← Resource Center
Intro
AI Foundations

Basic Principles: From Prompts to Agentic Swarms

25 min 6 sections 2 knowledge checks

A complete walkthrough of generative AI concepts, terminology, and architecture. From how LLMs work to the emerging world of autonomous agents. No technical background required.

Section 01

What is Generative AI?

You've heard the term hundreds of times. But what is it, precisely? Generative AI refers to artificial intelligence systems that can create new content: text, images, code, audio, and video. Unlike traditional software that follows rigid rules, generative AI learns patterns from vast amounts of data and produces original outputs based on those patterns.

The key distinction: traditional AI classifies and predicts (is this transaction fraudulent?), while generative AI creates and produces (draft a response to this client complaint).

T Traditional AI

Analyze data, classify inputs, predict outcomes. "Is this email spam?" "What is this customer's credit risk?" Rule-based or trained on labeled examples.

G Generative AI

Create new content from patterns learned during training. "Write a summary of this regulation." "Generate a risk assessment report." Trained on massive unlabeled datasets.

Key takeaway

Generative AI doesn't "understand" content the way a human does. It identifies statistical patterns in language and produces the most probable next output. This distinction matters when you're evaluating what to trust it with and what requires human judgment.

Section 02

Large Language Models: The Engine

At the heart of most generative AI products is a Large Language Model (LLM). These are neural networks trained on enormous volumes of text data, sometimes trillions of words from books, websites, code repositories, and public records.

The "large" in LLM refers to the number of parameters, the internal variables the model uses to make decisions. GPT-4, Claude, Gemini, and Llama are all LLMs. They differ in architecture, training data, and capabilities, but the underlying principle is the same: predict the most likely next token (word or word-piece) given everything that came before it.

How an LLM processes a request
Input
Your prompt
Tokenization
Breaks into pieces
Processing
Pattern matching
Generation
Token by token
Output
The response

When you type a question into ChatGPT or Claude, you're writing a prompt. The model processes that prompt through billions of parameters, identifies patterns relevant to your input, and generates a response one token at a time.

Terms to know
The smallest unit of text an LLM processes. Roughly 3/4 of a word in English. "Financial services" is approximately 4 tokens. Models have token limits (context windows) that cap how much information they can process at once.
The maximum amount of text (measured in tokens) a model can consider at once. Think of it as the model's working memory. A 128K context window can hold roughly 100,000 words. Larger context windows allow models to work with longer documents but increase cost and processing time.
When a model generates information that is factually incorrect but presented with confidence. Because LLMs produce statistically probable text rather than verified facts, they can and do fabricate data, cite nonexistent sources, or state regulations that don't exist. This is the single most important risk for financial services applications.
A parameter that controls how "creative" or "random" a model's output is. Low temperature (0.0-0.3) produces more predictable, conservative responses. High temperature (0.7-1.0) produces more varied and creative output. For compliance and governance work, lower temperatures are generally preferable.
Section 03

Prompts, Fine-Tuning, and RAG

There are three primary ways to customize how an LLM behaves for your organization. Understanding the differences is essential for evaluating vendor claims and making informed procurement decisions.

Prompt Engineering

Crafting specific instructions to get better outputs from a model. No changes to the model itself. Like giving a highly capable employee very detailed instructions. Fastest and cheapest approach.

Fine-Tuning

Training an existing model on your organization's specific data to permanently adjust its behavior. Like sending an employee to specialized training. Requires technical resources and curated data.

Retrieval-Augmented Generation (RAG)

Instead of training the model on your data, you give it access to your documents at the time of each query. The model searches your knowledge base, retrieves relevant information, and uses it to generate a response. Like giving an employee a well-organized filing cabinet. This is currently the most common approach for enterprise applications because it keeps your data separate from the model, reduces hallucinations, and doesn't require retraining.

Knowledge Check
Your firm wants to use an LLM to answer employee questions about internal compliance policies. The policies are updated quarterly. Which approach is most appropriate?
  • A Fine-tune the model on all compliance documents
  • B Use RAG to connect the model to a current document repository
  • C Write detailed prompts that include the full text of every policy
RAG is the right choice here. Your compliance policies change regularly, and RAG lets the model access the most current versions at query time without retraining. Fine-tuning would require retraining every quarter. Embedding full policies in prompts would exceed context window limits and become unmanageable.
Not quite. Think about what happens when the policies get updated. You need an approach that always reflects the most current information without requiring retraining or manual prompt updates.
Section 04

AI Agents: From Assistants to Autonomy

This is where the landscape is moving fastest, and where the governance implications are most significant.

An AI agent is an LLM that has been given the ability to take actions, not just generate text. Instead of simply answering a question, an agent can search databases, execute code, call APIs, send emails, or trigger workflows. The model decides which tools to use and in what order.

The AI capability spectrum

Basic Prompting

Human asks a question, model answers. No tools, no memory between sessions.

Example: "Summarize the key points of this regulation."

Tool-Augmented LLM

Model can use tools like web search, calculators, or document retrieval to enhance its responses.

Example: "Find the latest SEC guidance on AI disclosures and summarize the key requirements."

Single Agent

Model plans multi-step tasks, decides which tools to use, and executes them in sequence. Has a persistent goal.

Example: "Review all new vendor contracts from this quarter, flag any AI-related provisions, and draft a summary memo for the compliance team."

Multi-Agent Systems

Multiple specialized agents collaborate on complex tasks. One agent might research, another drafts, another reviews for compliance.

Example: A "research agent" pulls regulatory updates, a "drafting agent" writes the policy, and a "review agent" checks it against your existing framework.

Agentic Swarms

Large groups of agents with dynamic role assignment, operating with minimal human oversight. Agents can spawn new agents as needed. This is the frontier.

Example: An autonomous compliance monitoring system that identifies regulatory changes, assesses impact, drafts responses, and routes approvals, all in real time.
Governance implication

Each step up this spectrum increases capability and risk simultaneously. An agent that can take actions on behalf of your firm introduces liability questions, audit trail requirements, and accountability gaps that don't exist with basic prompting. Your governance framework needs to account for where on this spectrum each AI deployment sits.

Section 05

The Decisions That Matter

Now that you understand the core concepts, here's why they matter for your role. Every AI decision at your firm involves some combination of these considerations:

Data Residency

Where does your data go when it's sent to a model? Cloud-hosted models send data to external servers. On-premise models keep it inside your network. Hybrid approaches exist. Your regulatory obligations may dictate which is acceptable.

Model Selection

Open-source vs. proprietary. Large vs. small. General-purpose vs. specialized. Each has different cost, performance, and governance profiles. "The best model" is always "the best model for this specific use case."

Human-in-the-Loop

How much autonomy should the AI have? For high-stakes decisions (lending, compliance, client communications), human review is typically required. The question is where in the process that review happens and how it's documented.

Audit & Explainability

Can you explain why the AI made a specific recommendation? Regulators are increasingly requiring explainability for AI-driven decisions. Black-box models that produce outputs without traceable reasoning present significant compliance risk.

Section 06

Putting It All Together

Generative AI is not one thing. It's a spectrum of capabilities, from simple text generation to autonomous multi-agent systems. The terminology can be intimidating, but the underlying concepts are accessible once you see how they connect.

Knowledge Check
A vendor pitches you an "AI-powered compliance agent" that monitors regulatory changes and automatically updates your internal policies. What is the most important governance question to ask?
  • A What LLM does it use?
  • B How large is its context window?
  • C What actions can it take autonomously vs. what requires human approval?
  • D Is it built on open-source or proprietary technology?
This is the critical question. An agent that "automatically updates policies" is taking autonomous action on behalf of your firm. You need to know exactly what it can do without human approval, what requires sign-off, what the audit trail looks like, and who is accountable when it gets something wrong. The underlying model and architecture matter, but the governance boundaries matter more.
That's a reasonable technical question, but it's not the most important governance consideration. When a vendor describes autonomous action ("automatically updates"), the first question should always be about the boundaries of that autonomy and the accountability structure around it.
What you now know

You can now distinguish between traditional AI and generative AI, explain how LLMs work at a conceptual level, differentiate between prompting, fine-tuning, and RAG, and understand the spectrum from basic prompts to agentic swarms. Most importantly, you understand why these distinctions matter for governance, risk, and strategic decision-making at your firm.

Module Complete

You've completed "From Prompts to Agentic Swarms." You now have the foundational vocabulary and conceptual framework to evaluate AI initiatives, ask the right questions of vendors, and participate meaningfully in governance decisions.

Back to Resource Center
Intro
to navigate