AI-Ready Data Readiness Checklist

1. Data Quality Foundation

A model is only as good as its training data. These are the non-negotiables.

Completeness rates are measured and reported for all key datasets Data schemas are documented, enforced, and version-controlled Duplicate record rates are tracked and within acceptable thresholds Data validation rules are implemented at all ingestion boundaries Historical data covers at least 12 months for time-series use cases Ground truth or labelled data exists for supervised learning use cases Data distributions are profiled and documented per key feature Outlier detection and handling strategy is defined for training data

2. Data Access & Infrastructure

If data scientists need tickets to get data, AI projects die in sprint one.

All relevant data is accessible from a central location (lake, warehouse, or lakehouse) Data can be queried by ML engineers without an engineering ticket Query performance meets model training and iteration time requirements Data versioning is supported to enable reproducible experiments Data pipeline orchestration is in place (Airflow, dbt, Prefect, or equivalent) A feature store or feature registry is in place or on the near-term roadmap GPU or TPU compute is available for model training workloads Data is stored in ML-compatible formats (Parquet, TFRecord, Arrow) Streaming data is accessible if real-time or near-real-time inference is required

3. Data Governance & Compliance

AI amplifies governance problems. Find them before the model does.

Data ownership is assigned for every dataset used in AI/ML pipelines PII is identified, classified, and protected in all ML-bound datasets Data retention and deletion policies are automated and compliant Training data complies with applicable regulations (GDPR, CCPA, HIPAA) Data lineage is tracked from source system to model training Consent and licensing rights are confirmed for all training data sources A bias audit process is defined for training datasets Synthetic or anonymised data strategy exists for sensitive or scarce data

4. Feature Engineering & ML Readiness

The most common cause of model failure is a features problem, not a model problem.

Key predictive features are identified and documented by domain experts Feature definitions are agreed upon across business and data teams Features can be consistently computed at training time and inference time No target leakage exists in any training dataset Train/validation/test split strategy is defined and documented Class imbalance is assessed and a mitigation strategy is in place Temporal features are handled correctly to prevent look-ahead bias A feature importance baseline is documented Model evaluation metrics are aligned with business outcomes, not just ML metrics

5. Data Operations

Training once is a prototype. Training reliably is a data ops problem.

Data pipelines include automated quality checks at each stage Pipeline failures trigger alerts within acceptable SLA Data freshness for ML features meets inference latency requirements Retraining data pipelines are automated or fully scripted An experiment tracking tool is in use (MLflow, Weights & Biases, etc.) Model artifacts are versioned and stored in a model registry A/B testing or shadow mode infrastructure is in place or planned Data drift detection is implemented or on the near-term roadmap

6. Organizational Readiness

Data readiness without organizational alignment stalls every project eventually.

ML engineering and data engineering responsibilities are clearly defined A defined process exists for data scientists to request new or modified features Business stakeholders are involved in defining model success criteria A model governance and approval process is established before deployment Post-deployment model performance reviews are scheduled and staffed AI ethics and responsible AI guidelines are documented and communicated Budget is approved for ML infrastructure, tooling, and iteration work A named owner is accountable for AI data readiness across the organisation

Get the AI data readiness guide