About this page
Executive dashboard for clients' AI initiatives
Clients' whole AI agenda at a glance.
- AI initiatives, people reached and investment
- Trustworthy AI adoption in the project cycle
- Coverage of low-income and fragile settings
- Readiness heat map, attention list and upcoming missions
Link to this page
My work
Everything waiting for you, by role.
- Clearance decisions, assessments to do, client evidence to verify
- Overdue agreed actions and missions in the next 30 days
- Client counterparts see their own open actions
Client 360: one view per client government
Everything about one client in one place.
- Readiness radar against peers
- Policy, regulatory and cyber status
- Operations, AI systems and open actions
- Country brief and a preview of what the client sees
Link to this page
Portfolio of client AI initiatives across sectors
What clients do with AI, where, at which stage and for how many people.
- Sector × lifecycle stage matrix
- By type of AI and by funding instrument
- People reached, investment, annual savings
- Use cases replicated across clients
Link to this page
Results and value of client AI initiatives
What AI delivers for people and public finances.
- Reach against target
- Annual savings and service time reduction
- Development outcome per initiative
Link to this page
AI use-case library for replication across clients
Replicate what works.
- Evidence level: proven, promising, emerging
- Typical cost, impact and prerequisites
- Risks and safeguards
- Open source / Digital Public Goods
- Clients already using each use case
Link to this page
Digital and AI foundations by client
The foundations AI needs to work safely and at scale.
- Digital public infrastructure: ID, payments, data exchange
- Hosting and AI compute
- Local-language data and open data
- Biggest gap → foundational investment
Link to this page
Client register for AI governance engagements
The client portfolio.
- Region, income group and fragility
- Readiness score, operations and commitments
- AI systems and overdue actions per client
Link to this page
AI readiness diagnostic for governments
How ready each client is to govern and deploy AI safely.
- Policy and regulation, institutions and oversight, data and infrastructure, skills, cyber resilience
- Heat map across clients
- Weakest pillar → first recommended intervention
Link to this page
AI policy dialogue and regulatory status tracker
Where each client stands on AI governance instruments, and the dialogue behind it.
- AI strategy, AI law, data protection, oversight body
- Cyber strategy, AI in cyber strategy, CSIRT, sandbox
- Dialogue log with outcomes and next steps
Link to this page
Trustworthy AI built into the project cycle
Operations through the project cycle, with trustworthy AI gaps flagged.
- Identification, preparation, appraisal, approval, implementation, completion
- AI screening at concept and impact assessment at appraisal
- Trustworthy AI covenants and ISR ratings
Link to this page
Operational risk ratings including AI and cyber
Risk across operations.
- Nine risk categories and overall rating
- AI and cyber risks in design, capacity, E&S and other
Link to this page
Missions and aide-mémoires
Missions and what was agreed.
- Mission calendar by client and operation
- Key messages and agreed actions
- Printable aide-mémoire
Link to this page
Agreed actions tracked with clients
Follow-through with clients.
- Clients update status and evidence themselves
- Only the team verifies
- Overdue and high-severity first
Link to this page
Register of client AI initiatives
All client AI initiatives in one register.
- Any sector and type of AI: generative, predictive, computer vision, language, agents, forecasting
- Lifecycle from idea to scale, funding and budget
- Value: people reached, savings, faster services
- Trustworthy AI assessment and clearance with separation of duties
Link to this page
Trustworthy AI KPIs: fairness, explainability, robustness, privacy, security
Measured indicators with thresholds.
- Fairness: disparate impact, equal opportunity, FPR balance
- Explainability: coverage and stability
- Security and privacy: accuracy under attack, time to detect, ε
- Operations: data drift (PSI)
Link to this page
AI threat modelling with MITRE ATLAS and OWASP
Threats to each AI system and the controls that address them.
- OWASP Machine Learning and LLM Top 10
- MITRE ATLAS techniques
- Inherent and residual risk
- Attack hints for red-team testing
Link to this page
Procuring AI in the public sector
Safeguards for AI bought from third parties.
- Audit, data, incident, exit and escrow clauses
- Vendor lock-in rating
- Model clauses for AI contracts
Link to this page
Redress for automated decisions
People can understand and challenge automated decisions.
- Grievance redress coverage of AI decisions
- Appeals, overturn rate, time to resolve
- Recourse simulator: explanation, counterfactual, human review
Link to this page
AI incident register for client systems
AI incidents in client systems.
- Biased outcomes, hallucinations, leaks, attacks, misuse
- Reporter, severity and response
Link to this page
Live AI fairness audit: metrics and mitigation
A fairness audit you can run in the browser.
- Historical bias in training labels
- Protected groups: rural, female-headed, disability
- Mitigations compared: unawareness, reweighing, group thresholds
Link to this page
Explainability lab: SHAP, LIME and counterfactuals
Why the model decided this — for auditors and for citizens.
- Exact SHAP for the linear model
- LIME stability and agreement with SHAP
- Plain-language reasons and counterfactuals
- Global importance and proxy detection
Link to this page
Adversarial machine learning lab: evasion and poisoning
How attackers manipulate AI — and what stops them.
- Evasion: accuracy as manipulation grows
- Poisoning: fairness damage that passes accuracy checks
- Defences: registry verification of declared values, loss-based screening of training data
- Red-team suite for generative AI
Link to this page
Privacy lab: membership inference and differential privacy
Privacy risks of AI models and how to measure protection.
- Membership-inference attack (AUC)
- Differential privacy: ε versus accuracy
- Laplace mechanism and privacy budget
Link to this page
AI cybersecurity governance by country
AI in national cybersecurity.
- NIST CSF 2.0 functions per client
- AI risk in national cyber strategy
- AI-specific controls by function
Link to this page
AI and cybersecurity standards crosswalk
One control set, many frameworks.
- NIST CSF 2.0 and ISO/IEC 27001
- NIST AI RMF and ISO/IEC 42001
- OWASP AI Security and Privacy Guide and EU AI Act
Link to this page
AI regulatory options for governments
Advice on how to regulate AI, tailored to the client.
- Soft law, sectoral rules, horizontal law, co-regulation, standards
- Scored against country context
- Recommended path and printable policy note
Link to this page
Knowledge products on trustworthy AI
Knowledge delivered.
- Pipeline from plan to publication
- Reach by product
Link to this page
Capacity building on AI governance and AI security
Capacity built.
- Participants, women's share, satisfaction
- E-learning quiz on trustworthy AI basics
Link to this page
Partnerships and forums on trustworthy AI
Partnerships.
- Forums and standards bodies
- Academia, civil society, private sector
Link to this page
Concept notes and trust-funded programmes
Resource mobilisation.
- Pipeline by status and amount
- Clients covered
Link to this page
Reports
Printable reports generated from live data.
- Portfolio report, country brief, AI impact assessment, policy note, aide-mémoire
Trustworthy AI assessment methodology
How assessments are done.
- Principles, questions, evidence and standards
- Built into the project cycle
- Roles and separation of duties
Link to this page
Users and roles
Roles and client access.
- Client counterparts see their own country only
- Separation of duties enforced by the server
Audit log
Tamper-evident audit trail.
- Every change with old and new values
- SHA-256 hash chain with integrity check
- CSV export
Framework for client engagements on AI
How AI engagements with clients are organised.
- Seven pillars
- Engagement lifecycle: diagnose, identify, prepare, appraise, implement, scale
- Methods and templates per stage
- Anchored in OECD, NIST, ISO, EU AI Act and safeguards
Link to this page
Methodologies for AI in client operations
Step-by-step methods with outputs and standards.
- 14 methods with steps and outputs
- Linked to the tools on this site
- Standards behind each method
Link to this page
Training curriculum on AI for governments
Learning paths by audience.
- Seven audiences
- Modules with hours and formats
- Hands-on with the labs
Link to this page
Templates for working with clients on AI
Templates for every stage of work with a client.
- 15 templates across the engagement lifecycle
- Pre-filled with the selected client
- Open, print or download as Word
Link to this page
Internal library
The team's working library with access levels.
- Guidance, methods, reports, training material, contract clauses
- Versions and owners
- Access levels enforced by the server: internal, client-shareable, public
References and standards for trustworthy AI and AI security
Official sources behind the framework and methods.
- World Bank Group sources
- AI principles and regulation
- Risk management, security standards and open tools
Link to this page
Guiding principles for AI in client operations
The principles behind every engagement.
- Twelve principles consistent with the OECD AI Principles and UNESCO's Recommendation
- Where each principle is applied on this site
Link to this page
Algorithms and KPI catalogue for trustworthy AI
The algorithms and indicators behind assessments.
- 24 algorithms with formulas and where to try them
- 23 KPIs with definitions, targets, frequency and source
- Deepfake KPIs: APCER, BPCER, provenance coverage, time to takedown
Link to this page
AI and the Environmental and Social Framework: screening against the ten standards
How AI risks map onto the ten environmental and social standards.
- Matrix of AI initiatives × ESS1–ESS10
- Suggested ratings from the system profile, editable
- Energy (MWh) and carbon (tCO₂e) of AI workloads
- Feeds the ESCP and SEP templates
Link to this page
Workflows for AI in client operations
Who does what, in which order, by when.
- Clearance with separation of duties
- Agreed actions from mission to verification
- Incident handling within service levels
- Late items per workflow
Link to this page
AI governance frameworks: OECD AI Principles, EU AI Act, NIST AI RMF, ISO/IEC 42001
From international principles to evidence in each operation.
- OECD AI Principles and recommendations to governments
- EU AI Act risk classes and obligations
- NIST AI RMF: GOVERN, MAP, MEASURE, MANAGE
- ISO/IEC 42001 Annex A controls
- Gap report and agreed actions per initiative
Link to this page
ML models and their failure modes: overfitting, data leakage, drift, anomaly detection
How AI systems fail — and how to catch it early.
- Regression, tree ensembles, neural networks, clustering, anomaly detection, LLMs, RAG, fine-tuning, agents
- Overfitting and regularisation
- Data leakage from post-decision features
- Drift detected with PSI and KS
- Isolation Forest for payment fraud and its false alarms
Link to this page
AI operations: AI bill of materials, human oversight, fallback and kill-switch
Running AI safely in production.
- AI bill of materials with lock-in and safety checks
- Human review share and override rate
- Fallback drills and kill-switch runbook
- Availability and time to restore
Link to this page
Schedule and AI roadmap for client operations
When things happen, and in what order.
- Operations timeline (Gantt)
- Coming up in 30, 90 or 180 days
- Readiness index and engagement path
- Phased roadmap with documents for each phase
Link to this page