Artificial Intelligence
Definition:
Artificial intelligence refers to the technology that enables machines to perform tasks that typically require human intelligence, such as learning from information, recognising patterns, understanding language, making predictions and solving problems. AI can range from simple systems that automate specific tasks to advanced systems capable of generating content and analysing complex information.
Example:
A company using AI-powered chatbots to resolve customer queries, algorithms detecting financial anomalies, screening resumes for hiring purposes and creating personalised marketing campaigns.
AI governance
Definition:
AI Governance refers to the policies, processes and controls organisations establish to ensure that AI systems are developed and used responsibly, safely and ethically. It covers areas such as accountability, transparency, risk management, security and compliance throughout the AI lifecycle.
Example:
A bank plans to introduce AI for customer service and fraud detection. Before deployment, each use case must be reviewed by legal, privacy, risk and technology teams under the organisation’s AI governance framework and the underlying policies and controls.
Agentic AI
Definition:
Agentic AI refers to AI systems that can independently perform tasks by understanding objectives, making decisions, planning steps and taking actions with limited human involvement. Unlike traditional AI tools that mainly respond to human instructions, Agentic AI can proactively complete multi-step tasks using available tools and information.
Example:
A retailer wants to reduce e-cart abandonment during festive sales. He deploys agentic AI that identifies customers who are unlikely to complete a purchase, generates personalised offers and sends reminders, all without requiring the retailer to direct each step.
AI Agents
Definition:
AI agents (not to be confused with Agentic AI as there is an important distinction between the two explored in an earlier blog: AI Agents vs Agentic AI: Decoding the terms and why the difference matters for data protection) are designed to perform specific tasks with defined instructions. These agents function within those specifications and are limited in their adaptability.
Example:
An employee asks an internal HR agent how many leave days remain and whether they can carry them forward to next year. The agent retrieves the information from HR systems and provides an answer based on company policy.
Algorithm
Definition:
An algorithm is a set of instructions or rules that a computer system follows to analyse and solve problems or perform a task. Algorithms are the foundation behind many AI systems and determine how inputs are processed to generate outputs.
Example:
A ride- hailing application calculates fares using factors such as distance, demand, traffic conditions and time of day. The rules used to calculate the fare are the algorithm.
Automated decision making
Definition:
Automated decision making refers to decisions made by technology systems with no human involvement. These systems analyse data and apply predefined rules or AI models to reach outcomes, such as approving applications, identifying risks or recommending actions.
Example:
A fintech company uses software to instantly approve or reject loan applications based on predefined risk criteria. No employee reviews the application unless the customer appeals the decision.
Bias
Definition:
Bias occurs when an AI system produces unfair or inaccurate outcomes because of issues in the data, design, assumptions or processes used to develop the system.
Example:
A recruitment model is trained on historical hiring data where most successful candidates came from a small group of universities. Over time, the model starts favouring applicants from those institutions, even when candidates from other universities have similar qualifications.
Fine tuning
Definition:
Fine tuning is the process of further training an existing AI model on specific data so that it performs better for a particular purpose or industry.
Example:
A law firm takes a general-purpose AI model and trains it further using legal precedents, contracts, and regulatory guidance. As a result, the model provides responses that are more relevant to legal practice.
Foundation model
Definition:
Foundation models (also called General Purpose AI models) are large AI models trained on vast amounts of data and designed to perform a wide range of tasks. They can be adopted for different uses, such as generating text, analysing information, creating images, or supporting business processes.
Example:
GPT, Llama, DALL-E, Claude for generating images.
Frontier models
Definition:
Frontier models are highly advanced AI models that represent the latest developments in AI capabilities. These models can perform complex tasks involve reasoning, language understanding and problem solving.
Example:
A multinational company adopts an advanced AI model capable of analysing documents, interpreting images, processing voice recordings, and generating strategic recommendations. Its broad capabilities distinguish it from more specialised AI systems.
Large language model
Definition:
Large Language models (LLMs) are AI models trained on vast amounts of data to analyse and learn patterns, relationships among characters and phrases to perform text- based tasks. Generally, there are two types of LLMs:
(a) Generative models that create new data similar to their training data by learning and patterns of the training data.
(b) Discriminative models that maps input features to class labels and analyse patterns to categorise between different classes.
Example:
A customer service representative asks an LLM to summarise a long email thread and draft a response. The model understands the text and generates a coherent reply.
Hallucinations
Definition:
Hallucinations occur when an AI system generates output that appears to be accurate but is factually incorrect, incomplete or made up.
Example:
An employee asks an AI assistant whether a regulator recently issued a specific compliance requirement. The AI confidently references a circular that does not exist, creating the risk of an incorrect business decision.
Human in the loop
Definition:
Human in the loop is an approach where humans remain actively involved in the operation of an AI system by reviewing, approving or correcting its outputs before they are acted upon. Human oversight helps ensure that important decisions are not made solely by AI.
Example:
An insurance company uses AI to assess motor insurance claims and recommend outcomes. However, a claims officer reviews the recommendation and supporting evidence before the final decision is communicated to the customer.
Shadow IT
Definition:
Shadow IT refers to software, applications or technology used by employees without the knowledge or approval of their organisation’s IT or security teams.
Example:
A sales employee uploads customer information into a free online AI tool to prepare a presentation more quickly. The tool was not approved by the employer, and compliance teams are unaware of its use.
