[{"slug":"ai-product-lead","name":"AI product lead","family":"Lead & adopt","short":"Choose the problem, define success and own the trade-offs.","market":"ai","marketNote":"Enterprise AI adoption describes potential use across businesses. It does not count AI product vacancies or experienced product leaders.","outcome":"A service that solves a defined user problem, with evidence that its benefits outweigh its operating costs and failure risks.","skills":["Problem discovery and workflow mapping","AI evaluation and product metrics","Cost, latency and quality trade-offs","Launch decisions and accountable ownership"],"levels":["Writes a problem statement and tests whether AI is needed. Separates a model demo from a customer outcome.","Runs a bounded pilot with a baseline, acceptance criteria, a named owner and user feedback.","Owns a live product, monitors failure rates and costs, and uses evidence to expand or stop it.","Coordinates several products, sets investment criteria and retires systems that no longer meet user needs."],"evidence":["A problem brief with a non-AI alternative","A pilot decision showing baseline and outcome measures","A launch review with incident and cost data","A portfolio decision showing what was stopped and why"],"exercise":"Give the candidate a plausible AI feature request, limited evaluation data and a fixed budget. Ask for a go/no-go recommendation, missing evidence and a plan for the first four weeks.","warning":"A polished roadmap without a baseline, failure criteria or someone responsible for the outcome.","next":"Choose one workflow. Agree on the baseline, an acceptable error rate and the person who can stop the pilot.","sources":["nist","digcomp"]},{"slug":"applied-ai-engineer","name":"Applied AI engineer","family":"Build & run","short":"Build AI features that work beyond the demo.","market":"gen","marketNote":"Language-generation adoption is a signal of commercial use. It includes bought software and cannot measure the supply of engineers who can build it.","outcome":"A tested application that connects models, data and tools, with predictable failure handling and operating costs.","skills":["Application engineering and model APIs","Retrieval, grounding and data permissions","Evaluation datasets and regression tests","Tool permissions, latency and cost controls"],"levels":["Builds a small model-backed feature and explains its limits. Keeps credentials and sensitive data out of examples.","Adds retrieval or tools, builds test cases and checks access controls in a supervised pilot.","Ships a maintained service with versioned evaluations, monitoring, fallbacks and incident procedures.","Sets reusable engineering patterns across teams and evaluates model changes against shared quality and cost thresholds."],"evidence":["A working feature with documented limitations","A pilot with retrieval tests and permission checks","A deployment record, evaluation history and incident drill","A migration decision supported by regression results"],"exercise":"Provide a small document collection and a question-answering prototype. Ask the candidate to identify wrong answers, design an evaluation set and prevent retrieval of documents a user cannot access.","warning":"Success is judged only by a few convincing outputs, with no record of failures or test coverage.","next":"Collect twenty representative tasks, including failure cases, and turn them into a repeatable evaluation before adding more features.","sources":["nist"]},{"slug":"machine-learning-engineer","name":"Machine learning engineer","family":"Build & run","short":"Train, evaluate and maintain models for a defined task.","market":"ml","marketNote":"Machine-learning adoption describes enterprise use for data analysis. It is not a count of ML engineers, researchers or open positions.","outcome":"A reproducible model that performs adequately on relevant unseen data and remains useful as conditions change.","skills":["Statistics and experimental design","Training pipelines and feature engineering","Leakage prevention and error analysis","Deployment, drift and model monitoring"],"levels":["Reproduces a baseline and explains train, validation and test splits.","Builds a repeatable experiment, checks leakage and reports errors across relevant cases.","Owns a deployed model, monitors performance changes and documents retraining and rollback decisions.","Sets modelling standards across teams and chooses when a simpler model or different data strategy is preferable."],"evidence":["A baseline with clear dataset boundaries","Reproducible experiments and segment-level errors","Monitoring history and a tested retraining procedure","A model review showing a justified trade-off"],"exercise":"Give the candidate an apparently strong model result with a potential time leak. Ask them to audit the split, propose a valid baseline and explain the business cost of different errors.","warning":"A single accuracy score without the test population, leakage checks or an account of costly mistakes.","next":"Audit the split and baseline before spending more on training. Document the errors that would make the model unusable.","sources":["nist"]},{"slug":"data-engineer","name":"Data engineer for AI","family":"Build & run","short":"Make data usable, traceable and available to the right people.","market":"ict","marketNote":"ICT employment is a broad pool of adjacent technical skills. It includes many occupations unrelated to data engineering or AI.","outcome":"Reliable data pipelines with clear ownership, access rules, quality checks and reproducible inputs for AI systems.","skills":["Data modelling and transformation","Pipeline reliability and observability","Lineage, access control and retention","Dataset quality and versioning"],"levels":["Builds a simple pipeline and documents the source, schema and expected update frequency.","Adds automated quality checks, access rules and reproducible dataset versions for a pilot.","Operates pipelines with service expectations, failure alerts, lineage and recoverable backfills.","Designs shared data standards and resolves ownership and quality problems across business units."],"evidence":["A source-to-output data map","A versioned dataset with failing and passing quality checks","An incident record and successful pipeline recovery","A cross-team data contract with accountable owners"],"exercise":"Present a pipeline with duplicate records, a changing schema and inconsistent access rules. Ask for the first fixes, monitoring and a safe recovery plan.","warning":"A pipeline runs, but nobody can explain which data version produced the output or who should have access.","next":"Choose the data source causing the most rework. Define ownership, a data contract and the first quality check.","sources":["nist"]},{"slug":"mlops-platform-engineer","name":"MLOps / AI platform engineer","family":"Build & run","short":"Keep AI services deployable, observable and recoverable.","market":"ict","marketNote":"ICT employment is a proxy for adjacent infrastructure skills. It does not identify people experienced in operating AI systems.","outcome":"Teams can release and operate models with controlled access, repeatable deployments and a tested response to failure.","skills":["CI/CD and infrastructure automation","Model and configuration versioning","Observability and service reliability","Capacity, access and cost management"],"levels":["Deploys a test service and records its model, configuration and dependencies.","Creates repeatable environments with basic monitoring, access controls and deployment checks.","Operates live services with defined reliability targets, rollback procedures and cost alerts.","Runs a shared platform with measured developer adoption, capacity planning and tested recovery across teams."],"evidence":["A reproducible test deployment","A deployment pipeline that blocks a known bad release","A rollback drill and service dashboard","Platform adoption and reliability results across teams"],"exercise":"Give a scenario where a model update increases latency and changes output quality. Ask for detection, containment, rollback and the evidence needed to try again.","warning":"The service can deploy, but there is no known working version to restore or owner for an incident.","next":"Run one rollback drill and assign an owner to every alert that would wake someone up.","sources":["nist"]},{"slug":"ai-governance-lead","name":"AI governance & assurance lead","family":"Govern & evaluate","short":"Turn risk decisions into checks people can carry out.","market":"ai","marketNote":"Broader AI use can create assurance work, but adoption rates do not measure governance vacancies or the availability of qualified specialists.","outcome":"AI systems have identifiable owners, documented decisions and controls that are tested in the actual workflow.","skills":["System inventory and risk assessment","Evidence collection and independent challenge","Human oversight and incident response","Coordination with legal, security and domain specialists"],"levels":["Describes an AI system, its users, data and plausible failure consequences.","Creates a pilot risk assessment with owners, required evidence and a route for unresolved concerns.","Runs recurring reviews, checks controls in practice and tracks incidents and corrective actions.","Sets proportionate assurance across a portfolio and changes controls when evidence or requirements change."],"evidence":["A system record and risk map","A review showing evidence required before launch","A tested control and tracked corrective action","A portfolio review with escalation and closure decisions"],"exercise":"Present an AI assistant that summarises internal documents. Ask for an inventory record, the three most consequential failure scenarios and evidence that proposed controls work.","warning":"A policy document is treated as proof that the live system follows the policy.","next":"List the AI systems already used in one team. Assign owners and test one control on a real workflow.","sources":["nist"]},{"slug":"ai-adoption-lead","name":"AI adoption & learning lead","family":"Lead & adopt","short":"Help teams change their work and measure the result.","market":"auto","marketNote":"AI workflow automation is a signal of use in business processes. It is not a measure of change-management capability or successful adoption.","outcome":"People can use AI appropriately in recurring work, explain when to stop and demonstrate a useful change in the workflow.","skills":["Workflow analysis and learning design","Role-specific AI literacy","Manager coaching and behaviour change","Outcome measurement and feedback"],"levels":["Identifies a recurring task and demonstrates appropriate AI use with human review.","Runs a small learning pilot with practice tasks, baseline measures and manager involvement.","Maintains role-specific learning and support, measures sustained use and checks work quality.","Coordinates adoption across functions and revises the programme when work patterns or outcomes fail to improve."],"evidence":["A before-and-after workflow example","Practice results with a baseline and review criteria","Follow-up evidence of sustained use and work quality","A programme change prompted by observed results"],"exercise":"A team has attended training but rarely uses AI. Ask the candidate to diagnose the workflow, design one practice task and choose what to measure after four weeks.","warning":"Training attendance or licence activation is reported as proof that work improved.","next":"Interview five users about one recurring task. Measure the existing workflow before designing the next training session.","sources":["digcomp","nist"]},{"slug":"ai-research-scientist","name":"AI research scientist","family":"Research","short":"Create and test new methods with reproducible evidence.","market":null,"marketNote":"This release has no comparable research-talent measure. Enterprise adoption and ICT employment cannot identify frontier research capability, so no research-market ranking is provided.","outcome":"A defensible research contribution, with clear comparisons, reproducible experiments and an honest account of what remains unknown.","skills":["Mathematical and statistical reasoning","Literature review and experimental design","Reproducible research and ablation studies","Scientific communication and research ethics"],"levels":["Reproduces a published result and identifies assumptions and missing implementation details.","Tests a focused hypothesis against credible baselines and records negative results.","Leads a research project with reproducible experiments, ablations and independent scrutiny.","Shapes a research agenda, mentors researchers and decides which promising results warrant further resources."],"evidence":["A reproduction report including discrepancies","An experiment log with appropriate baselines","A reviewed contribution with code or a reproducibility package","An agenda and resource decision grounded in research results"],"exercise":"Provide a short paper claim and its reported benchmark result. Ask for alternative explanations, missing ablations and a feasible reproduction plan.","warning":"Novelty is asserted without strong baselines, or only the best runs are reported.","next":"Reproduce the strongest relevant baseline and record the conditions under which it fails.","sources":["nist"]}]