AI Applications in Government Training Course

Participants can explain how common AI methods work, where they fail and when they are appropriate.

• Supervised learning predicts labels, scores, categories or future numerical values
• Unsupervised learning groups records and detects unusual patterns without labels
• Natural language processing extracts meaning from text, speech and documents
• Generative AI produces text, images, code and structured content probabilistically
• Training, validation and test datasets separate learning from performance assessment
• Model outputs are estimates requiring context, thresholds and accountable human judgment

Participants can map AI applications to service journeys, operational decisions and affected populations.

• Service journey mapping identifies delays, errors, handoffs and information gaps
• Casework triage models prioritise files using transparent operational criteria
• Document intelligence extracts fields from permits, claims, invoices and correspondence
• Citizen assistants route enquiries while preserving escalation to human advisers
• Fraud analytics flags unusual transactions for investigation rather than automatic sanction
• Humanitarian targeting models require contextual validation before beneficiary selection

Participants can assess whether available data is suitable, lawful and sufficiently reliable for an AI use case.

• Data inventory records sources, owners, formats, access restrictions and retention periods
• Provenance analysis traces how records were collected, altered and transferred
• Completeness, accuracy, timeliness and consistency measures reveal data-quality weaknesses
• Representativeness checks identify groups missing from historical administrative records
• Data minimisation limits collection to fields necessary for stated decisions
• Sensitive attributes require controlled access, purpose limitation and documented justification

Participants can prepare a bounded AI problem statement with success measures and decision rules.

• Decision mapping specifies users, decisions, inputs, outputs and escalation routes
• Target variable design translates policy objectives into measurable prediction outcomes
• Baseline measurement compares AI proposals with existing manual or rules-based processes
• Success criteria combine service quality, timeliness, equity, cost and user acceptance
• Exclusion criteria identify decisions unsuitable for automated or AI-assisted treatment
• Assumption logs record dependencies, uncertainties, constraints and unresolved policy questions

Participants can use CRISP-DM to plan an AI project from business understanding through deployment.

• CRISP-DM business understanding defines objectives, constraints and measurable success criteria
• Data understanding documents sources, gaps, distributions, anomalies and access conditions
• Preparation workflows clean, join, label and document training data
• Modelling plans compare interpretable baselines with more complex candidate models
• Evaluation considers technical performance, operational impact and affected-person outcomes
• Deployment planning assigns ownership, controls, support processes and review intervals

Participants can interpret model performance measures and select thresholds appropriate to public decisions.

• Confusion matrices show true positives, false positives, false negatives and true negatives
• Precision and recall expose different harms in screening and prioritisation systems
• Threshold selection balances workload, missed cases, service delays and due-process risks
• Calibration checks whether predicted probabilities match observed outcome frequencies
• Baseline comparison tests whether models improve on rules, averages or staff judgment
• Error analysis examines recurring mistakes by case type, location and population group

Participants can assess potential rights impacts, unequal outcomes and mitigation options before deployment.

• Algorithmic impact assessment identifies affected groups, decisions, harms and safeguards
• Equality analysis compares error rates across protected and underserved populations
• Proxy variables can reproduce exclusion when direct sensitive attributes are absent
• Procedural fairness requires notice, explanation, review and accessible appeal routes
• Participatory assessment incorporates perspectives from service users and community representatives
• Mitigation options include redesign, threshold changes, human review and non-deployment

Participants can define privacy and security controls for AI data flows and model operations.

• Data protection impact assessment documents purpose, necessity, risks and mitigations
• Role-based access control restricts data and model functions by job responsibility
• Encryption protects sensitive records during storage, transfer and approved processing
• Retention schedules govern deletion, archiving and disposal of training datasets
• Prompt injection and data exfiltration threaten generative AI service deployments
• Incident response procedures define containment, notification, investigation and recovery actions

Participants can organise AI governance activities using the NIST AI Risk Management Framework.

• NIST AI RMF Govern function assigns policies, roles, accountability and culture
• Map function documents context, intended use, stakeholders and potential impacts
• Measure function tests performance, validity, reliability, fairness and security
• Manage function prioritises treatment actions, residual risks and decision records
• Risk registers connect technical findings with operational owners and deadlines
• Governance boards review high-impact systems before release and material change

Participants can translate AI management system requirements into practical organisational controls.

• ISO/IEC 42001 context analysis identifies mandates, stakeholders and management-system boundaries
• AI policy defines commitments for responsible development, acquisition and use
• Risk treatment plans assign controls, resources, owners and completion dates
• Documented information preserves decisions, evidence, approvals and system change histories
• Internal audit tests whether stated AI controls operate as designed
• Management review considers incidents, performance trends, resources and improvement actions

Participants can design human oversight arrangements that retain meaningful responsibility for consequential decisions.

• Human-in-the-loop design specifies review points before consequential decisions take effect
• Override protocols record reasons, evidence and authority for departing from recommendations
• Decision rights matrices distinguish model operators, reviewers, managers and accountable officials
• Explanation requirements vary by decision impact, audience and statutory obligations
• Appeal processes allow affected people to challenge AI-informed determinations
• Automation bias controls require independent review of high-risk recommendations

Participants can prepare AI procurement requirements that test supplier capability, transparency and accountability.

• Requirements specifications define intended use, prohibited uses, users and service levels
• Vendor questionnaires request training-data provenance, model documentation and testing evidence
• Evaluation criteria score accuracy, explainability, security, accessibility and support arrangements
• Contract clauses address audit access, incident reporting, change control and exit
• Acceptance testing verifies promised performance against representative operational scenarios
• Open-source components require licence review, maintenance plans and vulnerability management

Participants can assess and control generative AI applications for drafting, knowledge support and service interactions.

• Retrieval-augmented generation grounds responses in approved agency knowledge sources
• Prompt templates standardise tasks, constraints, source citation and escalation instructions
• Hallucination testing checks unsupported claims, fabricated sources and unsafe advice
• Content review workflows approve public-facing outputs before publication or action
• Knowledge-base governance sets ownership, update cycles and source-authority rules
• Sensitive-data controls prevent confidential material entering external model services

Participants can assess predictive models used for forecasting, prioritisation and resource allocation.

• Demand forecasting estimates caseloads, supplies, staffing needs and service volumes
• Risk scoring prioritises review queues without replacing evidence-based final determinations
• Early-warning models combine climate, market, health and mobility indicators
• Geospatial analysis identifies service gaps, exposure patterns and response locations
• Counterfactual caution prevents predictive correlations being treated as causal evidence
• Resource allocation rules document equity constraints alongside efficiency objectives

Participants can design proportionate AI-supported controls for integrity risks and suspicious activity review.

• Anomaly detection identifies unusual claims, payments, vendors and transaction sequences
• Network analysis reveals relationships among suppliers, beneficiaries, staff and intermediaries
• Investigation triage combines model scores with caseworker evidence and risk indicators
• False-positive management protects legitimate claimants from unnecessary delay or scrutiny
• Audit trails retain data, scores, reviewer actions and final disposition decisions
• Integrity controls separate detection, investigation, sanction and appeal responsibilities

Participants can develop an implementation plan that integrates technology, process, workforce and stakeholder changes.

• Operating model design assigns product, policy, data, technology and assurance responsibilities
• Phased rollout starts with controlled pilots, defined safeguards and measurable learning goals
• Change impact assessments identify altered tasks, workloads, skills and decision authority
• Training plans prepare staff to interpret outputs, escalate concerns and document exceptions
• Stakeholder engagement addresses unions, civil society, regulators and affected communities
• Communications plans explain purpose, limitations, safeguards and available redress channels

Participants can establish monitoring and assurance arrangements for deployed AI systems.

• Monitoring dashboards track accuracy, drift, fairness, workload and service outcomes
• Data drift detection identifies changing populations, behaviours, sources and data distributions
• Model change control requires testing, approval and documentation before production updates
• Periodic assurance reviews test governance evidence, controls and operational performance
• Incident classification distinguishes technical faults, harmful outputs, breaches and rights impacts
• Post-incident reviews assign corrective actions, accountable owners and follow-up deadlines

Participants can communicate AI use transparently to leaders, staff, service users and external stakeholders.

• Transparency notices explain purpose, data use, oversight and individual rights
• Plain-language model cards describe intended use, limitations and known performance concerns
• Briefing notes separate evidence, uncertainty, recommendation and decision authority
• Public consultation plans identify affected groups, questions and response commitments
• Media response protocols prepare factual statements for incidents and public concern
• Trust indicators include complaint patterns, appeal outcomes, user feedback and uptake

Participants can build, review and present a complete AI application dossier for a defined public-sector use case.

• Use-case dossier defines problem, public value, users, boundaries and success measures
• Data readiness assessment documents provenance, quality, access restrictions and representativeness
• Risk analysis covers rights, privacy, security, fairness, operational and supplier risks
• Governance design assigns oversight, decision rights, controls, escalation and assurance
• Implementation roadmap sequences pilot, procurement, training, monitoring and stakeholder engagement
• Peer review challenges assumptions, evidence, mitigations and residual-risk decisions