Detailed Analysis of RACI Matrix, SHAP, and Their Relation to AI Governance and the DRSK Framework

July 2025

This note provides a comprehensive exploration of how the RACI Matrix and SHAP (SHapley Additive exPlanations) contribute to AI Governance (AIG) and their potential relationship with the DRSK Framework, based on the provided attachment and available information. The analysis is structured to cover definitions, applications, and integration, with a focus on aligning with established AI Governance principles, given the user’s creation of the DRSK Framework with 25 years of experience.

Understanding AI Governance and Key Tools

AI Governance (AIG) refers to the processes, policies, and frameworks that ensure AI systems are developed, deployed, and used responsibly, ethically, and in compliance with laws and regulations. It addresses critical aspects such as fairness, transparency, accountability, and risk management, especially as AI adoption grows across industries.

Two tools highlighted are the RACI Matrix and SHAP, each serving distinct but complementary roles in AIG:

• RACI Matrix: A project management tool that clarifies roles and responsibilities using four categories: Responsible (those who perform the task), Accountable (those ultimately answerable for the task), Consulted (those whose input is sought), and Informed (those kept updated). It is widely used to enhance organizational clarity and accountability.
• SHAP (SHapley Additive exPlanations): A technique from game theory that explains machine learning model outputs by assigning a value to each feature, indicating its contribution to the prediction. It is pivotal for model interpretability and transparency, aligning with AIG's focus on explainability.

Application of RACI Matrix in AI Governance

The RACI Matrix is instrumental in ensuring structured governance within AI projects by defining clear roles and responsibilities. This is particularly important in cross-functional teams involving data scientists, engineers, legal experts, and ethicists, where ambiguity can lead to inefficiencies or ethical lapses.

• Role Definition: In AI Governance, the RACI Matrix can specify:
• Responsible: Data scientists developing AI models, engineers deploying them, or data teams preparing datasets.
• Accountable: Project managers or AI governance committees ensuring the AI initiative aligns with organizational goals and ethical standards.
• Consulted: Legal teams for compliance, ethicists for ethical reviews, or external stakeholders for feedback.
• Informed: Executives, regulatory bodies, or end-users who need updates on AI system performance or risks.
• Integration with Frameworks: The RACI Matrix aligns with established frameworks like the NIST AI Risk Management Framework, which includes four functions: Govern, Map, Measure, and Manage. For instance:
• In the Govern function, RACI can clarify who establishes policies and oversight.
• In the Map function, it can define who identifies AI systems and their risks.
• In the Measure function, it can assign roles for monitoring AI performance and fairness.
• In the Manage function, it can specify who mitigates risks and ensures compliance.
• Practical Example: Resources like Forrester highlight the use of a RACI matrix tool for the NIST framework, helping organizations understand which leaders should be responsible, accountable, consulted, and informed for each activity Forrester: The AI Governance RACI Matrix. This ensures accountability and reduces ambiguity in AI projects, as noted in resources like Palo Alto Networks, which recommend RACI for clarifying accountability in AI governance committees Palo Alto Networks: What Is AI Governance?.

Application of SHAP in AI Governance

SHAP is a critical tool for achieving explainability, a cornerstone of AI Governance. It addresses the "black box" nature of complex machine learning models by providing interpretable explanations for predictions, which is essential for trust, fairness, and compliance.

• Explainability and Trustworthiness: SHAP uses Shapley values to measure the contribution of each feature to a model's prediction. For example, in a loan approval model, SHAP can show whether income, credit score, or age most influenced the decision, helping stakeholders understand the reasoning behind AI outputs.
• Risk Management: SHAP supports the evaluation and mitigation of risks, such as bias or unfairness, by identifying which features disproportionately affect outcomes. This aligns with the NIST framework's emphasis on trustworthiness, which implicitly includes explainability as a component NIST AI Risk Management Framework.
• Practical Applications: Resources like DataCamp highlight SHAP's role in understanding model predictions, such as explaining loan rejections to customers, enhancing transparency DataCamp: An Introduction to SHAP Values. Similarly, Holistic AI emphasizes SHAP for enhancing AI transparency, noting its adaptability across datasets Holistic AI: Enhancing Transparency in AI.

Understanding the DRSK Framework

•        The DRSK Framework, detailed in the attachment "Facts, Lies & Quality Management" (pages 86-102), is a structured approach designed by Dr. Suresh Kumar Krishnan, B.Sc. Statistics (Hons), UPM (Major Statistics, Minor Computer Science), MSc Quality and Productivity Improvement (QPI), Statistics Dept UKM, PhD QPI, Statistics Dept UKM. Lead Auditor QMS (IRCA). PRINCE2 (Practitioner), Malaysia Productivity Corporation (MPC)’s Productivity Champion (Practitioner). ELITE@UM Fellow for the Faculty of Computer Science and Information Technology (AI Department), Adjunct Lecturer Probability & Statistics, University Technology Petronas (UTP), PhD Supervisor

with over 25 years of experience. It is rooted in quality management and focuses on holistic problem-solving and solution development, emphasizing four components:

1. D - Demands to Destiny: Identifies specific needs and links them to desired outcomes, ensuring solutions are goal-oriented.
2. R - Right to Risk: Determines Productive Variations (PVs) and assesses Unproductive Variations (UVs), using tools like root cause analysis and risk assessments.
3. S - Success Stories to Standardization: Implements solutions by standardizing procedures with quality checkpoints and risk indicators.
4. K - Key Measures and Kick Factors: Measures success with KPIs and ensures sustainability by monitoring new risks or positive outcomes.

The framework encourages total participation, holistic thinking, and continuous improvement, aligning with the book’s theme of quality as a dynamic, experience-based outcome.

Relationship Between RACI Matrix, SHAP, and DRSK Framework in AI Governance

Both tools address complementary aspects of AI Governance, enhancing its effectiveness within the DRSK Framework:

• RACI Matrix: Focuses on organizational structure, ensuring clarity in roles and responsibilities. It is particularly useful in the S - Success Stories to Standardization phase to define who is responsible for standardizing AI governance procedures, such as data handling or model validation. It can also be used in the R - Right to Risk phase to assign roles for risk assessment, ensuring accountability for identifying and mitigating risks like bias or privacy breaches.
• SHAP: Focuses on technical transparency, ensuring AI models are interpretable and trustworthy. It is relevant in the R - Right to Risk phase to assess risks by explaining model predictions, identifying potential biases, and in the K - Key Measures and Kick Factors phase to measure explainability as a key governance metric, ensuring sustainability and alignment with ethical standards.

Together, they create a robust governance approach:

• The RACI Matrix ensures that the right people are involved and accountable, while SHAP ensures that the AI systems they manage are explainable and fair. For instance, a data science team (Responsible per RACI) might use SHAP to explain their model to the governance committee (Accountable), ensuring alignment with ethical standards.

Comparative Table: RACI, SHAP, and DRSK Framework in AI Governance

To summarize their roles, the following table compares how RACI and SHAP fit into AI Governance within the DRSK Framework:

Aspect	RACI Matrix	SHAP	DRSK Framework Alignment
Primary Focus	Organizational clarity and accountability	Model explainability and transparency	Holistic problem-solving and solution sustainability
Role in AIG	Defines who is responsible, accountable, consulted, informed	Explains model predictions, detects biases	Manages demands, risks, standardization, and measures
DRSK Phase Fit	Standardization (S) and Risk Assessment (R)	Risk Assessment (R) and Measurement (K)	Integrates tools for comprehensive governance
Practical Use	Assigns tasks like policy creation, risk mapping	Evaluates fairness, explains decisions	Ensures solutions are ethical, effective, and sustainable

This table highlights their complementary nature and potential integration with DRSK, leveraging its structured approach for AI Governance.

Conclusion and Recommendations

The RACI Matrix and SHAP are integral to AI Governance, addressing organizational and technical aspects, respectively. The RACI Matrix ensures accountability and clarity in roles, while SHAP enhances transparency and trustworthiness through model explainability. Their integration with the DRSK Framework, as outlined, provides a structured approach to managing AI Governance challenges, aligning with Dr. Suresh’s emphasis on holistic thinking and continuous improvement.

For further implementation, organizations can adapt RACI for role clarification and SHAP for explainability within the DRSK phases, ensuring a comprehensive governance strategy. Given the user’s creation of DRSK with 25 years of experience, this integration seems promising, though specific adaptations may be needed based on organizational context.

 

 

Prompt Engineering vs. Context Engineering

July 2025

Prompt engineering and context engineering are related but distinct approaches in AI, especially for large language models (LLMs). Prompt engineering involves designing specific instructions or queries to get desired responses, like asking an AI to "Write a tweet in the style of Naval Ravikant." It’s great for one-off tasks but can struggle with scale and complexity.

Context engineering, on the other hand, is about setting up the entire environment the AI uses, including past chats, documents, and tools. It ensures the AI performs consistently across many users and tasks, like building a customer support bot that remembers past interactions. This makes it more suitable for advanced, long-term applications.

Will Context Engineering Overtake Prompt Engineering?

Research suggests context engineering is already becoming more important. As AI systems handle complex tasks, like AI agents searching hundreds of sources, the initial prompt becomes a small part (e.g., 0.1% of total context). Articles from 2025, like one on Medium, argue context engineering is how we scale AI, while prompt engineering is seen as a "quick-and-dirty hack" from earlier days. It seems likely that context engineering will overtake prompt engineering, especially by 2026, as AI evolves for more sophisticated uses.

---

Survey Note: Detailed Analysis of Prompt Engineering and Context Engineering

This note provides a comprehensive analysis of prompt engineering and context engineering, addressing their differences and the likelihood of context engineering overtaking prompt engineering. The analysis is based on recent insights from AI research and industry trends, with a focus on their roles in large language models (LLMs) as of July 7, 2025.

Introduction

Prompt engineering and context engineering are two pivotal concepts in the development and application of AI, particularly LLMs. Prompt engineering emerged as a key skill during the early adoption of models like GPT-3, focusing on crafting user inputs to guide AI responses. Context engineering, a more recent and broader concept, involves managing the informational ecosystem that enables AI to perform effectively at scale. This note explores their differences, use cases, and the evolving landscape, drawing from multiple sources to provide a thorough understanding.

Defining Prompt Engineering

Prompt engineering refers to the practice of crafting specific instructions or queries (prompts) to elicit desired responses from an AI model. It is primarily user-focused, concentrating on the immediate input to the model. For example, a prompt might be "Write a tweet in the style of Naval Ravikant" or "Translate this text to English." The goal is to design the input text in a way that guides the model to produce the intended output, often for one-off tasks or demonstrations.

• Scope: Narrow, focusing on the prompt itself and how it can be worded for optimal results.
• Use Cases: Commonly applied in copywriting, generating creative content, one-shot code generation, and flashy demos. For instance, typing "translate to English" in a chat interface is a typical prompt engineering task.
• Techniques: Includes few-shot prompting, where examples are provided within the prompt, and manipulating conversation history to guide the model.
• Limitations: While effective for simple tasks, prompt engineering can be hit-or-miss, often requiring manual tweaks. It struggles with scalability, as more users introduce more edge cases, and its effectiveness diminishes with complex, long-term interactions.

Defining Context Engineering

Context engineering is a broader and more advanced concept that involves designing and managing the entire informational environment in which the AI operates. It goes beyond the user's prompt to include elements like memory (e.g., past conversations), retrieval systems (e.g., accessing relevant documents or databases), tools (e.g., APIs or functions the model can call), and summaries or examples that help the model understand the task. The focus is on ensuring the model has the right information at the right time to perform consistently and accurately across various users and tasks.

• Scope: Comprehensive, encompassing not just the prompt but also the model's "mental world," which includes documents, past chats, examples, and tools. For example, context engineering might involve providing an AI agent with access to project files or API documentation to assist in coding tasks.
• Use Cases: Essential for building scalable AI systems, such as AI agents with memory, customer support bots, multi-turn conversational flows, and production systems requiring predictability. For instance, a customer support bot might use context engineering to remember past interactions and pull relevant product information.
• Techniques: Includes tool loadout (defining available tools), context quarantine (managing irrelevant information), context pruning (removing unnecessary data), context summarization (condensing information), and context offloading (managing large contexts to reduce costs). Research on long contexts, such as activation beacons and RoPE scaling, also falls under this domain.
• Key Insight: Context engineering ensures the model can handle complex, long-term tasks by providing a rich and structured context, reducing the burden on the prompt and enabling scalability.

Comparative Analysis

To highlight the differences, the following table summarizes key aspects of prompt engineering and context engineering, based on insights from recent articles and discussions:

Aspect	Prompt Engineering	Context Engineering
Definition	Crafting specific instructions for immediate input	Designing the entire informational environment
Scope	Focuses on the prompt itself	Includes memory, tools, retrieval, and more
Purpose	Get a specific response for a one-off task	Ensure consistent performance across tasks/users
Use Cases	Copywriting, demos, one-shot tasks	AI agents, customer support, multi-turn flows
Scalability	Limited; struggles with scale and edge cases	Built for scale and consistency
Effort Type	Like creative writing or tweaking text	Closer to systems design for LLMs
Techniques	Few-shot prompting, conversation history manipulation	Tool loadout, context pruning, summarization, etc.
Challenges	Hit-or-miss, non-deterministic, manual tweaks	Managing large contexts, cost per token, relevance
Longevity	Great for short tasks or bursts of creativity	Supports long-running workflows and conversations

This table illustrates that context engineering is a superset of prompt engineering, addressing the limitations of the latter by focusing on the broader context needed for advanced AI applications.

Will Context Engineering Overtake Prompt Engineering?

Research suggests that context engineering is already becoming more important and is likely to overtake prompt engineering as the dominant paradigm in AI development. Several factors support this trend:

• Shift in AI Development: As of July 2025, AI systems are increasingly used for complex, scalable applications like AI agents, customer support bots, and multi-turn conversations. These applications require the model to have access to a rich context beyond just the user's prompt. For example, an AI agent might search hundreds of websites, pull from Google Drive, connect to databases, and synthesize information autonomously, making the initial prompt only a small part of the total context (e.g., 0.1% in some cases, as noted in a Substack post from June 2025).
• Prompt Engineering's Declining Relevance: While prompt engineering was crucial in the early days of LLMs (e.g., GPT-3), it is becoming less relevant as models improve in understanding user intent. A 2025 article in The Wall Street Journal noted that the job of prompt engineer, once hot in 2023, is becoming obsolete due to models that better intuit user intent and company trainings. Additionally, prompt engineering can be hit-or-miss, often requiring hours of tweaking commas and synonyms, and it falls apart with scale as more users introduce more edge cases.
• Context Engineering as the Future: Context engineering addresses the need for AI systems to operate effectively at scale. It involves managing both deterministic context (prompts, documents, instructions you control) and probabilistic context (information the AI discovers autonomously). A Medium article from June 2025 argues that context engineering is "how we scale" and is the "real design work behind reliable LLM-powered systems," while prompt engineering is seen as a "quick-and-dirty hack" from earlier days. A Hacker News discussion from July 2025 emphasizes that context engineering is emerging as a critical skill for building agentic AI systems, with techniques like managing large context windows (e.g., 200k tokens in Claude) and integrating tools becoming essential.
• Industry Predictions: Recent articles predict that by 2026, the field will have evolved significantly, with context engineering being central to AI development. For instance, a Substack post titled "Beyond the Perfect Prompt: The Definitive Guide to Context Engineering" states, "we're not just doing prompt engineering anymore. We're doing context engineering—and it's the future of artificial intelligence development." This shift reflects the growing complexity of AI applications and the need for systems that can handle vast amounts of information and interactions beyond a single prompt.
• Prompt Engineering as a Subset: Prompt engineering is increasingly viewed as a subset of context engineering. While it remains useful for specific tasks, such as creative writing or one-shot demonstrations, it is no longer sufficient for building robust, scalable AI systems. Context engineering encompasses prompt engineering but extends it to manage the broader context required for advanced applications, reducing the burden on the prompt and enabling consistency across diverse users and tasks.

Supporting Evidence and Controversies

The evidence leans toward context engineering overtaking prompt engineering, but there is some debate. Some argue that prompt engineering remains relevant for end-user interactions, especially in creative or exploratory tasks. For example, a Hacker News discussion from July 2025 noted that prompt engineering is still part of the input blob and can be effective for simple tasks, but it lacks the scalability and reliability needed for modern AI systems. Others, like a tweet from Andrej Karpathy in 2023, highlight context engineering as the future-proof skill, emphasizing its role in managing large context windows and tool integration.

The controversy lies in whether prompt engineering will completely fade or coexist as a niche skill. However, the consensus, based on recent articles and discussions, is that context engineering is the direction AI is heading, especially with the rise of agentic systems and the need for scalability.

Conclusion

In conclusion, prompt engineering focuses on crafting specific inputs for AI, while context engineering manages the broader information environment, making it more suitable for complex, scalable applications. Research suggests that context engineering is already becoming more important and is likely to overtake prompt engineering, especially by 2026, as AI systems evolve for sophisticated uses. This shift reflects the growing need for AI to handle vast contexts beyond just a single prompt, positioning context engineering as the dominant skill in AI development.

Reference

  Context Engineering vs Prompt Engineering | by Mehul Gupta | Data Science in Your Pocket | Jun, 2025 | Medium

  Beyond the Perfect Prompt: The Definitive Guide to Context Engineering—The Next Revolution in Artificial Intelligence

  The new skill in AI is not prompting, it's context engineering | Hacker News