A comprehensive taxonomy of 432 AI-specific skill attributes mapped to the SFIA 9 framework.
Comparing different AI response styles, prompts or model configurations in controlled experiments to determine which delivers better user outcomes.
Designing training programmes that use AI to dynamically adjust content, pace and difficulty based on individual learner progress and performance.
Deliberately constructing prompts designed to bypass AI safety controls or elicit unintended behaviours, used to identify vulnerabilities before attackers do.
Systematically stress-testing AI agents by simulating hostile inputs, unexpected scenarios and edge cases to verify they behave safely under pressure.
Defining the boundaries within which an AI agent can act independently, including what decisions require human approval before execution.
Identifying and mitigating risks that arise when AI agents operate with reduced human oversight, including runaway actions and unintended consequences.
Specifying the observable behaviours an AI agent must demonstrate — and must not exhibit — before it is approved for production use.
Formal checkpoints in a release process where AI agent behaviour is validated against defined criteria before progressing to the next stage.
Managing and approving changes to how AI agents behave, ensuring modifications are reviewed, tested and traceable before deployment.
Verifying that updates to an AI agent have not degraded or altered previously working behaviours, similar to software regression testing.
Designing specific test cases that exercise an AI agent's decision-making across normal, boundary and failure conditions.
Tracking and controlling the settings, parameters, tool access and permissions that define how an AI agent operates in each environment.
The process of releasing an AI agent into a live environment and registering it within organisational systems for monitoring and governance.
Defining the moral and ethical limits within which an AI agent must operate, including topics it must refuse and actions it must not take.
Building AI agents using orchestration frameworks such as LangChain, CrewAI or AutoGen that manage tool use, memory and multi-step reasoning.
Architecting technical safeguards that constrain AI agent actions, such as output filters, tool-use restrictions and escalation triggers.
Assigning AI agents their own identity credentials and defining precisely which systems, data and actions they are authorised to access.
Integrating AI agent workflows using graph-based orchestration tools like LangGraph that manage state, branching and multi-agent coordination.
Designing the overall architecture that coordinates multiple AI agents, managing task delegation, inter-agent communication and result aggregation.
Defining explicit limits on what resources, tools and data an AI agent can access, following least-privilege principles.
Evaluating the potential harm an AI agent could cause when operating at the edges of its permitted autonomy and designing mitigations.
Determining which tasks within an organisation or process are best performed by AI agents versus humans, and designing the division of responsibilities.
Establishing processes to quickly revert an AI agent to a previous version or configuration when issues are detected in production.
Designing the internal reasoning patterns AI agents use — such as ReAct loops, chain-of-thought prompting and structured tool calling — to solve problems.
Defining the precise scope of work an AI agent is responsible for, including clear handoff points to humans or other agents.
Setting up which external tools, APIs and data sources an AI agent can invoke, and under what conditions it is permitted to use them.
Testing complete AI agent workflows from trigger to final output, validating that all steps, tool calls and decision points work correctly together.
Designing and packaging an AI agent as a standalone product or service offering, including its capabilities, interfaces, pricing and support model.
Rethinking existing business processes to take advantage of AI agents, replacing manual steps with autonomous agent actions where appropriate.
Providing expert guidance on how to structure systems that incorporate autonomous AI agents, including patterns for reliability, safety and scalability.
Using specialised frameworks like LangGraph, CrewAI or AutoGen to build, test and deploy AI agents that can reason, use tools and collaborate.
Using AI agents to automate project management tasks such as status updates, dependency tracking, risk flagging and stakeholder communications.
Pre-defined architectural templates for common AI agent workflows such as research-and-report, approval chains and multi-step data processing.
Designing the interaction patterns for when an AI agent needs to transfer control to a human, ensuring context is preserved and the transition is seamless.
Selecting and specifying the appropriate hardware accelerators — GPUs, TPUs or NPUs — needed to train or run AI models at the required performance level.
Understanding and applying the EU AI Act and other emerging regulations that govern AI development, deployment and use within organisations.
Using generative AI to draft advertising copy, headlines and calls-to-action, then refining outputs with human review for brand consistency.
Planning and delivering communications that build awareness, understanding and enthusiasm for AI tools being introduced across an organisation.
Measuring how widely and frequently AI tools and capabilities are being used across an organisation to gauge adoption success.
Identifying and addressing the reasons people resist using AI tools — such as fear, distrust or workflow disruption — through targeted interventions.
Advising organisations on the sequencing and prioritisation of AI initiatives, from quick wins to transformational programmes.
Tracking what AI agents are doing in real time — which tools they call, what data they access and what outputs they produce — for operational and security oversight.
Designing the rules and triggers that determine when an AI agent should stop and escalate to a human, and how to transfer context cleanly.
Assessing and comparing AI agent frameworks such as LangChain, CrewAI, AutoGen and Semantic Kernel for suitability against specific use cases.
Reusable templates for specifying what an AI agent must do, covering capabilities, constraints, safety requirements and integration points.
Using AI agents to create interactive, scenario-based training exercises where the agent plays roles such as customer, interviewer or adversary.
Using machine learning to automatically identify unusual patterns, outliers or emerging trends in data that would be difficult for humans to spot.
Implementing secure access controls for AI service APIs, managing API keys, OAuth tokens and role-based access to model endpoints.
Defining and negotiating the contractual terms for AI API usage, including uptime guarantees, rate limits, data handling and liability clauses.
Tracking, rotating and securing the API keys and licenses used to access AI services, preventing unauthorised use and cost overruns.
Managing AI models, datasets, prompts and configurations as organisational assets from creation through active use to retirement.
Introducing AI assistants such as Copilot or Claude into employee workflows with proper training, guidelines and support to drive effective use.
Using machine learning to automatically group customers or users into meaningful segments based on behaviour, preferences and characteristics.
Leveraging AI-powered advertising platforms from Meta, Google and others to automatically identify and target the most relevant audiences for campaigns.
Assessing whether AI systems produce outputs that disadvantage users with disabilities, different languages or diverse backgrounds.
Systematically comparing the costs, risks and benefits of building custom AI solutions in-house versus purchasing or subscribing to third-party AI services.
Using AI to automatically test variations of marketing campaigns and reallocate spend toward the best-performing creative, channels and audiences.
Visually mapping an organisation's current AI capabilities across functions and maturity levels to identify gaps and prioritise investment.
Standardised frameworks that describe the building blocks of AI capability — such as data, models, infrastructure, governance and talent — to guide planning.
Establishing a dedicated team or function to provide AI expertise, standards, reusable assets and guidance across an organisation.
Building networks of trained advocates across business units who promote AI adoption, share best practices and provide peer-to-peer support.
Structured guides that help teams assess and manage the organisational impact of introducing AI into existing roles, processes and ways of working.
Designing customer-facing conversational AI experiences that feel natural, resolve queries effectively and escalate to humans when needed.
Building and managing conversational AI solutions using platforms such as Azure Bot Service, Google Dialogflow or Amazon Lex.
Using AI-powered coding tools such as GitHub Copilot, Claude Code or Cursor that suggest, generate and refactor code within the developer workflow.
Structured definitions of the knowledge, skills and behaviours required for different AI-related roles and proficiency levels within an organisation.
Using AI to review bid documents and proposals against regulatory requirements, client specifications and organisational policies before submission.
Designing how AI components — such as models, embeddings and agents — connect with existing systems, data flows and user interfaces.
Measuring and reporting the energy consumption and carbon emissions associated with training and running AI models.
Forecasting the infrastructure and API costs of AI initiatives, including GPU compute time, token consumption and storage for models and embeddings.
Using AI tools to automatically verify that digital content meets accessibility standards such as WCAG for users with diverse needs.
Tagging AI-generated or AI-assisted content with clear provenance labels so audiences can distinguish it from purely human-created material.
Using AI tools to review, improve and polish written content — checking for clarity, tone, grammar, style consistency and audience fit.
Using generative AI to produce marketing content such as email copy, social media posts, landing page text and ad variations at scale.
Deploying AI systems to automatically detect and flag inappropriate, harmful or off-brand content before or after publication.
Design patterns for how AI copilots present suggestions, ask for clarification and integrate into the user's workflow without disrupting their flow.
Designing structured learning programmes that teach AI skills — from prompt engineering and tool use to model development — matched to organisational needs.
Legal agreements that govern how AI service providers handle, store and process organisational data, covering privacy, retention and cross-border transfers.
Maintaining detailed logs of AI system inputs, reasoning steps and outputs so that decisions can be reviewed, explained and challenged after the fact.
Using AI-powered design tools such as Adobe Firefly, Canva AI or Figma AI to generate, iterate and refine visual designs from prompts and references.
Identifying coordinated or AI-generated misinformation campaigns by detecting patterns in content, timing and distribution networks.
Tracking, comparing and documenting AI experiments — including model configurations, training runs and evaluation results — to support reproducibility.
Using large language models to brainstorm, explore and evaluate potential product features by analysing user feedback, market data and competitive offerings.
Automated pipelines that compute, store and serve the input features required by machine learning models for training and real-time inference.
Designing intermediary services that route, rate-limit, log and secure traffic between applications and AI model endpoints.
Organising and running time-boxed innovation events where teams rapidly build AI prototypes and proofs of concept to explore new use cases.
Quantifying how frequently an AI model generates factually incorrect, fabricated or unsupported statements across different use cases and domains.
Documenting and tracking the specific scenarios where AI hallucinations pose business risk, along with mitigations and monitoring controls.
Creating visual content using text-to-image AI tools such as Midjourney, DALL-E or Stable Diffusion, guided by detailed prompts and style references.
Structured evaluations of how an AI system will affect stakeholders, processes, jobs, fairness and rights — typically conducted before deployment.
Ensuring that AI components used in systems where failure could cause harm — such as medical devices, transport or industrial controls — meet stringent safety standards.
Automatically adjusting the compute resources allocated to AI model serving based on real-time demand, balancing cost against response time.
Using tools like TensorRT, ONNX Runtime or vLLM to make AI models run faster and more efficiently at inference time without sacrificing accuracy.
Evaluating proposed AI projects against criteria such as business value, feasibility, data readiness and risk to determine investment priority.
Building structured repositories of organisational knowledge — documents, FAQs, procedures — optimised for AI retrieval using embeddings and indexing.
Using AI tools such as Salesforce Einstein to automatically score sales leads by likelihood to convert and forecast pipeline revenue.
Organisation-wide education initiatives that build foundational understanding of AI concepts, capabilities, limitations and responsible use across all roles.
Diagnostic tools that allow individuals to evaluate their own AI knowledge and skills, identifying areas for development.
Using AI-powered marketing platforms such as Google AI and Adobe Sensei to analyse campaign performance, attribution and customer behaviour.
Deploying AI-enhanced marketing automation platforms to personalise customer journeys, optimise send times and automate lead nurturing.
Structured models for evaluating how advanced an organisation is in its AI adoption across dimensions such as strategy, data, talent, technology and governance.
Comparing an organisation's AI maturity against industry peers or standards to identify relative strengths and gaps.
The specific performance, safety, fairness and operational requirements an AI model must meet before it is approved for production deployment.
Restricting who and what can invoke AI models, including role-based permissions, network controls and usage quotas.
Connecting applications to AI model endpoints via APIs, handling authentication, request formatting, response parsing and error management.
Evaluating the downstream effects of updating or replacing an AI model, including impacts on dependent systems, outputs and user experience.
Packaging AI models and their dependencies into containers using Docker, Triton or similar tools for consistent, portable deployment.
Continuously tracking whether an AI model's performance is degrading over time as real-world data diverges from the data it was trained on.
Investigating why an AI model's performance has degraded — whether due to data distribution changes, upstream data issues or concept shift.
Understanding and defending against techniques where adversaries attempt to recreate a proprietary AI model by systematically querying its API.
Responding to production failures in AI systems — such as outages, quality degradation or safety incidents — with structured diagnosis and remediation.
Operating dashboards and alerting systems that flag when AI model performance metrics fall outside acceptable thresholds.
Standardised reports that track AI model accuracy, latency, cost, fairness and reliability metrics over time for stakeholder review.
Assessing third-party AI model providers on criteria including model quality, pricing, data handling, compliance, support and roadmap.
Establishing policies and processes for what gets recorded in an AI model registry, who can publish models and how lifecycle stages are managed.
Operating a centralised catalogue of AI models that tracks versions, metadata, ownership, deployment status and dependencies.
Using platforms such as MLflow or Azure Machine Learning to maintain a searchable, versioned registry of all organisational AI models.
Automated workflows that take an AI model from validated artefact through staging, testing and approval to production deployment.
Testing how AI model serving performs under increasing load to identify bottlenecks and determine capacity limits before they affect users.
Deploying and operating the infrastructure that hosts AI models for real-time inference, using serving platforms such as Triton Inference Server or vLLM.
Day-to-day operation of AI model serving platforms, including monitoring, scaling, troubleshooting and applying updates.
Administering the platforms and infrastructure used to serve AI models, including resource allocation, security patching and configuration management.
Structured approaches for verifying that AI models meet requirements for accuracy, fairness, robustness and safety before release.
Tracking changes to AI models, datasets and experiments using version control tools such as DVC or MLflow to ensure reproducibility.
Systematically testing AI models for security weaknesses such as prompt injection susceptibility, data leakage and adversarial input sensitivity.
Using AI to capture, interpret and generate realistic human motion for animation, gaming and simulation applications.
Structured templates for identifying and evaluating potential AI use cases, capturing the problem, data availability, expected value and feasibility.
Proven architectural patterns for coordinating AI components — such as chaining prompts, routing between models and managing agent collaboration — using frameworks like LangChain.
Defining the key performance indicators that measure whether AI initiatives are delivering their intended business outcomes and value.
Tracking the origin of AI-generated outputs — which model, version, prompt and data produced them — for accountability and debugging.
Processes where subject matter experts review and verify AI-generated outputs for accuracy and relevance before they are used or published.
Applying AI to detect patterns, anomalies and connections in large datasets during cybercrime, fraud or compliance investigations.
Systems that use machine learning to tailor content, recommendations and experiences to individual users based on their behaviour and preferences.
Managing and maintaining cloud AI platforms such as Azure AI Studio or AWS SageMaker, including user access, resource allocation and service configuration.
Evaluating and comparing cloud AI platforms and tools — such as Azure AI, AWS SageMaker, Google Vertex AI and Anthropic — against specific requirements.
Defining the evaluation criteria for selecting AI platforms, covering model availability, security, compliance, cost, integration and vendor support.
Using AI-enhanced tools such as Copilot in Power BI to create and interact with dashboards that track portfolio performance using natural language.
Quantifying the time savings, quality improvements and throughput gains achieved by introducing AI tools into specific workflows.
Defined standards for assessing how effectively individuals can work with AI tools, from basic prompt writing to advanced agent development.
Using AI assistants such as Copilot or Claude to help draft project plans, estimate effort, identify risks and generate status reports.
Optimising how databases and search systems handle AI-related queries — including vector similarity searches and embedding lookups — for speed and efficiency.
Structured evaluations of an organisation's preparedness to adopt AI, covering data maturity, infrastructure, skills, governance and culture.
Organised exercises where a team deliberately attempts to find flaws, biases and vulnerabilities in AI systems before they are exploited in the real world.
Specific tools and methodologies used to systematically test AI systems for security vulnerabilities, bias, harmful outputs and failure modes.
Verifying that AI systems meet applicable legal and regulatory requirements such as the EU AI Act, GDPR, sector-specific rules and organisational policies.
Monitoring upcoming AI-related legislation, standards and regulatory guidance across jurisdictions to prepare the organisation for compliance.
Using AI-powered research tools such as Elicit or Consensus to find, summarise and synthesise findings from academic and technical literature.
Assessing the accuracy, relevance, completeness and appropriateness of AI-generated responses through systematic human or automated review.
Anticipating how existing job roles will change as AI capabilities grow, and planning transitions in responsibilities, skills and organisational structures.
Creating structured arguments — supported by evidence — that an AI system is acceptably safe for its intended use in a specific context.
Designing and implementing technical constraints that prevent AI systems from taking unsafe actions, even when facing novel or adversarial inputs.
Using natural language processing to extract sentiment, opinions and emerging trends from customer reviews, social media and survey responses.
Using AI tools to analyse search intent, optimise content for search engines and identify keyword opportunities to improve organic visibility.
Using AI copilots integrated into service management platforms to assist with ticket classification, knowledge retrieval and resolution suggestions.
Managing the service accounts and machine identities that AI systems use to authenticate with other services, ensuring proper access controls.
Designing training experiences that use AI to create realistic simulated scenarios — such as customer interactions or crisis situations — for hands-on practice.
Predicting which AI-related skills an organisation will need in the future based on technology trends, strategy and planned AI initiatives.
Structured career and learning paths that guide individuals from foundational AI awareness through to advanced specialisation in AI disciplines.
Comparing the AI skills an organisation currently has against those it needs, identifying gaps that must be closed through hiring, training or partnering.
Monitoring and recording individuals' AI skill levels over time to measure development progress and identify further training needs.
Using AI to analyse individuals' skills profiles and match them to suitable roles, projects or development opportunities.
Ready-made guides for engaging different stakeholder groups — executives, end-users, regulators — on AI initiatives with appropriate messaging.
Visual planning tools such as BMADS canvases that help teams map out AI strategy components including vision, capabilities, data and governance.
Verifying the security and integrity of the AI supply chain — including pre-trained models, third-party datasets and open-source components.
Using AI models to approximate complex simulations or physical processes, enabling faster exploration of design alternatives and what-if scenarios.
Using AI to automatically analyse survey responses — extracting themes, sentiment and key findings from open-text answers at scale.
Using AI tools to identify, attract and shortlist candidates with AI-related skills from job boards, networks and internal talent pools.
Structured evaluation criteria for determining whether an AI solution is technically viable given available data, infrastructure, skills and constraints.
Using AI to identify security threats in real time — such as anomalous network traffic, suspicious behaviour or malware — and automate initial response actions.
Optimising network configurations to minimise latency for AI workloads, particularly for real-time inference and edge AI deployments.
Ensuring that data used to train AI models complies with privacy regulations such as GDPR and the EU AI Act, including consent, minimisation and rights management.
Publishing clear information about how AI systems work, what data they use, their known limitations and how decisions are made.
Measuring how much users trust AI system outputs and how confident they feel relying on AI-generated recommendations in their work.
AI-powered tools that provide personalised tutoring, coaching and feedback to learners, adapting to their pace, style and knowledge gaps.
Understanding and managing AI service licensing models based on token consumption, API calls or usage tiers rather than traditional seat licences.
Organisational policies that define how AI tools may and may not be used, covering data handling, approved tools, quality checks and accountability.
Facilitated sessions that bring together business and technical stakeholders to identify, evaluate and prioritise potential AI applications.
Structured methods for estimating the return on investment of AI use cases, accounting for development costs, ongoing compute, risk and measurable benefits.
Standardised criteria and scoring models for assessing AI technology vendors on capability, security, compliance, pricing and strategic fit.
Comparing AI vendor performance against agreed metrics such as model quality, uptime, latency, support responsiveness and cost efficiency.
Monitoring and managing expenditure on AI services, APIs and platforms across the organisation to control costs and optimise investment.
Planning how the workforce will evolve as AI takes on more tasks, including reskilling programmes, role redesign and change support.
Managing the scheduling, scaling and resource allocation of AI workloads across GPU-enabled Kubernetes clusters.
Adapting agile sprint planning for AI and ML projects, accounting for the iterative and experimental nature of model development.
Using natural language processing models to automatically categorise text, documents or records into predefined classes.
Using AI to select, organise and schedule content for publication across channels based on audience engagement patterns and relevance.
Using AI tools to process and analyse large volumes of digital evidence, identifying relevant patterns and connections faster than manual review.
Using AI to analyse incident symptoms, logs and historical patterns to suggest probable root causes and accelerate resolution.
Using AI to transcribe, code and analyse user research interviews, identifying themes and patterns across multiple sessions.
Using AI tools such as Otter.ai or Microsoft Copilot to transcribe meetings, capture action items and generate summaries automatically.
Using AI to help tailor development and delivery methodologies to specific project contexts, team sizes and organisational constraints.
Using AI to analyse incident and problem data to detect recurring patterns, predict future issues and suggest preventive actions.
Using AI to draft, structure and refine sales proposals and pitch documents, drawing on organisational templates and past successful bids.
Using AI to automatically classify and tag records based on content, applying retention schedules and compliance categories.
Using AI assistants such as M365 Copilot to draft emails, schedule meetings and manage administrative correspondence.
Tools that use AI to evaluate individuals' competencies through adaptive questioning, scenario-based exercises or portfolio analysis.
Using AI to help estimate story points, identify dependencies, balance team capacity and suggest sprint compositions based on historical velocity.
Using AI to automatically classify, prioritise and route incoming incidents or requests to the most appropriate team or individual.
Using AI to analyse employee performance data, identify trends and generate insights that support management decision-making.
Using AI tools to verify that business processes execute correctly by analysing process data and flagging deviations from expected behaviour.
Designing team compositions that integrate AI agents and copilots as team members, redefining how humans and AI collaborate on shared objectives.
Using machine learning to analyse customer loyalty programme data, predict churn and identify the most effective retention interventions.
Using AI to automate the buying, placement and optimisation of digital advertisements across platforms in real time.
Applying AI techniques to systematically identify the underlying causes of business problems, system failures or process deviations.
Security information and event management platforms enhanced with AI — such as Microsoft Sentinel or Splunk AI — that detect threats and automate response.
Using generative AI to create animation sequences, character assets, textures and visual effects for games, film and interactive media.
AI tools that analyse datasets and automatically suggest the most appropriate chart types and visual encodings to communicate the data effectively.
Informing audiences when content has been created or substantially modified by AI, in line with organisational policy and emerging regulations.
Understanding and implementing the legal and regulatory requirements for disclosing AI involvement in content creation.
Detecting whether a piece of content — text, image, audio or video — was generated by AI, using technical analysis or watermark detection.
Identifying AI-generated synthetic media — such as manipulated video, audio or images of real people — using forensic analysis and detection tools.
Using AI to identify fraudulent transactions, claims or activities by detecting patterns and anomalies that differ from legitimate behaviour.
Using AI to automatically produce written explanations and commentary about data findings, turning charts and metrics into business narratives.
Using large language models such as Claude or ChatGPT to draft learning materials, quizzes, case studies and training exercises.
Using AI to create first drafts of proposals, one-pagers, brochures and sales collateral that can be refined by subject matter experts.
Using AI to automatically compile project or service status reports from data in tracking tools, reducing manual report writing effort.
Using AI coding tools such as GitHub Copilot or Claude Code to automatically generate unit tests, integration tests and test data.
Design approaches purpose-built for products where AI is the core capability — not an add-on — covering UX, feedback loops and graceful degradation.
Using AI to automate IT operations tasks such as incident detection, root cause analysis, remediation and capacity optimisation.
Deploying observability and operations platforms — such as Dynatrace or Datadog AI — that use machine learning to detect and resolve infrastructure issues.
Designing database queries that are optimised for AI workloads, including vector similarity searches, approximate nearest neighbour lookups and hybrid search.
Delivering training that automatically adapts content, pacing and assessments in real time based on the learner's demonstrated understanding.
Learning management systems enhanced with AI capabilities for personalised recommendations, automated content tagging and learner analytics.
Using AI to model and simulate business processes, enabling teams to test changes and predict outcomes before implementing them.
Classifying organisational data assets by their suitability for AI use, considering quality, completeness, sensitivity and licensing constraints.
Tailored frameworks for managing the unique risks of AI projects — such as data dependency, model uncertainty and ethical considerations — within scope management.
Independent assessments of how algorithmic systems affect different populations, checking for unintended discrimination, harm or unfairness.
Integrating AI-powered anomaly detection into business intelligence dashboards to automatically highlight unexpected deviations in metrics.
Using AI to automatically discover and generate exploits for security vulnerabilities, accelerating penetration testing and red team exercises.
Using AI to automatically summarise risks, assumptions, issues and dependencies from project tracking tools into concise status updates.
Cloud platforms such as Google Vertex AI and AWS SageMaker Autopilot that automate model selection, hyperparameter tuning and training for non-specialist users.
Reviewing scenarios where AI systems make decisions without human involvement to ensure safety, fairness and alignment with organisational values.
Systematically testing AI models for biased outputs across demographic groups, protected characteristics and different input populations.
Using the BMADS (Business Model, Architecture, Design, Strategy) framework with AI agents to collaboratively generate and refine enterprise architecture artefacts.
An AI agent configured with the BMADS methodology that assists architects in developing business and technology architecture documentation.
Deployment strategies that expose new AI models to a small percentage of traffic (canary) or run them in parallel with the current model (shadow) to validate performance.
Conducting usability tests specifically designed to evaluate how effectively users can interact with AI chatbots and agents to complete tasks.
Managing AI training and inference workloads across cloud platforms, handling resource provisioning, job scheduling and cost optimisation.
Mapping out the end-to-end customer journey through conversational AI interactions, identifying friction points and opportunities for improvement.
Operating and maintaining conversational AI systems including chatbots, voice assistants and virtual agents across their lifecycle.
Designing natural and effective conversational interfaces — including dialogue flows, personality, error handling and multi-turn interactions.
AI-powered assistants that allow users to ask questions in natural language and receive answers drawn from organisational knowledge bases.
Analysing transcripts and metrics from AI conversational interfaces to understand user behaviour, identify failures and improve interaction quality.
Embedding AI copilots and assistants into products so that users can interact with AI capabilities natively within the product experience.
Microsoft's AI assistant for security professionals that helps investigate threats, analyse incidents and take remediation actions using natural language.
Embedding AI copilots — such as M365 Copilot or GitHub Copilot — into existing business and development workflows to augment human productivity.
Using AI copilots to help service desk and support staff diagnose issues faster by suggesting solutions based on historical incident data.
Using AI copilots to equip sales teams with real-time insights, talking points, competitive intelligence and personalised content during engagements.
Using AI assistants to guide support staff through diagnostic steps, suggest likely causes and recommend fixes based on symptoms and system data.
Redesigning job roles to incorporate AI copilot capabilities, redefining which tasks are done by the human, which by the copilot, and which collaboratively.
Using AI copilots in tools like Power BI to explore data through natural language questions, generating charts and insights without writing queries.
Integrating AI copilots into daily employee workflows — such as email, document creation and data analysis — to reduce manual effort and improve quality.
Using Anthropic's Cowork tool to automate file management, data processing and repetitive desktop tasks through AI-driven workflows.
Coordinating AI governance practices — policies, standards and oversight — across different business units to ensure consistency and reduce duplication.
AI assistants that support customer service agents in real time by suggesting responses, retrieving knowledge and summarising customer history.
Managing the processes and tools used to label, annotate and validate training data for supervised machine learning models.
Implementing controls to prevent AI systems from exposing sensitive or confidential data in their outputs, training data or API interactions.
Using workflow orchestration tools such as Apache Airflow or Dagster to schedule, monitor and manage data pipelines that feed AI systems.
Assessing datasets against criteria such as volume, quality, completeness, labelling and bias to determine their suitability for AI model training.
Developing AI models using deep learning frameworks such as PyTorch or TensorFlow that provide tools for building, training and deploying neural networks.
Analysing suspected deepfake content — manipulated video, audio or images — using forensic techniques to determine authenticity in investigative contexts.
Using specialised tools and techniques to detect AI-generated synthetic media by analysing artefacts, inconsistencies and statistical signatures.
Creating AI-enhanced digital twins — virtual replicas of physical systems or processes — that can predict behaviour, test changes and optimise performance.
Training AI models across multiple GPUs, machines or data centres to handle large datasets and complex models that exceed single-machine capacity.
Deploying AI models to run directly on edge devices — such as sensors, cameras or mobile devices — rather than in the cloud, for low-latency local inference.
Evaluating and selecting hardware suitable for running AI inference at the edge, considering power consumption, form factor and processing capability.
Optimising AI models for efficient execution on edge devices with limited compute, memory and power resources.
Designing network architectures that support AI processing at the edge, including data flow between edge devices, fog nodes and cloud services.
Integrating AI models into embedded systems and firmware, optimising for the constrained resources of IoT devices and industrial controllers.
Building automated pipelines that convert text, images or other data into numerical vector representations for use in AI search and retrieval.
Managing and maintaining vector indexes that store embeddings, including operations such as building, updating, partitioning and optimising for query performance.
Designing the structure and metadata schemas for vector embeddings, determining what gets embedded, how chunks are defined and what metadata to attach.
Defining organisation-wide standards for AI platforms, tools and services to ensure consistency, interoperability and governance across teams.
Implementing search systems that use vector embeddings to understand the meaning of queries and documents, delivering more relevant results than keyword search.
Using AI tools to help design experiments — including hypothesis formulation, variable selection and statistical power analysis.
Using platforms such as Weights & Biases or MLflow to log, compare and visualise AI experiment parameters, metrics and results.
Using tools such as SHAP and LIME to explain why AI models make specific predictions, making their decision-making transparent to stakeholders.
Using toolkits such as Microsoft Fairlearn or IBM AIF360 to detect and measure bias in AI model outputs across different population groups.
Creating, transforming and selecting the input variables that machine learning models use to make predictions, often the most impactful part of model development.
Operating centralised feature stores — using tools like Feast or Tecton — that manage reusable, versioned features for ML model training and serving.
Systematically evaluating and comparing foundation models — such as GPT, Claude, Gemini and Llama — against standardised tasks and organisational requirements.
Adapting a pre-trained foundation model to perform better on specific tasks or domains by training it further on curated, task-specific datasets.
Using large language models such as Claude, ChatGPT or Gemini to create first drafts of written content including reports, articles, emails and documentation.
Using generative AI to create game assets, narratives, dialogue, levels and procedural content, accelerating game development workflows.
Applying generative AI to develop marketing strategies — including market analysis, persona creation, messaging frameworks and campaign concepts.
Time-boxed rapid experiments that test whether generative AI can solve a specific business problem, producing a working demonstration within days.
Building generative AI models using architectures such as GANs, diffusion models and transformer networks to create new content from learned patterns.
Managing clusters of GPU servers for AI workloads, including CUDA programming environments, NCCL communication libraries and job scheduling.
Allocating, monitoring and optimising GPU resources across AI workloads to maximise utilisation and minimise costs.
Forecasting the GPU and TPU compute capacity needed for AI training and inference workloads, planning procurement and cloud reservations accordingly.
Designing and building the physical or cloud infrastructure needed to run GPU and TPU clusters for large-scale AI model training.
Using GPU hardware to accelerate numerical computations — such as matrix operations and simulations — that underpin AI model training and scientific computing.
Approaches to reducing the environmental impact of AI, including choosing efficient model architectures, optimising training runs and using renewable-powered compute.
Studying how people interact with AI systems to understand usage patterns, pain points, trust dynamics and opportunities for improving the experience.
Established design patterns for building effective interfaces between humans and AI systems, covering feedback, control, transparency and error recovery.
Designing organisational operating models that define how human workers and AI systems share responsibilities, make decisions and escalate issues.
Breaking down work into component tasks and determining which are best suited to AI, which to humans, and which benefit from human-AI collaboration.
Modelling future workforce scenarios that account for AI automation and augmentation, projecting changes in headcount, skills mix and cost.
Designing workflows where humans review, validate or approve AI outputs at critical decision points before they are acted upon.
Designing AI interfaces and interactions that are usable and equitable for people of all abilities, backgrounds and levels of digital literacy.
Combining AI capabilities — such as NLP, computer vision and decision-making — with traditional automation to handle complex, judgment-based processes.
Designing graph databases using platforms like Neo4j to represent and query complex relationships between entities, supporting AI reasoning and RAG systems.
Designing knowledge graph structures that AI retrieval systems can traverse to find contextually relevant information for augmenting LLM responses.
Analysing the costs of using LLM APIs — including token pricing, volume discounts and infrastructure alternatives — against the business value delivered.
Integrating large language model APIs from providers such as OpenAI, Anthropic and Google into applications using their SDKs and client libraries.
Evaluating and selecting LLM API providers by comparing pricing, model capabilities, rate limits, data policies and contractual terms.
Architectural patterns for building applications powered by large language models, covering prompt management, context handling, caching and fallback strategies.
Documented procedures for supporting LLM-powered applications in production, covering common failure modes, debugging approaches and escalation paths.
Systematically testing large language models against standardised tasks and custom evaluations to assess their suitability for specific use cases.
Using quantitative metrics such as BLEU and ROUGE scores alongside human evaluations to measure the quality of LLM outputs.
Implementing gateway services — such as LiteLLM or Portkey — that provide a unified API layer across multiple LLM providers for routing, fallback and cost control.
Implementing safety and quality filters — using tools like Guardrails AI or NVIDIA NeMo Guardrails — that validate and constrain LLM outputs before they reach users.
Techniques and tools for identifying when a large language model generates factually incorrect, fabricated or unsupported content in its responses.
Curated collections of proven patterns for integrating LLMs into enterprise systems, covering authentication, error handling, retry logic and context management.
Studying techniques that attempt to circumvent LLM safety restrictions, used defensively to strengthen model guardrails and improve alignment.
Measuring and comparing the response times and request handling capacity of LLM endpoints under realistic workload conditions.
Monitoring public LLM benchmarks and leaderboards — such as the LMSYS Chatbot Arena — to stay informed about relative model performance.
Structured approaches for assessing the quality of LLM outputs across dimensions such as accuracy, relevance, safety, coherence and helpfulness.
Defined quality standards and reference examples that LLM outputs are measured against for specific use cases and domains.
Minimum acceptable quality levels for LLM outputs in production — such as accuracy rates, rejection rates and safety scores — used as deployment gates.
Regularly sampling and manually reviewing a portion of LLM outputs in production to monitor quality, catch issues and identify improvement opportunities.
Setting the operational parameters of LLMs — such as temperature, top-p, max tokens and system prompts — to achieve desired output behaviour.
Testing AI systems for vulnerability to prompt injection attacks, where malicious input attempts to override the system prompt or manipulate behaviour.
Quickly testing whether an LLM can solve a specific problem by building minimal prototypes, before committing to full development.
Managing service level agreements with LLM API providers, tracking uptime, latency, throughput and support responsiveness against commitments.
Structured methodologies for evaluating the security posture of LLM deployments, covering data exposure, access controls, prompt injection and output risks.
Providing expert guidance on selecting the right LLM for specific use cases, based on performance benchmarks, cost, compliance and integration requirements.
Calculating the total cost of ownership for LLM-based solutions, including API fees, infrastructure, development, maintenance and opportunity costs.
The process of training or further optimising large language models, including data preparation, compute management and hyperparameter tuning.
Using large language models to help analyse complex information, summarise findings and explore research questions interactively.
Using LLMs such as Claude or ChatGPT to draft, structure and refine bid responses, ensuring compliance with tender requirements.
Using LLMs to search, summarise and synthesise relevant academic and technical literature, accelerating the research review process.
Using LLMs to transform data analysis results into compelling written narratives that communicate insights to non-technical audiences.
Using LLMs to help analysts draft, refine and challenge requirements by generating questions, identifying gaps and suggesting alternatives.
Using AI assistants such as Claude or ChatGPT as research and analysis partners, helping to gather information, test hypotheses and draft findings.
Using LLMs to help design simulation parameters, scenarios and validation criteria for scientific and engineering models.
Using large language models to analyse text data — such as reports, news and social media — to identify emerging trends and patterns.
Using LLMs to automatically classify, categorise and tag content or records based on their meaning, applying consistent taxonomies at scale.
Using LLMs to generate and evaluate multiple future scenarios for strategic planning, exploring different assumptions and their implications.
Operational practices specific to managing LLMs in production, including prompt versioning, model updates, cost monitoring, quality tracking and incident response.
Using LLMs as thinking partners to generate ideas, explore possibilities and challenge assumptions during creative and strategic sessions.
Using LLMs to gather, synthesise and analyse publicly available information about competitors, markets and industry trends.
Using LLMs to analyse open-ended survey responses, extracting themes, sentiment and actionable insights from large volumes of qualitative data.
Using AI security assistants such as Microsoft Copilot for Security to analyse threat data, investigate alerts and recommend response actions.
Building or configuring connectors that use the Model Context Protocol to give AI agents secure, structured access to external data sources and tools.
Designing MCP server implementations that expose organisational tools and data to AI agents in a standardised, secure and well-documented way.
Designing how MCP servers connect AI agents to enterprise systems — such as databases, APIs and SaaS platforms — for tool-use capabilities.
Developing MCP servers that expose organisational tools, data sources and capabilities to AI agents via the Model Context Protocol standard.
Automated pipelines that take ML models from development through testing, validation and deployment, ensuring consistent and repeatable delivery.
Using MLOps platforms such as MLflow or Kubeflow to manage the full lifecycle of ML models including experimentation, training, deployment and monitoring.
The collection of tools used to operationalise ML — such as MLflow for tracking and Weights & Biases for experiment management — integrated into development workflows.
Evaluating the risk that an AI model produces biased outputs that could discriminate against or disadvantage certain groups.
Creating standardised documentation for AI models that describes their purpose, training data, performance characteristics, limitations and ethical considerations.
Using cloud platforms such as AWS SageMaker or Google Vertex AI to automate the packaging, testing and deployment of AI models to serving endpoints.
Techniques for making AI models smaller and faster — distillation transfers knowledge to a smaller model, quantisation reduces numerical precision.
Reducing the computational cost of running AI models through techniques like knowledge distillation, quantisation and pruning while maintaining acceptable quality.
Setting up and configuring the API endpoints through which applications access deployed AI models, including scaling, routing and timeout settings.
Creating visual explanations of AI model decisions using techniques such as SHAP plots that show which input features most influenced a prediction.
Tracking and verifying the complete history of an AI model — including training data, code, parameters and transformations — for compliance and reproducibility.
Specific methods for reducing the numerical precision of AI model weights and activations — such as INT8 or INT4 quantisation — to improve inference speed.
Designing the architecture for how AI models are served in production — including load balancing, caching, failover and multi-model routing strategies.
Maintaining a complete version history of AI models along with their training data, configuration and performance metrics across iterations.
Tools that allow users to query databases and datasets using plain English questions, with AI translating natural language into data queries.
Using AI features in BI tools — such as Power BI Copilot or Tableau AI — to ask questions about data in natural language and receive visual answers.
Using automated techniques to discover optimal neural network architectures for specific tasks, rather than relying solely on manual design.
Algorithms used to train neural networks effectively — including gradient descent variants, learning rate schedulers and regularisation techniques.
Using natural language processing to enable intelligent search across records management systems, finding relevant records based on meaning rather than keywords.
Specifying the quality attributes AI systems must meet beyond functionality — including response time, accuracy thresholds, fairness constraints and availability.
Running machine learning models directly on mobile and edge devices using frameworks such as TensorFlow Lite or Apple Core ML for offline, low-latency inference.
Designing systems that use ML to deliver personalised content, recommendations and experiences tailored to individual user behaviour and preferences.
Delivering learning experiences that AI dynamically customises for each individual based on their knowledge level, learning style and progress.
Neural networks that incorporate physical laws and constraints into their architecture or training process, improving predictions in scientific domains.
Implementing automated scanning to detect and protect personally identifiable information in data flowing through AI training and inference pipelines.
Using automated machine learning platforms to build predictive models without deep ML expertise, automating feature selection, model training and evaluation.
Building ML models that predict which customers are likely to stop using a product or service, enabling proactive retention interventions.
Using machine learning to forecast future demand for products, services or resources based on historical patterns and external signals.
Using AI to predict infrastructure failures, capacity issues and performance degradation before they impact services.
Using AI to detect emerging patterns in incident data that indicate a systemic problem is developing, before it causes widespread impact.
Applying AI to predict service issues, forecast demand and optimise resource allocation in technology service management.
Using AI-enhanced process mining tools such as Celonis to automatically discover, analyse and optimise business processes from event log data.
Applying AI-powered process mining from platforms like Celonis or UiPath to visualise actual process flows, identify bottlenecks and recommend improvements.
Managing changes to AI system prompts and configurations through formal change control processes, ensuring changes are reviewed, tested and traceable.
Checkpoints that evaluate whether prompts produce outputs meeting defined quality standards before they are used in production systems.
The practice of crafting effective prompts for LLMs and linking multiple prompts together in sequences to accomplish complex multi-step tasks.
Formal credentials that validate an individual's proficiency in designing effective prompts for large language models and AI systems.
Organisational guidelines and best practices for writing, formatting, testing and maintaining prompts used in AI-powered applications.
Training programmes that teach employees how to write effective prompts for AI tools, moving from basic queries to advanced techniques.
Technical measures that protect AI systems from prompt injection attacks, where malicious input attempts to override system instructions.
Specialised tools that systematically test AI applications for prompt injection vulnerabilities by simulating various attack vectors.
Automated testing that verifies prompt changes have not degraded the quality of AI outputs compared to previously validated baselines.
Formal testing where stakeholders evaluate whether AI system prompt-response pairs meet business requirements and quality standards.
Designing reusable, parameterised prompt templates that can be composed, versioned and maintained as part of an AI application's codebase.
Tracking changes to prompt templates over time using version control, enabling rollback and comparison of prompt effectiveness across versions.
Designing the user-facing aspects of prompt interactions — including input guidance, example suggestions and result presentation — for optimal usability.
Establishing production workflows where content is drafted, reviewed and refined using structured prompts across generative AI tools.
Using AI tools to generate architecture diagrams, component designs and technical documentation from natural language descriptions of requirements.
Using text prompts to direct AI design tools in creating visual assets, layouts and design concepts, iterating through natural language feedback.
Using LLMs to transform high-level requirements or feature descriptions into structured user stories with acceptance criteria.
Designing retrieval-augmented generation systems that combine vector search with LLMs to produce responses grounded in organisational knowledge.
Building end-to-end RAG systems including document ingestion, chunking, embedding, vector storage, retrieval and LLM response generation.
Selecting, preparing and maintaining the document collections that feed RAG systems, ensuring content is accurate, current and well-structured.
Designing the data flow from source documents through chunking, embedding, indexing and retrieval that powers a RAG-based AI system.
Advising on when to use retrieval-augmented generation versus model fine-tuning, based on factors such as data volume, update frequency and accuracy needs.
Using RAG systems to retrieve relevant information from organisational knowledge bases and augment AI responses with factual, sourced content.
Quickly building functional prototypes of AI-powered solutions using LLMs to validate ideas and demonstrate value before full development.
Building real-time data streaming pipelines — using technologies like Apache Kafka — that feed AI models with live data for continuous inference.
Training AI models using reinforcement learning techniques — such as RLHF, PPO or DPO — where the model learns from human feedback and reward signals.
Cross-functional bodies that oversee the ethical development and deployment of AI within an organisation, reviewing high-risk use cases and setting policies.
Assessment tools that rate AI systems against responsible AI criteria — including fairness, transparency, safety and accountability — to track compliance.
Methods for aligning AI model behaviour with human preferences and values, primarily through reinforcement learning from human feedback.
AI assistants integrated into sales tools — such as M365 Copilot or Gong AI — that provide real-time coaching, insights and administrative support.
Using large language models to explore and develop strategic scenarios by generating narratives, challenging assumptions and modelling different futures.
Using generative AI to create and run simulations of business scenarios, testing the potential impact of different decisions and market conditions.
Using NLP to automatically classify user feedback — reviews, comments, support tickets — by sentiment, identifying positive, negative and neutral opinions.
The risk posed by employees using unauthorised AI tools — such as personal ChatGPT accounts — for work tasks without organisational oversight or data controls.
Generating artificial datasets that mimic real data patterns without containing actual personal information, enabling AI development while protecting privacy.
Creating artificial datasets using statistical methods or AI models to supplement real data for training, testing or privacy-preserving scenarios.
The risk of relying on external AI providers whose services could change, degrade or become unavailable, affecting dependent organisational systems.
Measuring the concrete benefits of AI adoption by tracking how much time is saved and how much output quality improves compared to pre-AI baselines.
Predicting future LLM API token consumption based on usage patterns, planned features and growth, to inform budgeting and capacity planning.
Managing AI training data through its full lifecycle — from collection, labelling and versioning through to retirement — ensuring quality and compliance.
Automating the collection, cleaning, transformation and delivery of training data to AI model development workflows.
Ensuring that data used to train AI models is accurate, complete, representative and free from harmful biases through systematic quality controls.
Applying knowledge from a pre-trained model to a new but related task or domain, reducing the data and compute needed for effective performance.
Designing data models that represent information as numerical vectors, defining dimensions, metadata structures and relationships for AI retrieval systems.
Administering vector databases — such as Pinecone, pgvector or Qdrant — including index management, performance tuning, backups and access control.
Designing vector database schemas and architectures — using platforms like pgvector, Pinecone or Qdrant — to support efficient similarity search for AI applications.
Deploying and operating the infrastructure needed to run vector databases — such as Pinecone or Weaviate — for AI search and retrieval workloads.
Building automated pipelines that convert source data into vector embeddings, managing chunking, model selection and index updates at scale.
Operating and maintaining vector stores — such as Pinecone, Chroma or Weaviate — that hold the embeddings used by AI search and RAG systems.
Using machine learning to predict agile team delivery velocity based on historical sprint data, helping improve planning accuracy.
