Artificial Intelligence (AI)

What it means

An AI system combines one or more models with data, software, interfaces, and operational processes to produce outputs for explicit or implicit objectives. Systems vary in autonomy and in whether they adapt after deployment; responsibility for selecting, deploying, and overseeing them remains with people and organizations.

How to calculate it

AI is a technology category, so it has no universal formula. Evaluation must use task-specific performance plus risk and operational measures. Task performance | Accuracy, precision, recall, calibration, or domain-specific utility on representative data Impact | Severity and distribution of beneficial and harmful outcomes, including affected groups Operations | Latency, availability, drift, override rate, incidents, and cost under real use

What counts / what does not

Define the system, use context, affected people, and lifecycle rather than labeling a model alone. Include | Machine-learning and knowledge-based systems that infer outputs influencing physical or virtual environments Exclude | Every deterministic rules engine, ordinary database query, or fixed automation merely because it is complex State explicitly | Objective, inputs, model and version, outputs, human role, deployment context, affected parties, monitoring, and retirement

What moves the number

Results depend on problem framing, data, model design, evaluation coverage, system integration, human interaction, and post-deployment change. Representative data and tests determine which conditions are actually covered. Deployment context can make the same model acceptable in one use and unsafe in another. Human oversight needs authority, information, time, and an effective fallback. Monitoring detects drift, misuse, changing populations, and unexpected impact.

• Representative data and tests determine which conditions are actually covered.
• Deployment context can make the same model acceptable in one use and unsafe in another.
• Human oversight needs authority, information, time, and an effective fallback.
• Monitoring detects drift, misuse, changing populations, and unexpected impact.

When it helps

Clarifies whether AI adds enough value over simpler software or process change. Sets evidence and oversight proportional to the consequence of error. Assigns accountability across provider, deployer, operator, reviewer, and affected-user channels.

• Clarifies whether AI adds enough value over simpler software or process change.
• Sets evidence and oversight proportional to the consequence of error.
• Assigns accountability across provider, deployer, operator, reviewer, and affected-user channels.

How to use it

• AI is broader than machine learning, while generative AI is a narrower family within AI.
• A model is a component; the deployed AI system includes data, interfaces, people, and operations.
• Benchmark performance does not guarantee fitness in a new context.
• Human oversight must be designed as a functioning control, not a label.
• After deployment, monitor context, impact, anomalies, overrides, and stop conditions as well as performance.

Decision cautions

Do not move from a compelling demonstration to consequential use without contextual evaluation. Training data can encode gaps or biases that aggregate accuracy hides. Confident outputs may still be incorrect, unsafe, or unsupported. Automation can magnify harm through scale, feedback loops, and over-reliance. Monitoring and incident response must continue after launch.

• Training data can encode gaps or biases that aggregate accuracy hides.
• Confident outputs may still be incorrect, unsafe, or unsupported.
• Automation can magnify harm through scale, feedback loops, and over-reliance.
• Monitoring and incident response must continue after launch.

Read with

Combine usefulness, reliability, impact, and governance evidence. Task utility | Whether outputs improve the intended decision or workflow Robustness and calibration | Behavior under variation and confidence quality Group and impact analysis | Who benefits, who bears errors, and how severely Override and incident rates | Whether human controls and escalation work Drift and cost | Sustainability under changing data and load

Example

A support team evaluates an AI system that routes incoming cases. It compares the system with rules-based routing on representative languages and issue types, measures recall for urgent cases, reviews error impact by customer group, and sets a confidence threshold for human triage. The rollout begins in one queue with logging, override, fallback, and incident procedures. Expansion depends on observed benefit and bounded harm, not demonstration quality alone.

Compare with

AI system | Infers outputs for objectives with varying autonomy Machine learning | Techniques that learn patterns from data; a subset used in many AI systems Generative AI | AI that produces content such as text, images, audio, or code Traditional automation | Executes explicitly specified rules without the same inference behavior Model | Core inference component inside a broader operational system

Common mistakes

• AI means a system understands like a person. Output quality does not establish human-like comprehension.
• All AI learns continuously. Many deployed systems remain fixed until deliberately updated.
• Human approval automatically makes a system safe. Oversight fails without context, authority, and time.

Frequently asked questions

Sources

Related topics