Artificial Intelligence for Public Sector Training Course

Participants can assess whether available data is suitable, lawful and sufficiently controlled for an AI initiative.

• Data inventory covering sources, owners, sensitivity, retention and permitted use
• Quality dimensions: accuracy, completeness, timeliness, consistency and uniqueness
• Provenance records tracing collection methods, transformations and dataset versions
• Representative sampling checks for protected groups and service-user segments
• Data minimisation applied to model inputs, prompts and training datasets
• Access controls using role-based permissions, logging and separation of duties
• Data protection impact assessment triggers for personal data processing

Participants can prepare a ranked AI use-case portfolio supported by value, feasibility and risk evidence.

• Use-case canvas defining users, decisions, outputs, constraints and success measures
• Impact-effort matrix comparing service value, data readiness and delivery complexity
• Feasibility assessment covering process stability, data volume and technical dependencies
• Risk triage based on affected persons, decision consequences and reversibility
• Baseline measurement before automation using throughput, accuracy and satisfaction metrics
• Human-centred service design requirements for staff workflows and user recourse
• Prioritisation criteria separating pilots, procurement opportunities and prohibited applications

Participants can explain key AI system types, limitations and control requirements to non-technical stakeholders.

• Supervised learning predicts categories, scores, demand and likely outcomes
• Unsupervised learning identifies clusters, anomalies and unusual transaction patterns
• Generative AI produces text, code, images and structured summaries
• Large language model limitations: hallucination, prompt sensitivity and stale knowledge
• Retrieval-augmented generation connects approved sources to generated responses
• Evaluation datasets test accuracy, relevance, safety and factual consistency
• Prompt controls restrict sensitive inputs, unsafe requests and unauthorised disclosures

Participants can establish governance roles, decision gates and evidence requirements for AI systems.

• NIST AI RMF functions: Govern, Map, Measure and Manage
• ISO/IEC 42001 management system roles, policies and documented controls
• Accountability map assigning business owner, model owner and risk owner
• AI inventory records systems, purposes, suppliers, datasets and risk ratings
• Decision gates from discovery through deployment, review and retirement
• Model documentation using model cards, datasheets and decision logs
• Escalation routes for ethical concerns, incidents and material model changes

Participants can design safeguards for fairness, transparency, contestability and meaningful human review.

• Algorithmic impact assessment identifies rights impacts and affected populations
• Fairness measures compare error rates, outcomes and access between groups
• Disparate impact analysis tests whether rules create unequal adverse effects
• Explainability matched to audiences: caseworkers, service users, auditors and leaders
• Human-in-the-loop design defines authority, competence and intervention thresholds
• Appeal processes provide reasons, correction routes and independent review
• OECD AI Principles applied to transparency, accountability and human-centred values

Participants can specify and evaluate AI supplier requirements throughout procurement and contract management.

• Requirements specification defines intended use, exclusions and acceptance criteria
• Supplier due diligence examines training data, model provenance and subcontractors
• Evaluation scorecards weight accuracy, explainability, security and implementation capability
• Contract clauses address data ownership, audit rights and change notification
• Service levels define response times, availability and incident reporting duties
• Evidence requests include test results, bias reports and model documentation
• Exit planning covers data return, model handover and service continuity

Participants can create a monitoring plan that detects declining performance, unfair outcomes and operational failures.

• Pre-deployment testing compares model results against defined acceptance thresholds
• Shadow mode runs predictions without affecting live service decisions
• Performance dashboards track accuracy, precision, recall and false-positive rates
• Data drift detection identifies material changes in production input patterns
• Concept drift monitoring tests whether relationships between inputs and outcomes changed
• Bias monitoring compares outcomes and errors across protected group proxies
• Incident playbooks define containment, investigation, notification and recovery actions

Participants can identify security and resilience controls for public sector AI systems and generative AI services.

• Threat modelling identifies assets, adversaries, attack paths and control priorities
• Prompt injection risks tested against connected tools, documents and workflows
• Data poisoning controls protect training pipelines and feedback collection
• Access logging records prompts, outputs, model actions and administrative changes
• Security testing covers API exposure, secrets management and third-party integrations
• Business continuity plans define manual fallback for unavailable AI services
• Retention schedules govern prompts, outputs, audit records and personal data

Participants can produce a practical 90-day plan for a governed AI initiative.

• Implementation roadmap sequences discovery, assessment, procurement, testing and deployment
• RACI matrix assigns delivery, approval, assurance and escalation responsibilities
• Stakeholder engagement plan identifies staff, unions, residents and oversight bodies
• Change impact assessment maps workflow changes, training needs and role redesign
• Communications plan explains system purpose, limits, safeguards and appeal routes
• Benefit tracking links operational measures to approved investment assumptions
• Ninety-day action plan assigns owners, milestones, evidence and review dates