Course Home ▸ Module 2

Responsible Workflow & Data Foundations

Transform your Week 1 insights into governed workflows and synthetic datasets. Redesign real processes with human-in-the-loop checkpoints, ensuring every AI touchpoint is auditable, compliant, and ready for prototyping.

Workflow Setup Guide Jump to Lab 02
Time Commitment 7–9 hours this week

2h concept study, 4h workflow + data lab, 1h team review, 1h optional tooling practice.

Primary Tools Miro/Excalidraw · draw.io · Jupyter · Codex CLI

Optional: n8n, Airbyte, or Prefect for pipeline visualisation.

Deliverables Future-state workflow · Synthetic data pack · Governance updates

All traceable to Week 1 hypotheses.

Setup & Inputs

1. Consolidate Week 1 Artefacts
  • Discovery backlog with prioritised opportunities.
  • Opportunity canvas PDF.
  • Governance checklist with risk owners.
  • Any stakeholder feedback gathered since Week 1.

Identify the top hypothesis you will shepherd through workflow redesign. Highlight dependencies and assumptions to validate this week.

2. Toolkit Preparation
  • Diagramming: ensure access to Miro, Lucidchart, or open-source draw.io.
  • Data notebooks: set up a Jupyter environment (python -m venv venv && source venv/bin/activate, then pip install pandas faker sdv openpyxl).
  • Install mdit-py-plugins if you want to auto-render Markdown diagrams.
  • Security: confirm you can store synthetic data locally without breaching policy.
3. Reference Material (80 minutes)
  • Microsoft Responsible AI Standard (Aether) – governance principles.
  • NIST AI Risk Management Framework – functions Map/Measure/Manage/Govern.
  • "Responsible Synthetic Data" (Gartner, 2024) – fidelity vs privacy trade-offs.

Capture policy requirements that apply to your industry (GDPR, HIPAA, PCI DSS, etc.).

Current-State Mapping Manual steps, controls Future-State AI Workflow Human + AI swimlanes Controls & Guardrails Synthetic Data Factory Datasets + coverage report
Module 2 output: future-state workflow, governed synthetic datasets, updated risk controls.

Learning Outcomes

Translate hypotheses into governed workflows and data assets.

  • Map current and future-state workflows with explicit human/AI responsibilities.
  • Generate synthetic datasets that reflect real scenarios while protecting sensitive information.
  • Document data lineage, quality checks, and compliance impacts.
  • Update governance frameworks to reflect workflow and data changes.

Concept Briefings

Responsible AI Workflow Design

Embed controls into the workflow instead of bolting them on later. Define trigger points where human oversight is mandatory.

Reference
  • Microsoft Responsible AI Standard v2.
  • NIST AI Risk Management Framework (Map/Measure/Manage/Govern).
  • WEF AI Governance Alliance: "Human-Centered AI".

Synthetic Data Strategies

Use synthetic data to stress-test edge cases and accelerate prototyping without exposing sensitive records. Balance fidelity with privacy by documenting assumptions.

jupyter cell from sdv.single_table import CTGANSynthesizer from sdv.metadata import SingleTableMetadata metadata = SingleTableMetadata() metadata.detect_from_dataframe(df_real) synthesizer = CTGANSynthesizer(metadata) synthesizer.fit(df_real) df_synth = synthesizer.sample(num_rows=2000)

Data Lineage & Compliance

Trace data flow from ingestion to AI output. Identify regulatory touchpoints (GDPR, HIPAA, SOC2) and embed checkpoints for approvals.

gemini prompt lineage "Generate a tabular data lineage for an AI-enabled onboarding workflow covering source systems, transformations, storage, model input, model output, and audit trail."

Guided Exercise Timeline

Step 1 · Current-State Deep Dive (90 min)

Document existing workflow

Use swimlane diagrams to capture roles, systems, decision points, and controls. Identify pain points validated by Week 1 personas.

  • Tool: Miro or draw.io swimlane template.
  • Output: artifacts/week2/current_state.png with legend.
  • Include cycle times and failure modes for each step.
Step 2 · Future-State Design (150 min)

Overlay AI & human responsibilities

Create a dual-lane diagram showing AI services, human oversight, data input/output, and control checkpoints. Define clear RACI assignments.

  • Tool: use layered diagrams; highlight new decision points in contrasting colour.
  • Output: artifacts/week2/future_state.svg.
  • Document guardrails (confidence thresholds, manual override triggers).
Invite a teammate to play "risk officer" and review the future-state map for missing controls.
Step 3 · Synthetic Data Factory (120 min)

Generate datasets and coverage reports

Replicate data structures required by your workflow. Include common and edge cases, ensuring bias and privacy considerations are recorded.

  • Notebook: notebooks/synthetic_data.ipynb.
  • Outputs: data/synthetic/training_dataset.parquet, data/synthetic/coverage_report.json.
  • Document assumptions and fidelity limitations.
Step 4 · Governance & Lineage Update (60 min)

Extend risk register and lineage

Update your Week 1 governance checklist with workflow and data insights. Capture lineage diagrams and regulatory mapping.

  • Artefacts: artifacts/week2/data_lineage.md, updated governance file.
  • Highlight approvals required for production deployment.
  • Tag backlog items with new prerequisites.

Lab 02 · Governed Workflow Package

Produce the process and data assets needed to progress into prototyping.

Inputs
  • Top-ranked opportunity from Week 1.
  • Existing process documentation and metrics.
  • Regulatory requirements and policy documents.
Outputs
  • Current and future-state workflow diagrams.
  • Synthetic dataset package with coverage report.
  • Updated governance and data lineage artefacts.
Collaboration
  • 30-minute review with risk/compliance peer.
  • Async feedback thread in Slack #week2-build.
  • Optional: n8n or Prefect flow share for automation preview.

Execution Steps

  1. Create lateral swimlanes for current and future states; export to PNG/SVG.
  2. Catalogue data required per step. Build a data dictionary capturing field name, type, source, sensitivity, and owner.
  3. Generate synthetic data using SDV/CTGAN; run python scripts/coverage_report.py to summarise distribution coverage and bias checks.
  4. Document data lineage using Markdown tables or tools like yEd. Include ingestion, storage, transformation, model input, model output, and audit logging.
  5. Update governance checklist with new risks (model drift, data leakage) and mitigations.

Validation Checkpoints

  • Future-state workflow shows explicit AI components, human approvals, and fallbacks.
  • Synthetic dataset coverage report includes bias metrics (e.g., demographic parity) and notes on synthetic realism.
  • Data lineage identifies systems of record, integration points, and security classifications.
  • Governance checklist references specific policies and named owners.

Reflection & Submission

Submission Checklist

  • artifacts/week2/current_state.png
  • artifacts/week2/future_state.svg
  • data/synthetic/training_dataset.parquet
  • data/synthetic/coverage_report.json
  • artifacts/week2/data_lineage.md (with diagram link if external).
  • Updated governance checklist (artifacts/week1/governance_checklist.md with Week 2 section).
  • Reflection (artifacts/week2/reflection.md) – highlight biggest workflow/design insight.

Assessment Rubric

Workflow Clarity (30%): Visual accuracy, role delineation, control mapping.
Data Quality (30%): Synthetic fidelity, coverage reporting, documented assumptions.
Governance Integration (25%): Risk updates, lineage completeness, compliance references.
Collaboration Evidence (15%): Peer review notes, risk officer feedback, iteration log.

Submission Process

Commit all artefacts to your Week 2 branch (git checkout -b week2), push to private remote, and open a pull request summarising changes. Upload diagrams and datasets to the portal. Post a 2-minute Loom or written summary in Slack to share your new workflow with the cohort.

Troubleshooting & FAQ

My future-state workflow feels unrealistic.

Cross-check with personas and opportunity scores. If AI is overused, scale back to augmentation (e.g., AI drafts, human finalises). Validate assumptions with a domain expert.

Synthetic data doesn't match real distributions.

Tune the synthesizer: adjust epochs, enforce constraints (min/max, categorical distributions). Consider hybrid synthetic data (seed with anonymised aggregates). Document deviations.

How do I evidence compliance?

Link each workflow step to policy references (e.g., GDPR Article 25). Add control IDs to your governance checklist. Capture consent and audit mechanisms within the lineage doc.

Further Study & Next Steps

Recommended Resources

  • Google Cloud: Responsible AI with Vertex synthetic data lab.
  • Accenture: "AI Governance Playbook" (2024).
  • Open-source toolkits: SDV, Gretel, YData synthetic data libraries.

Prepare for Module 3

Export your synthetic dataset into a vector-friendly format (CSV/Parquet) and ensure you can connect to a notebook environment for RAG experiments. Collect sample knowledge sources for the use case (policies, knowledge base articles, transcripts).

Preview Module 3 →