termEngineering & architecture

Artificial Intelligence (AI)

Artificial intelligence is the broad field of building systems that infer from inputs to produce predictions, content, recommendations, or decisions toward human-defined objectives.

Download Markdown (.md)

01What it means

Artificial intelligence (AI) describes machine-based systems that use inputs to infer how to generate outputs such as predictions, content, recommendations, or decisions. AI is an umbrella term: it includes rule-based and machine-learning approaches, while generative AI is one narrower family. Calling a feature AI says little about its mechanism, evidence, or fitness for a particular use.[1][2]

An umbrella capability, not one mechanism AI turns inputs into inferred outputs inside a governed product system
Human-defined objective + decision rights
Inputs
Data, prompts, signals, rules, or environment
Inference mechanism
Rule-based, statistical, or learned model
Outputs
Prediction, content, recommendation, or decision
Evaluation
Quality, fairness, safety, cost, and latency
Human controls
Oversight, correction, appeal, and withdrawal
Operations
Monitoring, incidents, change, and retirement

Calling a feature AI does not establish understanding, accuracy, autonomy, neutrality, or fitness for a particular use.

02Why it matters

AI features can change product capability, cost, risk, and user expectations. Shared vocabulary helps teams separate a desired outcome from a fashionable implementation, make uncertainty visible, and assign responsibility for data, evaluation, oversight, and harm.

03When to use it

Consider AI when the task involves patterns, variation, or scale that fixed rules handle poorly and when outputs can be evaluated against a meaningful baseline. Prefer deterministic software when rules are stable, exact behavior is required, or the consequence of an unpredictable output is unacceptable.

04How it works

  1. Define the user problem, affected people, objective, constraints, and a non-AI baseline.
  2. Identify the inputs, outputs, decision rights, and acceptable uncertainty for the use case.
  3. Choose the simplest technical approach that can meet the objective and risk threshold.
  4. Evaluate quality, failure modes, fairness, privacy, security, accessibility, cost, and latency with representative cases.
  5. Deploy with proportional oversight, user controls, monitoring, incident response, and a way to withdraw the capability.

05A practical example

Example

A support product considers AI for ticket routing. The team first compares a learned classifier with existing rules, measures misrouting by issue type and customer segment, keeps agents able to correct assignments, and ships only after the model improves resolution time without increasing harmful routing errors.

06Where it breaks

AI is not a single capability and does not imply understanding, autonomy, accuracy, or neutrality. Performance depends on data, design, context, and changing operating conditions. Some failures are difficult to predict or explain, and human review can be ineffective when volume, time pressure, or automation bias is high.

Common mistakes

  • Starting with a mandate to add AI instead of a validated user problem and comparison baseline.
  • Using AI as a precise technical description without naming the model, inputs, outputs, and role in the product.
  • Treating fluent or confident output as evidence of correctness.
  • Assigning nominal human oversight without time, information, authority, or a safe fallback.

07Sources

  1. Artificial Intelligence Risk Management Framework (AI RMF 1.0)National Institute of Standards and Technology · Government frameworkArtificial Intelligence Risk Management Framework (AI RMF 1.0)
  2. AI PrinciplesOECD · Intergovernmental principlesAI Principles
View saved topics

Start typing to search titles, definitions, categories, and frameworks.