Agentic AI
AI systems that can plan and take a sequence of actions towards a defined goal, within assigned permissions and controls. Business use requires clear limits, monitoring, escalation routes and human accountability.
Glossary Terms
Algorithms
Defined procedures used to process information or solve a problem. An algorithm’s usefulness depends on its purpose, inputs and assumptions; the existence of an algorithm does not establish that its output is reliable.
Anomaly Detection
Methods for identifying observations that differ from an expected pattern. An unusual result is a signal to investigate, rather than automatic proof of an error, fraud or a business opportunity.
Artificial Intelligence
Computer-based systems that infer from inputs to produce outputs such as predictions, recommendations, content or decisions. In business, assess AI against a defined problem, useful outcomes and appropriate human oversight.
Attention Mechanism
A machine learning mechanism that weights relationships between elements of an input when producing an output. Attention can improve information processing, but its weights do not by themselves explain a model’s reasoning.
Autoencoder
A neural networkA model made of connected computational units with adjustabl trained to reconstruct its input through a compressed internal representation. It can support tasks such as feature learning or anomaly detectionMethods for identifying observations that differ from an exp More, with performance assessed against the intended application.
Backpropagation
A method for calculating how a neural network’s parameters contribute to its error during training. Those gradients are used by an optimisation method to adjust parameters; useful training still depends on data and evaluation.
Bag of Words
A text representation based on the presence or frequency of words, usually without preserving their order. It can support simple classification or retrieval tasks but loses much of the text’s context.
Batch Size
The number of training examples processed together before an update to a model’s parameters. The choice affects memory use and training behaviour and should be tested for the specific model and environment.