Artificial Intelligence: Foundations
and Concepts
A structured introduction to AI systems, evaluation, and governance.
How to use this document
Each chapter is designed to stand alone. Read it end-to-end for a structured introduction, or
jump to the checklists and templates to apply the ideas immediately.
Date: 05 March 2026
Format: Executive primer + practical templates
Length: 11+ pages
AI Foundations 2026-03-05
Table of Contents
Section Page
What AI is 2
History 3
AI stack 4
Problem types 5
Evaluation 6
Responsible AI 7
Operating model 8
Checklist 9
Case study 10
Glossary 11
Note: page numbers are indicative because text may reflow depending on viewer settings.
Prepared for: General audience Page 1
AI Foundations 2026-03-05
1. What AI is (and what it is not)
Artificial Intelligence (AI) is a family of methods that enable software to perform tasks that normally
require human judgement, such as recognizing patterns, making predictions, planning actions, or
generating language. AI is not a single technology: it is an ecosystem of data, models, computing
infrastructure, and human processes.
Practical definition
AI systems map inputs to outputs under uncertainty. The output can be a label (classification), a
number (regression), a decision (policy), or new content (generation).
In business contexts, AI is valuable when it reduces decision latency, increases decision quality, or
automates repetitive cognitive work.
Common misconceptions
bullet
AI is not automatically 'autonomous' - most systems operate under human-defined objectives and
constraints.
bullet
AI does not guarantee truth - many models optimize for likelihood, not correctness.
bullet
More data is not always better - relevance and governance matter as much as volume.
Rule of thumb
If you cannot describe how the output will be used in a decision or workflow, the 'AI use case' is
not yet well defined.
Prepared for: General audience Page 2
AI Foundations 2026-03-05
2. A short history: from rules to learning
AI progressed through waves: symbolic systems based on explicit rules, statistical learning from data,
and deep learning with large-scale neural networks. Each wave improved different capabilities and
had different infrastructure requirements.
Era Core idea Typical strengths Typical limits
Symbolic AI Hand-coded rules andExplainable
logic reasoning, constraints Brittle, hard to scale
Machine Learning
Learn patterns from labeled
Prediction
dataaccuracy, flexibility Needs quality data, monitoring
Deep LearningLarge neural networksPerception, language, representation Compute intensive, less transparent
Hybrid approaches
Combine rules + learning
Robustness + control Higher design complexity
Today, many high-performing solutions are 'hybrid' in practice: learned models are surrounded by
guardrails, business rules, and human review processes.
Prepared for: General audience Page 3
AI Foundations 2026-03-05
3. The AI stack: data to decisions
A reliable AI solution is usually less about a single model and more about an end-to-end system. The
stack below is a useful way to align technical and business stakeholders.
Layer Purpose Typical artifacts
Data sources Collect signals from operations Sensors, apps, logs, documents
Ingestion Move/validate data APIs, ETL/ELT, streaming
Data layer Store and structure data Data warehouse/lakehouse, schemas
Model layer Learn or infer patterns Training code, model registry
Decision layer Turn predictions into actions Rules, optimization, approvals
Delivery Embed into workflows Dashboards, alerts, automations
Operations Run safely over time Monitoring, drift checks, audits
Why this matters
Most AI failures are system failures: missing data, unclear ownership, weak monitoring, or lack
of adoption - not 'bad algorithms'.
Prepared for: General audience Page 4
AI Foundations 2026-03-05
4. Problem types and how to choose them
Most AI initiatives can be classified into a small set of problem types. Selecting the right type early
reduces cost and increases delivery speed.
Common problem types
bullet
Classification: assign a category (fraud vs non-fraud, defect vs ok).
bullet
Regression: predict a numeric value (demand, energy usage, remaining life).
bullet
Forecasting: predict a time series (hourly occupancy, sales).
bullet
Anomaly detection: flag unusual behavior (sensor drift, security events).
bullet
Optimization: recommend the best action (scheduling, setpoints).
bullet
Generation: create content (summaries, reports, designs, code).
Selection criteria
Start from the decision, not the data. Define who uses the output, what they do with it, and what
'better' means.
Prefer problems where feedback loops exist: outcomes can be measured and the model can be
improved.
Prepared for: General audience Page 5
AI Foundations 2026-03-05
5. Measuring performance: accuracy is not enough
Model evaluation must match the business risk. A model with high accuracy can still be useless if it
misses rare events or generates too many false alarms.
Key metrics by problem type
bullet
Classification: precision, recall, F1, ROC-AUC; calibrate probabilities when needed.
bullet
Regression: MAE/MAPE for interpretability; RMSE when large errors are costly.
bullet
Forecasting: backtesting by seasonality; evaluate at the decision horizon.
bullet
Anomalies: alert rate, time-to-detect, and operator workload.
Operational metrics (often overlooked)
bullet
Decision latency: time from signal to action.
bullet
Adoption: percentage of decisions using the AI output.
bullet
Override rate: how often humans disagree and why.
bullet
Stability: drift, outages, and re-training frequency.
Good practice
Define a baseline (manual process or simple heuristic) and measure improvement against it.
'Better than before' is the real KPI.
Prepared for: General audience Page 6
AI Foundations 2026-03-05
6. Responsible AI: risk, safety, and trust
AI introduces new categories of risk: bias, privacy, security, and model errors at scale. Responsible
AI is the combination of governance processes and technical controls that reduce these risks.
Key risk themes
bullet
Data privacy and consent (especially for personal data).
bullet
Bias and disparate impact across user groups.
bullet
Security threats (prompt injection, data leakage, adversarial inputs).
bullet
Explainability and accountability for high-stakes decisions.
bullet
Reliability over time (drift, changes in environment).
Controls that scale
bullet
Document data lineage and model purpose.
bullet
Use human-in-the-loop review for high-impact outputs.
bullet
Implement access controls, logging, and audit trails.
bullet
Test models on edge cases and sensitive cohorts.
Prepared for: General audience Page 7
AI Foundations 2026-03-05
7. Operating model: who owns what
AI initiatives succeed when ownership is clear. The model is only one component: data pipelines,
product UX, and business processes need equal attention.
Role Primary responsibilities Typical outputs
Business owner Define outcomes, approve trade-offs Use case charter, KPIs
Data owner Data quality and access Schemas, access policies
ML/AI team Model development and evaluation Models, experiments
Product/UX Adoption and workflow design Screens, alerts, journeys
IT/Platform Reliability, security, scaling CI/CD, monitoring
Risk/Legal Compliance and governance Assessments, approvals
Tip
For each data source and KPI, name a single accountable owner. Shared ownership is often no
ownership.
Prepared for: General audience Page 8
AI Foundations 2026-03-05
8. AI readiness checklist (quick self-assessment)
Use the checklist below to evaluate whether an AI initiative is ready to start, or whether foundational
work is needed first.
bullet
A clear decision point exists, with a named user and workflow.
bullet
Baseline performance is known (manual, rules-based, or legacy model).
bullet
Data sources are identified, accessible, and governed.
bullet
Data quality expectations are documented (latency, completeness, accuracy).
bullet
Success metrics include both model metrics and operational metrics.
bullet
The deployment path is defined (dashboard, API, automation), including security.
bullet
Monitoring and incident response are planned from day one.
bullet
Stakeholders agree on who owns outcomes, data, and operations.
If you only do one thing
Write a one-page use-case brief: decision, user, data, baseline, target metric, and how the
output is delivered.
Prepared for: General audience Page 9
AI Foundations 2026-03-05
9. Mini case study: predictive maintenance (generic)
Predictive maintenance uses sensor and maintenance data to estimate failure risk and schedule
interventions before breakdowns occur.
Typical workflow
bullet
Collect equipment signals (vibration, temperature, run hours).
bullet
Build features and train a model to predict failure or remaining useful life.
bullet
Generate risk scores and recommended inspection windows.
bullet
Integrate with CMMS/work orders and track outcomes.
Benefits and pitfalls
bullet
Benefits: fewer outages, optimized spare parts, less reactive labor.
bullet
Pitfalls: poor labeling of failures, changing operating conditions, and alert fatigue.
What makes it work
Tie every alert to an action owner and a measurable outcome (e.g., avoided downtime hours).
Prepared for: General audience Page 10
AI Foundations 2026-03-05
10. Glossary and further reading
Term Plain-language meaning
Model A learned function that maps inputs to outputs.
Training The process of fitting a model using historical data.
Inference Using a trained model to produce outputs on new data.
Drift When the environment changes and the model degrades.
Feature A measurable attribute used as model input.
Pipeline Automated steps that move/transform data.
Governance Policies and controls ensuring safe, compliant use.
Further reading suggestions: introductory ML texts, applied MLOps guides, and organizational AI
governance playbooks. Choose sources that match your industry and regulatory constraints.
Prepared for: General audience Page 11