Catalog
concept#AI#Data#Governance#Security

Artificial Intelligence (AI)

A high-level concept for building systems that perform tasks requiring human-like perception, learning and decision-making.

Artificial intelligence (AI) is the field of study and practice that develops systems capable of perception, learning, reasoning and autonomous decision-making.
Established
High

Classification

  • High
  • Technical
  • Architectural
  • Advanced

Technical context

Data platforms (data lake / warehouse)CI/CD pipelines for ML modelsMonitoring and observability systems

Principles & goals

Data-first: valid, representative data are the foundation of any solution.Transparency and explainability: models and decisions must be understandable.Iterative approach: prototype, measure, improve in short cycles.
Discovery
Enterprise, Domain, Team

Use cases & scenarios

Compromises

  • Bias in training data leads to unfair or discriminatory predictions.
  • Privacy breaches from improper data handling.
  • Overreliance on automated decisions without human oversight.
  • Ensure versioning of data, features and models.
  • Introduce metrics for fairness, robustness and privacy.
  • Include human oversight in critical decision paths.

I/O & resources

  • Historical training data and labels
  • Feature pipelines and metadata
  • Infrastructure for training and deployment
  • Trained models and validation reports
  • Production prediction endpoints
  • Monitoring and audit logs

Description

Artificial intelligence (AI) is the field of study and practice that develops systems capable of perception, learning, reasoning and autonomous decision-making. It spans symbolic methods, statistical machine learning and deep neural networks. AI enables automation, predictive analytics and new product capabilities across domains.

✔

  • Automation of repetitive tasks and scaling of decision-making.
  • Improved prediction accuracy through data-driven models.
  • New business opportunities and product capabilities through intelligent systems.

✖

  • Dependence on data quality and availability.
  • Limited explainability of complex models such as deep neural networks.
  • High resource demands for training and operating large models.

  • Prediction accuracy (e.g. F1 score)

    Measures model prediction quality taking precision and recall into account.

  • Inference latency

    Time between request and available prediction result in production.

  • Model drift rate

    Frequency and magnitude of model performance degradation over time.

Personalization in streaming services

Recommendation algorithms use user behavior to individualize content and increase engagement.

Predictive maintenance in industry

Sensor data and models predict failures, reduce downtime and optimize maintenance cycles.

Fraud detection in finance

Machine learning systems identify anomalies in transactions to reduce financial risk.

1

Define problem and success criteria.

2

Collect, clean data and develop suitable features.

3

Develop prototype, measure, validate and scale incrementally.

⚠️ Technical debt & bottlenecks

  • Unstructured feature pipelines hinder reproducibility.
  • Missing test and validation automation for models.
  • Outdated dependencies in ML toolchains increase maintenance costs.
Data integration across heterogeneous sourcesCompute capacity for model trainingAvailability of ML expertise
  • Using AI as sole decision authority in safety-critical cases.
  • Training models on non-representative historical data.
  • Using sensitive personal data without anonymization.
  • Overestimating model stability over time (ignoring drift).
  • Lack of collaboration between domain experts and data scientists.
  • Insufficient assessment of ethical and legal implications.
Data engineering and feature engineeringMachine learning and model evaluationProduct and domain knowledge for problem framing
Data availability and qualityLatency requirements for predictionsScalability of training and inference infrastructure
  • • Privacy and compliance requirements
  • • Budget and infrastructure limits
  • • Lack of representative training data