60%.

Boost model explainability

Meaningful features are easier to explain than opaque engineered ones.

Example: "average order value" is explainable to a regulator; "feature_47" is not.
Exam anchor — the feature rulesThe sheet lists four: work with subject matter experts, use the same features for training and testing, avoid unnecessary features, and use feature flags for rollbacks (so a bad feature can be switched off without redeploying the model).[Module 6 sheet, p.430]

4.3 Model training

In plain words: the team trains the churn model, then scores it on 20,000 records it never saw during training. Then they run it in shadow mode on live traffic for two weeks before anyone acts on its predictions.

4.4 Determine system architecture (L3)

SplitPurpose (sheet definition)Plain-words example
TrainingTo train the ML modelThe model learns patterns from 70% of the data
ValidationTo fine-tune parameters and prevent overfitting before final evaluation using the test datasetThe team tunes hyperparameters on 15% of the data, watching for the point where the model starts memorising
TestingTo test performance with new data to ensure the model generalises wellThe final 15% is touched once — if the model scores well here, it generalises
In plain words — challenger models: instead of trusting one model, the team builds three: a logistic regression (interpretable), a gradient-boosted tree (accurate), and a challenger neural net. The champion is chosen on test-set performance and explainability — and the challengers stay ready to swap in if the champion degrades.

4.5 Training the AI model

In plain words: the team fine-tunes a pre-trained language model on 5,000 curated support tickets (fine-tuning) rather than training from scratch — cheaper and faster, but they inherit the base model's biases and must re-evaluate.

4.6 Human oversight of AI development — 3 LoD (revisited)

The sheet repeats the 3 Lines of Defence at the model-development stage — the same structure applies throughout the lifecycle:

Exam noteThe 3 LoD appears in both the Plan block and the Model Dev block of the sheets — expect it in scenario questions about who approves, who challenges, and who audits.[Module 6 sheets, p.430–431]
5
Curriculum block 5 · Lesson 1 (L1)

Model Test & Evaluate

Testing is where claims become evidence. The sheet covers the six testing types, metric and threshold surveillance, and the system-audit checklist.

5.1 Test and validate the AI system (L1) — the six testing types

Brittleness

The model breaks or degrades on small input variations it wasn't trained on.

Example: a vision model misclassifies a stop sign when a sticker is added — the adversarial-example failure.

Uncertainty

The model is overconfident when it should be unsure — confidence scores don't match reality.

Example: a medical triage model says 99% confident on a case that's actually ambiguous.

Hallucinations

Generative models produce confident but fabricated content.

Example: a support chatbot invents a refund policy that doesn't exist.

Embedded bias

Bias baked in from training data or design choices surfaces in outputs.

Example: a résumé screener penalises "women's" club memberships — the Amazon 2018 failure.

Catastrophic forgetting

Fine-tuning on new data destroys previously learned knowledge.

Example: fine-tuning a model on legal text makes it forget how to do maths it previously handled.

False positives

The model flags things that aren't real — costly in fraud, safety, and moderation contexts.

Example: a fraud model blocks 3× more legitimate transactions than the old system.
Exam memory hook — the six testing typesBrittleness, Uncertainty, Hallucinations, Embedded bias, Catastrophic forgetting, False positives — "BUH-ECF". Each one is a failure mode the exam describes in a scenario; you name the type.[Module 6 sheet, p.431]

5.2 Metric and threshold surveillance

MetricWhat it measuresPlain-words example
AccuracyOverall share of correct predictions95% of fraud flags were right — but see precision/recall for the real picture
PrecisionOf items flagged, how many are real80% of flagged transactions were actually fraud — 20% were false alarms
RecallOf real cases, how many were caughtThe model caught 70% of actual fraud — 30% slipped through
F1 scoreHarmonic mean of precision and recallA single number balancing false alarms against missed cases
Bias metricsDisparity across groups (e.g., equalized odds, demographic parity)Approval rates for group A vs group B — a 15-point gap triggers review
In plain words — thresholds: the team sets "recall ≥ 85% and precision ≥ 60%" as the go-live bar, and "recall < 80% for 3 consecutive days" as the retraining trigger. Thresholds are set before deployment, not after the damage.

5.3 System audits

The sheet's audit checklist — six items that make up a full system audit:

Algorithmic impact assessments

Re-run or verify the AIA covers the deployed system's actual impacts.

Example: the audit confirms the hiring model's AIA was updated when the tool expanded to a new role type.

Bias and fairness testing

Independent subgroup analysis on live data.

Example: the auditor runs its own demographic breakdown rather than trusting the vendor's report.

Explainability and interpretability evaluations

Verify the model's decisions can be explained to stakeholders.

Example: can the team explain why a specific loan was rejected, in plain language?

Data governance and quality review

Check lineage, quality, and lawful basis of the data in use.

Example: the audit traces 50 random training records to their sources and consent bases.

Verification of compliance with regulations

Confirm the system meets applicable law (PDPA, GDPR, EU AI Act).

Example: the chatbot's Art. 50 AI disclosure is verified present and correct.

Confirmation of accountability and human oversight

Verify named owners and oversight controls actually operate.

Example: the audit confirms the kill-switch owner is named, trained, and reachable.

5.4 Test and validate — the discipline (L1)

In plain words: a high-risk credit model gets adversarial testing, threat modeling, and a full incident review before go-live; a low-risk product recommender gets a lighter but still documented test pass. Testing effort follows risk.
6
Curriculum block 6 · Lesson 1 (L1) · also Module 5

Monitor & Maintain

Deployment is the start, not the end. This block covers how to monitor, what to document, and how to communicate — to regulators and consumers.

6.1 Manage and monitor AI systems (L1)

In plain words: the bank's AI inventory lists 14 systems, each scored high/medium/low risk. The high-risk credit model gets daily drift monitoring and monthly audits; the low-risk recommender gets weekly checks. Security follows the same industry standards as the rest of the bank.

6.2 Documentation (L1)

Create model cards

Standardised disclosure: intended use, out-of-scope uses, disaggregated performance, caveats.

Example: the credit model's card states it is not validated for self-employed applicants — a documented out-of-scope use.

Create counterfactual explanations (CFE)

Explain a decision by showing what would have changed it: "if your income were $1,000 higher, you would have been approved."

Example: a denied applicant is told the single factor that would flip the decision — the GDPR Art. 22-style explainability tool.

Remediation of adverse impacts

Document how harms are corrected: appeals, manual review, compensation, retraining.

Example: the bank's remediation playbook covers wrong rejections: 24-hour manual review and retroactive correction.

Method of deployment

Record the platform used: cloud, on-site, or hybrid.

Example: the model runs on-premise for data-sovereignty reasons; the deployment method is documented for audit.
Exam anchor — counterfactual explanationsA CFE answers "what would need to change for a different outcome?" — the most practical explainability tool for individual decisions. The exam loves pairing it with model cards (system-level) as the two documentation artefacts.[Module 6 sheet, p.431]

6.3 Create a communication plan (L1)

AudienceFocus (sheet)Plain-words example
RegulatorsConsider compliance and disclosure obligations; ensure transparency and explainability; document data and risk classificationsThe bank files its AIA, model card, and risk classification with the regulator before the credit model goes live
ConsumersFocus on transparency; communicate the functionality of the AI systemThe chatbot tells users they are talking to an AI, what it can do, and how to reach a human
In plain words — two audiences, two messages: regulators get the evidence (compliance, risk classifications, explainability documentation). Consumers get the plain truth (this is AI, here's what it does, here's how to appeal).
7
Curriculum block 7 · Key concerns & governance considerations

De-commissioning AI Systems

The end of the lifecycle is governed too. The sheet lists ten de-commissioning concerns — each one an exam scenario waiting to happen.

Residual risk management

Risks don't vanish when the system is switched off — data, dependencies, and decisions live on.

Example: a retired credit model's past decisions still affect customers; appeals must still be handled.

Data disposal and retention

Delete or retain per policy and law — not all data can be deleted (regulatory retention).

Example: training data is deleted after 90 days, but transaction records are kept 7 years per MAS rules.

Model archiving

Preserve the model, its version, and its documentation for future reference and audits.

Example: the model binary, model card, and test results are archived so a 2028 audit can still inspect the 2026 system.

Documentation obligations

De-commissioning itself must be documented — who decided, when, and why.

Example: the decommissioning record notes the model was retired due to persistent drift and names the approver.

Communications

Tell affected parties the system is being retired — users, consumers, and regulators as needed.

Example: customers are emailed that the AI loan pre-approval service is ending and how to access alternatives.

Knowledge retention

Capture the team's knowledge before it walks out the door.

Example: the model owner writes a handover note covering quirks, failure modes, and lessons learned.

Security risks

Retired systems can be attacked — orphaned APIs, exposed data stores, forgotten credentials.

Example: the retired model's API endpoint is decommissioned, not just ignored — an unpatched endpoint is a breach waiting to happen.

Downstream dependencies

Other systems may depend on the model's outputs — map them before switching off.

Example: the churn model feeds the retention team's dashboard; retiring it without a replacement breaks the dashboard.

Third-party AI

If the system is a vendor product, de-commissioning involves the vendor: data return, deletion, contract exit.

Example: the vendor contract requires data deletion certificates within 30 days of termination.

Governance checklist

Run a formal checklist before, during, and after de-commissioning — the sheet's own governance tool.

Example: the checklist covers: approver named? data disposed? model archived? dependencies mapped? parties notified?
Exam anchor — de-commissioning is not "switch it off"The sheet's ten items are the exam's favourite "what did they forget?" scenario. The classic misses: downstream dependencies (something else still uses the output) and security risks (the orphaned endpoint).[Module 6 sheet, p.431]
8
Self-test · Curriculum-based exam traps

Eight Exam Traps

TRUE / FALSE — every question drawn directly from the Module 6 sheets. Instant feedback, best score persists.

9
Exam-day readiness

Before Thursday — The Checklist

Tap each item as you master it. The goal: fill every box from memory, not recognition.

§
Citation ledger

Sources