§ A–ZReference

The AI terms every leader should know.

A plain-language dictionary of the most commonly used artificial intelligence terms, written for executives and operators, not engineers. From machine learning and large language models to agents, infrastructure, and governance.

Terms: 98
Categories: 8
Format: A–Z
Author: Cybernomics

Search the dictionary

98 terms

A

11 terms

Agentic AI: AI designed to act with a degree of autonomy, breaking a goal into steps, using tools, and adapting as it goes, instead of producing a single one-shot response.
AI Agent: An AI system that can take actions toward a goal on its own, deciding which steps to take, calling tools, and reacting to results, rather than just answering a single question.
AI Alignment: The work of making sure an AI system's goals and behavior match human intentions and values, so it does what people actually want rather than just optimizing a narrow target.
AI Ethics: The study and practice of building and using AI responsibly, addressing issues like fairness, privacy, transparency, accountability, and potential harm.
AI Governance: The policies, controls, and oversight an organization puts in place to manage how AI is built, deployed, and monitored, so it stays compliant, safe, and accountable.
Algorithm: A defined set of step-by-step instructions a computer follows to solve a problem or complete a task. In AI, algorithms are the procedures used to learn patterns from data.
Artificial General Intelligence: A hypothetical AI that can understand, learn, and perform any intellectual task a human can, across domains. It does not exist today; current systems are narrow and task-specific.
Artificial Intelligence: The broad field of building computer systems that perform tasks normally requiring human intelligence, such as understanding language, recognizing images, or making decisions.
Attention Mechanism: A technique that lets a model weigh which parts of the input matter most for each piece of output. It is the core idea behind transformer models and modern language AI.
Automation: Using technology to perform tasks with little or no human effort. AI extends automation to tasks that involve judgment, language, or perception, not just fixed rules.
Autonomous Agent: An AI agent that can pursue a goal over multiple steps with minimal human intervention, deciding for itself what to do next within set boundaries.

B

5 terms

Backpropagation: The core training method for neural networks. It measures how wrong the model's output was and adjusts the internal weights backward through the network to reduce future errors.
Batch Processing: Running many AI requests together as a group rather than one at a time. It is slower to respond but cheaper, useful for work that does not need an instant answer.
Bias (Fairness): Systematic unfairness in an AI system's outputs, often because the training data reflected human or historical inequities. It can lead to discriminatory or skewed results.
Big Data: Datasets so large or complex that traditional tools struggle to process them. AI thrives on big data because more examples generally improve what a model can learn.
Black Box: A system whose internal reasoning is hard or impossible to inspect. Many AI models are black boxes: they produce answers without an easy explanation of how they got there.

C

7 terms

Chain-of-Thought: A prompting and model technique where the AI works through a problem step by step before answering, which tends to improve accuracy on reasoning-heavy tasks.
Classification: A machine learning task that sorts inputs into categories, such as labeling an email as spam or not spam, or routing a support ticket to the right team.
Clustering: An unsupervised technique that groups similar items together without predefined labels, useful for discovering natural segments in customers, documents, or behavior.
Computer Vision: The field of AI that enables machines to interpret images and video, powering tasks like object detection, facial recognition, quality inspection, and document scanning.
Context Window: The maximum amount of text (measured in tokens) a language model can consider at once. Anything beyond the window is forgotten unless re-supplied.
Copilot: An AI assistant embedded in a workflow or tool that helps a person do their job, suggesting code, drafting text, or surfacing information, while the human stays in control.
Cost per Outcome: Measuring AI spend against useful results, such as cost per resolved ticket or completed workflow, rather than raw cost per token or per request. The better gauge of value.

D

5 terms

Data Labeling: The process of tagging raw data with the correct answers, like marking which images contain a defect, so a supervised model can learn from the examples.
Data Pipeline: The automated flow that moves data from its sources through cleaning and transformation to where a model or application can use it. Reliable pipelines are foundational to AI.
Deep Learning: A branch of machine learning that uses many-layered neural networks to learn complex patterns. It powers most modern advances in language, vision, and speech.
Diffusion Model: A type of generative model that creates images (and other media) by starting from random noise and gradually refining it into a coherent result. Common in AI image tools.
Distillation: Training a smaller, cheaper model to mimic a larger one. It preserves much of the capability while cutting cost and latency, useful for deployment at scale.

E

5 terms

Edge AI: Running AI directly on a local device, like a phone, camera, or sensor, instead of in the cloud. It reduces latency and can improve privacy.
Embedding: A numerical representation of text, images, or other data that captures meaning, so that similar items sit close together. Embeddings power search and recommendations.
Ensemble Learning: Combining the predictions of several models to get a more accurate or robust result than any single model would produce on its own.
Epoch: One full pass of the training algorithm through the entire training dataset. Models typically train for many epochs to gradually improve.
Explainable AI: Techniques that make an AI system's decisions understandable to people, so users can see why a model produced a given output, important for trust and compliance.

F

7 terms

Feature: An individual measurable input a model uses to make predictions, such as a customer's age, purchase history, or the words in a document.
Feature Engineering: The craft of selecting and transforming raw data into the inputs a model can learn from most effectively. Often a major driver of model performance.
Federated Learning: A training approach where models learn across many devices or sites without the raw data ever leaving them, helping protect privacy and meet data-residency rules.
Few-shot Learning: Giving a model a handful of examples in the prompt so it can perform a task correctly, without any retraining. Contrasts with zero-shot, where no examples are given.
Fine-tuning: Taking a pretrained model and training it further on your own specific data so it performs better on your particular tasks, tone, or domain.
Foundation Model: A large model trained on broad data that can be adapted to many downstream tasks. Large language models and major image models are examples.
Function Calling: The ability of an AI model to invoke external tools, APIs, or functions, such as looking up data or sending an email, so it can act, not just generate text.

G

6 terms

Generative Adversarial Network: A model design where two networks compete, one generating fake data and one trying to detect it, driving the generator to produce increasingly realistic output.
Generative AI: AI that creates new content, text, images, audio, code, or video, rather than only analyzing or classifying existing data.
GPU: A specialized processor that performs many calculations in parallel. GPUs are the workhorse hardware for training and running modern AI models.
Gradient Descent: The optimization method that gradually adjusts a model's parameters to reduce its error, like walking downhill step by step toward the lowest point.
Grounding: Connecting a model's responses to trusted, verifiable sources or real data, so its answers are factual and traceable rather than invented.
Guardrails: Rules, filters, and checks placed around an AI system to keep its behavior safe, on-topic, and compliant, blocking harmful, off-limits, or inaccurate outputs.

H

3 terms

Hallucination: When an AI confidently produces information that is false or fabricated. A key reason outputs from generative models need verification before use.
Human-in-the-loop: A design where people review, approve, or correct an AI system's outputs at key points, balancing automation with human judgment and accountability.
Hyperparameter: A setting chosen before training that controls how a model learns, such as learning rate or number of layers, as opposed to the values the model learns on its own.

I

1 term

Inference: Using a trained model to produce an output for new input, the moment the AI actually answers a question or makes a prediction. The ongoing cost of running AI in production.

L

2 terms

Large Language Model: An AI model trained on vast amounts of text to understand and generate human-like language. The technology behind chat assistants and most generative text tools.
Latency: The delay between sending a request to an AI system and getting a response. Low latency matters for interactive, real-time experiences.

M

8 terms

Machine Learning: A branch of AI where systems learn patterns from data and improve at a task without being explicitly programmed with every rule.
Memory (Agents): An agent's ability to retain information across steps or sessions, such as past instructions or results, so it can act consistently over time rather than starting fresh each turn.
MLOps: The practices and tooling for deploying, monitoring, and maintaining machine learning models in production reliably, the AI equivalent of DevOps.
Model: The trained system that has learned patterns from data and can make predictions or generate output. The core artifact produced by machine learning.
Model Card: A short document describing a model's intended use, performance, limitations, and risks, helping teams use it responsibly and meet governance expectations.
Model Drift: The gradual decline in a model's accuracy as the real world changes and no longer matches its training data. It is why deployed models need ongoing monitoring.
Multi-agent System: A setup where several AI agents collaborate or divide work, each handling part of a task and coordinating toward a shared goal.
Multimodal: An AI system that can work with more than one type of input or output, for example understanding text, images, and audio together.

N

4 terms

Named Entity Recognition: An NLP technique that finds and labels specific items in text, such as people, companies, dates, or dollar amounts.
Narrow AI: AI built to do one specific task or a narrow set of tasks well. All AI in use today is narrow AI, in contrast to the hypothetical general intelligence.
Natural Language Processing: The field of AI focused on understanding and generating human language, underpinning translation, chatbots, search, and text analysis.
Neural Network: A model loosely inspired by the brain, made of interconnected layers of simple units that together learn to recognize complex patterns in data.

O

3 terms

Optical Character Recognition: Technology that converts images of text, like scanned documents or photos, into machine-readable, editable text.
Orchestration: Coordinating the multiple models, tools, and steps an AI workflow needs, deciding what runs when, so a complex task completes reliably.
Overfitting: When a model learns the training data too closely, including its noise, and performs poorly on new, unseen data. A common pitfall in machine learning.

P

5 terms

Parameter: An internal value a model learns during training. Modern large models have billions of parameters, which together encode what the model knows.
Pretraining: The initial, large-scale training of a foundation model on broad data, before any task-specific fine-tuning. It gives the model its general capabilities.
Prompt: The input or instruction you give a generative AI model to get a response. The quality and clarity of the prompt strongly shape the output.
Prompt Engineering: The practice of crafting and refining prompts to get more accurate, useful, or consistent results from a generative model.
Prompt Injection: An attack where hidden or malicious instructions are slipped into an AI's input to make it ignore its rules or leak information. A key security risk for AI applications.

Q

1 term

Quantization: Reducing the numerical precision of a model's parameters to make it smaller and faster to run, usually with minimal loss in quality.

R

7 terms

Random Forest: A popular machine learning method that combines many decision trees and averages their results for more accurate, stable predictions.
Red Teaming: Deliberately stress-testing an AI system by trying to make it fail, produce harmful output, or be misused, in order to find and fix weaknesses before launch.
Regression: A machine learning task that predicts a continuous number, such as a price, demand forecast, or expected revenue, rather than a category.
Reinforcement Learning: A training approach where a model learns by trial and error, receiving rewards or penalties for its actions and improving its strategy over time.
Reinforcement Learning from Human Feedback: A method of refining models using human ratings of their outputs, helping align responses with what people find helpful, safe, and appropriate.
Responsible AI: An organizational approach to designing and operating AI that is fair, transparent, secure, and accountable throughout its lifecycle.
Retrieval-Augmented Generation: A technique that lets a language model pull in relevant information from your documents or databases at answer time, improving accuracy and grounding responses in real data.

S

6 terms

Sentiment Analysis: An NLP technique that detects the emotional tone of text, such as whether a review or message is positive, negative, or neutral.
Speech Recognition: Technology that converts spoken language into written text, powering voice assistants, transcription, and ambient documentation tools.
Structured Data: Data organized in a defined format, like rows and columns in a database or spreadsheet, that is easy for software to query and analyze.
Supervised Learning: Training a model on labeled examples, where each input comes with the correct answer, so it learns to predict that answer for new inputs.
Synthetic Data: Artificially generated data that mimics real data. It is used to train or test models when real data is scarce, sensitive, or expensive to collect.
System Prompt: The behind-the-scenes instruction that sets a model's role, rules, and tone for a conversation, shaping how it responds to every user message.

T

7 terms

Temperature: A setting that controls how random or creative a generative model's output is. Lower values give focused, predictable answers; higher values give more varied ones.
Text-to-Speech: Technology that converts written text into natural-sounding spoken audio, used for voice assistants, accessibility, and audio content.
Token: A small chunk of text, often a word or part of a word, that language models read and generate. AI usage and pricing are usually measured in tokens.
Tokenization: The process of breaking text into tokens so a model can process it. How text is tokenized affects both cost and how the model interprets the input.
Training Data: The examples a model learns from. Its quality, quantity, and representativeness largely determine how well the resulting model performs.
Transformer: The neural network architecture, built around the attention mechanism, that underpins modern language models and much of today's generative AI.
Turing Test: A classic thought experiment proposing that a machine is intelligent if a human cannot reliably tell its responses apart from a person's in conversation.

U

3 terms

Underfitting: When a model is too simple to capture the patterns in the data, so it performs poorly even on the examples it was trained on.
Unstructured Data: Information without a predefined format, such as emails, documents, images, and audio. Most enterprise data is unstructured, and AI is key to making it usable.
Unsupervised Learning: Training a model on data without labeled answers, so it discovers structure, groupings, or patterns on its own.

V

1 term

Vector Database: A database built to store and search embeddings, finding items by meaning rather than exact keywords. A common building block for retrieval and semantic search.

Z

1 term

Zero-shot Learning: When a model performs a task it was not given any examples for, relying on its general training to figure out what to do from the instruction alone.

Put the vocabulary to work

Knowing the terms is step one. Getting ready is the work.

Cybernomics turns AI from scattered experiments into a durable advantage, with the economics, workflow, and governance to back it up.

Find Your Firm’s Capacity See a CPA Capacity Audit