Module 4

Agent Planning Cycle is the three-stage operational workflow of autonomous agents comprising: 1) Goal Decomposition, 2) Tool Invocation, and 3) Reflection/Error Correction.

Agentic AI vs. Generative AI is generative AI produces single-turn static content on demand, whereas Agentic AI executes goal-directed, multi-step actions across systems with feedback loops and state awareness.

Approval Checkpoint (Validation Gate) is a mandatory programmatic stop in an agent workflow where human confirmation is required before state-changing or high-consequence operations (e.g., sending emails, deleting files, financial transfers) proceed.

Chain-of-Thought (CoT) Prompting is prompting a model to break down complex reasoning problems into explicit step-by-step intermediate thoughts before providing a final answer.

Constraints & Guardrails is explicit negative rules and boundaries specified in a prompt (e.g., length limits, forbidden phrases, required formats) that restrict AI outputs to corporate standards.

Context Drift is the gradual degradation of response quality or loss of original prompt constraints that occurs when long multi-turn conversations exceed optimal attention spans.

Context Window is the finite number of tokens (words and sub-words) an LLM can retain and process across input prompts, conversation history, and output generation within a single session.

Data Cleaning is the systematic process of identifying, correcting, or removing inaccurate, duplicate, or improperly formatted records from a dataset before analysis or formula modeling.

Data Processing Agreement (DPA) is a legally binding contract between an organization and an AI vendor stipulating that customer prompt data will not be logged, reviewed by human contractors, or used for model training.

Data Redaction (Sanitization) is the mandatory operational practice of removing, anonymizing, or replacing confidential business names, financial records, and PII with generic placeholders prior to submitting AI queries.

Deterministic vs. Probabilistic Systems is deterministic software produces identical outputs for identical inputs (e.g., standard code or spreadsheets), whereas probabilistic LLMs output likelihood-weighted variations.

Few-Shot Prompting is a prompt engineering technique that includes one or more high-quality input-output examples directly inside the prompt to establish exact formatting and tone patterns.

Fine-Tuning is the process of further training a pre-trained base model on a specialized domain dataset to adapt its style, vocabulary, or task performance.

Formula Syntax Verification is the practice of prompting an LLM to explain the step-by-step logic, range boundaries, and error conditions of complex spreadsheet formulas before applying them to production workbooks.

Generative Artificial Intelligence (Generative AI) is computer systems powered by deep learning models that generate new text, code, imagery, or data patterns from natural language prompts rather than merely analyzing or classifying existing datasets.

Grounding is the practice of constraining model responses strictly to authoritative reference documents, verified databases, or explicitly provided context packets to eliminate hallucinations.

Hallucination is a generative AI phenomenon where a model produces plausible-sounding but factually false, unverified, or fabricated statements with high confidence.

Human-in-the-Loop (HITL) is a governance architecture requiring qualified human professionals to review, verify, and validate AI-generated outputs before they are acted upon or released into business operations.

Importance of Data Cleaning refers to the critical operational requirement of sanitizing, formatting, and deduplicating raw datasets prior to analysis to prevent inaccurate calculations, broken spreadsheet formulas, and AI hallucinations (the 'Garbage-In, Garbage-Out' principle).

Intellectual Property (IP) Safeguards is corporate policies and technical controls ensuring proprietary source code, internal business strategies, and patentable materials are never transmitted to public AI platforms.

Knowledge Worker refers to a professional whose primary job responsibilities involve acquiring, analyzing, creating, and communicating information, ideas, and decisions (such as analysts, project managers, technical writers, and operations leads) rather than manual physical labor.

Large Language Model (LLM) is a machine learning model trained on vast text corpora that predicts the statistically most probable next tokens in a sequence based on prompt context and attention mechanisms.

Model Bias is systematic errors, unfair stereotypes, or unrepresentative viewpoints embedded in AI outputs resulting from skewed historical training data or uncalibrated alignment tuning.

Multi-Turn Prompting is the iterative practice of refining, steering, and expanding AI outputs through sequential conversation turns rather than expecting perfection in a single prompt.

Next-Token Prediction is the core statistical mechanism of generative text models where probability distributions determine each sequential word or character sub-unit without internal subjective reasoning.

Over-Reliance Risk is the cognitive hazard of passively accepting AI-generated prose, code, or calculations without independent human critical scrutiny and verification.

Personally Identifiable Information (PII) is sensitive data that can identify an individual (e.g., social security numbers, medical records, client addresses) that must be redacted before sending prompts to external AI models.

Prompt Engineering is the disciplined practice of structuring, refining, and constraining natural language instructions to guide generative models toward accurate, consistent, and mission-aligned outputs.

Prompt Injection is a cybersecurity vulnerability where untrusted external text or user input overrides system instructions to manipulate model behavior or bypass safety guardrails.

RCTC Framework is a structured prompt architecture organizing instructions into four critical components: Role (persona), Context (background), Task (actionable deliverable), and Constraints (rules and limits).

ReAct Pattern (Reasoning and Acting) is an agent execution loop where the model explicitly generates a Thought (reasoning), executes an Action (tool call), and monitors the Observation (output) before deciding the next step.

Retrieval-Augmented Generation (RAG) is an architectural pattern that dynamically retrieves relevant document excerpts from external knowledge bases and injects them into the prompt context to ground model answers.

Role Specification is the component of a structured prompt that establishes the expertise, tone, audience perspective, and professional persona for the AI model to adopt during output generation.

SUMIFS Function is an Excel/Google Sheets formula that calculates the sum of cells meeting multiple specified criteria across distinct evaluation ranges (`SUMIFS(sum_range, criteria_range1, criteria1, ...)`).

Supervised Autonomous Workflow is an operational model where AI agents perform repetitive multi-step research and drafting autonomously within strict permission boundaries, while humans maintain final executive authority.

Synthetic Data is artificially generated text or data created by machine learning algorithms that mimics the statistical characteristics of real-world data without exposing private source records.

System Prompt is persistent, top-level instructions or custom guidelines provided to an AI model that set durable behavioral guardrails, tone, and operational boundaries across entire sessions.

Tabular Synthesis is using AI to restructure unstructured meeting notes, emails, or qualitative survey responses into structured tables with standardized column schemas.

Task-Oriented AI Agent is an autonomous software entity powered by an LLM that executes multi-step objectives by iteratively planning actions, invoking tools or APIs, and observing environment results.

Temperature is a model generation hyperparameter controlling response randomness; lower values (0.0 to 0.3) generate deterministic, focused text, while higher values (0.7 to 1.0) increase diversity.

Token is the fundamental unit of text processed by language models; typically corresponds to approximately 4 characters or 0.75 words in English.

Tool Use (Function Calling) is the capability of an LLM to identify when external actions are required (e.g., querying a database, calculating math, web searching) and emit structured parameters to invoke external code.

VLOOKUP / XLOOKUP is lookup functions used to search for a value in one column of a data table and return a corresponding value from another column in the same row.

Zero-Retention Enterprise Endpoint is a secure corporate API tier where cloud AI providers guarantee that submitted prompt payloads are discarded immediately after response delivery with zero logging.

Zero-Shot Prompting is supplying an AI model with direct task instructions without providing any illustrative examples or sample outputs.