The TD Stories AI glossary

How well do you understand AI?

According to the results of a 2025 survey of 2,500 Canadians conducted by Ipsos on behalf of TD, roughly four out of five respondents gave themselves a grade of C or below when it came to their knowledge of AI.

Consider this your study guide to potentially help improve your grade.

As AI becomes more ingrained in daily life, it can sometimes feel dizzying to keep up with AI terminology. What’s the difference between machine learning and deep learning? What is agentic AI? What makes something AI in the first place?

To help demystify this new and seemingly omnipresent technology, TD Stories has collected the definitions of some of the more common terms and ideas to help promote AI literacy—and hopefully bump some of those Cs to As. Take a read of our AI glossary below.

Artificial intelligence

Most of us don’t really think about how our brains work or how we actually think our thoughts (unless we’re feeling particularly philosophical).

But some AI experts look at that thinking process and ask different questions. For instance: how can I break down a thought process into steps a computer could understand and copy?

So what does that look like? Imagine you’re working on a jigsaw puzzle. Maybe you're the kind of person who starts with the edge pieces, building a border before working your way in. Or maybe you like to focus on a particular area or colour first.

These kinds of thought processes can be described to computers in lines of code. What makes AI different from other computer programs is that rather than following strict instructions to solve the puzzle, it is programmed to first evaluate the type of puzzle, access different ways of solving it, and then determine the best way to do so.

In that way it is “thinking” and “learning” based on the problem-solving skills developed in its code.

Machine learning

Machine learning is a subset, or type, of artificial intelligence. With machine learning, AI tools learn to identify patterns within the data sets they’ve been trained on.

Remember the game “spot the differences” that you might have played with a children’s menu at a restaurant? Kids are presented with two side-by-side images and must find the differences between them.

It seems like a simple enough game – after all, young children play it – but it showcases important cognitive tasks. One person might say “if the colour of the fire hydrant is red in one image, and yellow in another, that qualifies as a difference.”

Machine learning does similar things, but much faster than any preschooler – or anyone, really.

It can go through lots of examples to spot all the differences very quickly. Each time machine learning goes over the examples it’s being trained on, it “learns” and improves.

With enough training, it can spot more complicated examples, patterns and nuances, much like a child who has moved on from “spot the difference” to more sophisticated games.

Here are some examples of what machine learning can do:

• Fraud detection: By spotting unusual patterns in client spending or transactions
• Recommendation systems: Suggest what music or video you might enjoy next based on your user history and what other users have liked
• Spam detection: Categorize emails based on certain characteristics and filtering out content that appears to be suspicious because it is not in keeping with the usual characteristics of the other emails received.

Deep Learning

Deep learning is a type of machine learning that can handle far more data and that data can be unstructured. This means it can handle information that may not be neatly organized or identified, like categorizing an image that lacks a description, or the text in a PDF document.

With deep learning, artificial neural networks – algorithms inspired by how the human brain works – learn from vast amounts of data to perform tasks. They will do these tasks repeatedly, changing the approach slightly to try and improve each time.

Deep learning requires more computing power than machine learning, and depending on the tasks, it can take much more time. However, it requires much less human intervention than machine learning.

Real world examples of deep learning:

• Image classification: AI tools can be trained to recognize and classify different kinds of images and the objects in those images. You might see this in emerging AI-powered technology that can help health professionals analyze X-rays.
• Voice-activated assistants: Deep learning currently powers some voice assistants, such as the ones built into some smartphones. With deep learning, voice assistants can learn an individual's speech patterns, convert spoken words into text, filter out background noise, and respond to particular verbal prompts
• Maintenance: Using sensor data on machinery, deep learning can help to analyze business needs, like when equipment maintenance can be optimized

Large Language Models (LLMs)

Have you ever had a conversation with a close friend and anticipated their next sentence?

It might be because based on your experience with that person, you have a good sense of their tone and the words they tend to use. It’s like you know what they’re going to say before they say it.

Computers can get pretty good at this, too.

LLMs are a type of neural network that uses statistics and deep learning techniques to analyze massive amounts of data, usually text. They are programmed to then use this data to predict what comes next in a sentence when prompted.

LLMs are good at reading and writing text and providing information in text-based interactions.

LLMs are commonly used for the following types of applications:

• Chatbots
• Document summarization
• Translating languages

Knowledge Management System

A Knowledge Management System, or KMS, is like the "brain" of an organization. It helps makes sure knowledge doesn't get lost and is accessible to the right people at the right time.

A KMS can help an organization capture, organize, and retrieve knowledge to help employees with decision making and efficiency. TD, for instance, has a lot of data that colleagues might need to access, including policies, procedures, reports and historical financial information. With a KMS-powered virtual assistant available, colleagues can more easily access this data.

Examples of AI-powered KMS include:

• The TD Securities AI Virtual Assistant is a generative-AI assistant that helps employees sort through information so the employee can better synthesize what's relevant for their clients
• Contact centre representatives at TD piloted an AI-powered virtual assistant trained with the Bank's policies and procedures to help answer client questions more quickly

Generative AI or GenAI

This type of AI is trained on massive amounts of data, including text (such as books and newspaper articles), images, videos, computer code and music. It then uses deep learning to analyze patterns, relationships, and structures within that data.

What distinguishes GenAI is that with this extensive data, it is programmed to create new content when given a specific request, or prompt, from a user.

For instance, when using a gen AI chatbot, you could ask it to write a birthday greeting in the style of William Shakespeare. The GenAI tool would analyze all of Shakespeare’s work and a plethora of birthday cards to spit out an Elizabethan-inspired birthday greeting, likely in iambic pentameter.

Predictive AI

Have you ever wondered how, when you look at your email inbox, the email platform can detect what belongs in your spam folder? Predictive AI uses both machine learning and statistical analysis to seek data patterns from historical information.

Predictive AI differs from GenAI in a couple of ways. GenAI needs more data, while predictive AI can generally use smaller and more targeted data sets.

While both generative and predictive AI can make inferences based on existing data, generative AI can use that process to make new content, while predictive AI can only use it to share a prediction or insight with the user.

Three examples of Predictive AI:

• Meteorology: Using historical data, predictive AI can help predict weather patterns
• Business: Predictive AI can analyze patterns in business data to help predict sales demand
• Healthcare: Predictive AI can be used in healthcare to help profile potential patient risks by examining clinical data and medical records

Agentic AI

Agentic AI is a field that refers to AI systems designed to act autonomously, taking actions to achieve specific goals with minimal human intervention.

When given a goal to work towards, agentic AI can proactively work through multi-step problems to arrive at a solution.

Three examples of the types of things agentic AI can do:

• Simple business tasks, like processing invoices or scheduling meetings
• Agentic AI can power self-driving cars predicting hazards and updating directions based on evolving traffic
• Support patients by automatically re-ordering prescriptions and setting up recurring doctor appointments

Foundation Models

Foundation models are almost exclusively created with deep learning, are trained on large data sets, and can be fine-tuned on a specific data set relevant to a certain field of study or organization. The pre-training of a foundation model is important because it allows the model to perform a wide range of tasks with less re-training needed.

Think of it like an MBA graduate. They learn a variety of skills specific to business administration – math, writing, leadership, how to read a balance sheet, and more. Even though the focus of the courses are on business, these skills can be used in a variety of industries, and so once fully trained, an MBA grad can work in various roles across industries and quickly adapt to new tasks with these foundational skills.

Similarly, foundation models, can provide more efficiency for each particular AI solution in organization because they can personalized and scaled more quickly and for less.

Three examples of what foundation models can do:

• Business: Foundation models are well suited to understand user preferences, and can use this information to provide specific recommendations to clients and personalized marketing
• Astronomy: Astronomers are developing foundation models to search the skies more efficiently. Early tests show that the foundation model approach for image analysis in astronomy can mean lower error rates and costs
• Medicine: Multiple companies are working to develop a foundation model to improve MRIs. The goal is that by training neural networks on a massive dataset of medical information and images, MRI image quality, scan times and diagnostic workflow can be improved.