Data and AI Strategy Weekly - December 1, 2024
Learning AI by Using AI, Andrew Ng on Agents, Personalizing styles on Claude, Claude Analysis Tool, State of AI Agents Report, Model Context Protocol, Five Leadership Strategies for AI Adoption
💡What if the key to mastering AI wasn’t about waiting for the perfect tool, but learning to disrupt your routines and habits?
This week, we’re diving into how to harness AI tools to save time and transform how you work and think.
Discover six reflective questions to supercharge your AI use and uncover hidden opportunities in your daily workflows.
Learn how to customize AI tools like Claude to fit your brand and communication style.
Explore key insights on agentic AI workflows and why they’re reshaping the future of productivity.
Plus, actionable takeaways from Andrew Ng’s Snowflake Build keynote, insights into AI agent adoption from LangChain, and leadership strategies to foster successful AI collaboration in your teams.
Learn AI by Using AI
Over the last six months, I have continually ramped up my AI tool use by taking a step back at what I am doing manually and seeing what I can accelerate or even automate. You don’t know what you don’t know until you experiment. Our habits and routines are so ingrained that I have tapped my value of curiosity to explore capabilities within these AI tools and new ones.
But here’s the thing: AI doesn’t just reward curiosity—it demands it. You're missing the point if you’re waiting for the perfect tool to transform your workflows magically. The real transformation happens when you disrupt your patterns, challenge what’s “working,” and look for ways to do it better, faster, smarter.
With AI tools constantly evolving and new ones released weekly, your ability to learn these tools by integrating them into your workflows becomes critical to your productivity and impact. I would argue that it is essential to your livelihood and competitive advantage.
Realizing our full potential requires you to go further and faster toward your vision. I am a tools nerd using ChatGPT, Claude, Gemini, and Perplexity desktop and mobile apps throughout the day. I chat and talk to these tools as friendly collaboration partners, sharing when they are awesome and what needs improvement.
Here are a few questions to reflect on this week to accelerate your AI use:
1 - Awareness of Current Routines
What tasks or processes in my workday feel repetitive or manual?
Are there parts of my workflow where I lose the most time or focus?
2 - Goal Setting
What would I achieve if I could free up 20% of my time by automating specific tasks?
What outcomes in my work or personal life could be improved with more speed, accuracy, or creativity?
3 - Curiosity and Exploration
What are some common tasks in my industry that others use AI tools to optimize?
Have I explored the features of the AI tools I already can access? What might I be missing?
4 - Embracing Experimentation
What’s one small, low-risk task I can test an AI tool on today?
How can I build a habit of regularly experimenting with new tools or features?
5 - Mindset Shift
What assumptions am I holding about my current way of working that might limit my openness to change?
What might I learn by trying an AI tool, even if it doesn’t work perfectly the first time?
6 - Outcome Measurement
How will I know if an AI tool adds value to my work? What metrics or results can I track?
Have I compared how long a task takes with and without AI to understand the impact?
Personalizing Claude’s Responses to Your Style
Anthropic released a new capability, “custom styles,” that enables you to tailor Claude’s responses to your unique style and workflows. Customizing responses can include communication preferences, tone, and structure. A consistent style is essential to reinforce your personal brand when delivering content across multiple channels like LinkedIn, Substack, and X.
Predefined Styles
Claude now includes four preset options to customize the output:
Normal - this is the base out-of-the-box tone and style.
Formal - this style is more professional, with clear and polished responses.
Concise - this style allows for shorter and more direct responses.
Explanatory - this style is more of a teaching style for learning new concepts and ideas.
While using these four present styles is quick and easy for most tasks, configuring a custom style to model your identity and brand is even more valuable for longer thought leadership pieces.
Custom Styles
There are two ways to create a custom style.
Upload an example of your writing, such as a document or post (e.g., a LinkedIn post or article). This approach is suitable if you have curated your style and want to reinforce it consistently through new content.
Describe your desired style, and Claude will create a custom style that fits your requirements. This method provides more control through specific instructions Claude uses to generate content. There are five options for describing your style.
Define style objective - here, you describe the primary purpose and goal this writing style seeks to achieve. Examples include teaching, persuasion, and inspiration.
Tailor to an audience: This method tailors content to a specific target audience, including relevant information, needs, and preferences. Connecting with an audience requires the right words and tone. I expect to use this method for my audience, primarily people in tech, specifically data and AI.
Voice and tone: Here, you specify the characteristics of the response, including formality, emotion, personality, and the emotional connection you seek to engage your audience with.
Describe generally - this method is where you describe in specific detail how you want responses generated.
Custom instructions - this method provides Claude with detailed and exact instructions on how to generate new content.
As I experiment with custom styles, I imagine using a few styles that align with my content. I envision one " teaching " style that will be explanatory and guide the reader through the core idea and how to use or apply it. This will include examples to reinforce the concept. This style will be suitable for conveying data and AI concepts I teach at UC Berkeley or through articles. I envision a “coaching” style that is more inspirational, spiritual, and emotional and leads the reader through an inner journey to empowering their best selves.
Ideas for a Style Template
These are a few of my ideas as I thought about where to start creating my style template.
Tone and Voice
Formality level (formal, semi-formal, casual, etc.)
Emotional tone (friendly, authoritative, empathetic, etc.)
Point of view (first person, second person, third person)
Example requirements:
"Use a semi-formal tone while maintaining warmth. Write in the first person plural ('we') when addressing the reader. Express enthusiasm through measured but engaging language."
Language Characteristics
Vocabulary preferences (simple, technical, industry-specific)
Sentence structure (short and direct, complex, varied)
Paragraph length (concise, detailed, mix)
Example:
"Prefer straightforward vocabulary with occasional industry terms when necessary. Vary sentence length but favor clarity over complexity. Keep paragraphs under 4 sentences."
Content Formatting
Section organization
Use of transitional phrases
Emphasis techniques (bold, italics, etc.)
Example:
"Begin each major section with a brief overview. Use clear transitions between paragraphs. Emphasize key points with italics rather than bold text."
Specific Conventions
Preferred terms and phrases
Words or phrases to avoid
Special formatting requirements
Example:
"Use 'customers' instead of 'clients.' Avoid passive voice and buzzwords. Include a key takeaway at the end of each section."
Style Examples
Good example:
"Our latest research shows that sustainable practices can increase profitability by 25%. This finding suggests that businesses should consider environmental impact when making strategic decisions."
Poor example:
"The data indicates that implementing eco-friendly protocols resulted in enhanced monetary outcomes for organizational entities."
Use Cases
I expect to use the custom styles feature to improve the consistency and quality of my writing, including LinkedIn posts, articles, and this Substack newsletter. Specifying your writing style will also help reinforce your brand through tone, format, and language consistency. If you struggle to encapsulate your style, you may need to clarify and refine your identity first.
In upcoming posts, I will share lessons learned and the structure of my custom instruction.
New Claude Experimental Features
You can discover early versions of new features by clicking on the “Feature Preview” link within the menu in the lower left of the app.
Data Analysis Tool: You can leverage Claude as a BI developer and data analyst by uploading data and asking it to conduct an analysis, such as evaluating sales data. Claude will write and run code to process the data, run the analysis, and generate a data visualization in real time. There is no need for sophisticated BI tools like Tableau when you can have Claude do the work for you.
Mathematical Equations: Claude now supports LaTex rendering to generate complex mathematical equations consistently.
Model Context Protocol by Anthropic
A popular trend for AI assistants like Claude and ChatGPT is integrating these tools with systems where data reside, such as document repositories like Google Drive and One Drive, as well as business applications. Simplifying the integration and access to these systems is the aim of an open-source project Anthropic released called “Model Context Protocol (MCP).”
The Model Context Protocol is an open standard that enables developers to build secure, two-way connections between their data sources and AI-powered tools. The high-level pattern allows developers to build against a standard protocol instead of building connectors for each data source. This will help AI systems maintain context by executing tasks between tools and datasets.
An example is Google Drive, which is now integrated into the Claude UI. For more details, see the Model Context Protocol on GitHub.
Andrew Ng on AI Agents and Trends
Here are my key takeaways from Andrew Ng’s presentation at this week’s Snowflake Build conference.
LLM-based development of machine learning apps has significantly reduced the time to value compared to traditional development. What used to take months can now take days.
Fast development leads to fast experimentation, which is a path to invention. It’s easier to build multiple prototypes quickly and choose the one that creates the most value. Fast iteration is becoming an important design pattern for fast-moving AI teams.
AI prototyping has accelerated, but there are bottlenecks in the end-to-end lifecycle for building and deploying production applications such as DevOps.
“Move fast and be responsible” is the new mantra.
Agents
Agentic AI is the most important technology trend right now.
Non-Agentic AI (Zero Shot) workflow - giving the LLM a prompt to generate a response.
An agentic workflow is executed in a series of steps, with thinking/researching and revising the response in a loop before providing the output. This process takes longer but generates a more robust response with significant performance improvement.
Agentic Reasoning Design Patterns:
Reflection: This pattern takes the output of your request to the LLM and feeds it back into the LLM to critique and evaluate its output. The LLM may identify mistakes or ways to improve the output quality. This is an iterative loop until the output achieves the desired level of performance. This pattern can include another LLM evaluating the output of the LLM that generated the response.
Tools Use (API calls): this pattern uses prompts to generate a request for an API call to decide what tool it needs to complete the task and deliver an output. Examples include doing a web search or writing and executing code.
Planning/Reasoning: A complex request is sent to an LLM, evaluated, and then decomposed into a sequence of actions to deliver on the complex task.
Multi-Agent Collaboration: you prompt the LLM to play different roles at different points in time. In this pattern, multiple agents interact with each other to accomplish the task. This may include specialized LLMs that are optimized for specific tasks. Think of this pattern as taking a main task and breaking it down into subtasks to improve performance.
Large Multimodal Model (LMM) Agents: For tasks that include multiple modes, such as vision and text, performance is improved when the primary task is broken down into steps and an iterative workflow of planning and reasoning is used.
Vision agents can be used to write code that processes videos or images to generate responses that retrieve video clips or images given a specific request.
The AI Stack has been extended with an agentic orchestration layer on top of the foundation models to coordinate agents and develop applications. Foundation models from leading companies such as Open AI, Anthropic, and Meta are powered by cloud infrastructure and semiconductor layers.
Four AI Trends:
Agentic workflows consume a lot of tokens and will benefit from services that generate faster, cheaper tokens by firms such as SambaNova and Groq.
LLMs are evolving from basic question-and-answering capabilities to models that enable agentic workflows, including planning/reasoning (e.g., OpenAI o1 models), tools use, and performing computer actions (e.g., Claude Computer Use).
Data Engineering will become essential for acquiring and staging non-structured data (e.g., text, video, and images) to create more value.
The image and video processing revolution has started and will unlock even more value with visual-based applications.
State of AI Agents by LangChain
The “State of AI Agents” report by LangChain offers valuable insights into the current landscape of AI agent adoption and application. Here are my seven key takeaways:
Widespread Adoption Across Industries: Approximately 51% of surveyed professionals have implemented AI agents in production, with mid-sized companies (100–2,000 employees) leading at 63%. Notably, 90% of respondents from non-tech sectors are either using or planning to deploy AI agents, indicating broad industry interest. 
Primary Use Cases: AI agents are predominantly utilized for research and summarization (58%), personal productivity enhancement (53.5%), and customer service improvement (45.8%). These applications highlight the role of AI in automating routine tasks and augmenting knowledge work. 
Emphasis on Control Mechanisms: Organizations prioritize tracing and observability tools to ensure reliable AI agent performance. Many implement guardrails and human oversight, and larger enterprises favor read-only permissions to mitigate risks. 
Challenges in Deployment: Performance quality is the primary barrier to AI agent deployment, overshadowing concerns like cost and safety. Smaller companies, in particular, cite performance reliability as a significant hurdle. 
Notable AI Agent Applications: Innovative platforms such as Cursor, Perplexity, and Replit exemplify successful AI agent integration. They offer advanced coding assistance and AI-powered search capabilities. 
Emerging Trends: Organizations are exploring complex AI agent capabilities, including multi-step task management, automation of repetitive tasks, effective task routing, and human-like reasoning. However, challenges persist in understanding and controlling agent behavior. 
Future Outlook: The report anticipates continued growth in AI agent adoption, focusing on enhancing performance quality, establishing robust control mechanisms, and expanding use cases across various industries.
See LangChain’s State of AI Agent report for more details.
You can learn about AI agents for free through DeepLearning.AI short courses at beginner and intermediate levels.
Five Leadership Strategies for AI Adoption
Creating value from AI requires more than technology. It requires leaders who invest in leading change management that shapes a culture of continuous learning and embracing AI capabilities that augment employee skills. The Harvard Business Review article “Set Your Team Up to Collaborate with AI Successfully” by Tomas Chamorro-Premuzic outlines five strategies that leaders can use to set up their workforce for successful human/AI collaboration:
1 - Develop an AI Augmentation Strategy
AI should automate everything possible, but true value comes from what humans can achieve with their creativity and empathy after leveraging AI. For example, recruiters can use AI for repetitive tasks, freeing them to focus on human-centric activities like aligning candidates with opportunities and understanding client needs. Success hinges on rethinking how roles evolve and equipping employees with skills that amplify the value AI creates.
2 - Focus Performance Evaluations on Output
Organizations must reward outcomes, not effort, especially as AI enables employees to do more with less. Adapt metrics to incentivize productivity gains and reskilling to encourage transparency in AI usage. This approach prevents “faking busyness” and ensures employees view AI as a tool for growth, not a threat to their jobs.
3 - Cultivate Uniquely Human Skills
The rise of AI elevates the importance of emotional intelligence, creativity, and the ability to engage critically with AI outputs. Encourage curiosity, question-asking, and a thoughtful approach to AI-generated insights. Just as a home-cooked meal surpasses microwaved food, employees should aim to create outputs beyond what AI alone can produce.
4 - Invest in Mid-Level Managers
Mid-level managers are the linchpins of AI adoption. Equip them with technical know-how and soft skills to navigate the complexities of the AI age. These managers bridge strategy and execution, impacting engagement, productivity, and overall organizational success.
5 - Promote AI Experimentation
Create a culture where employees feel safe experimenting with AI without fear of failure. Provide incentives like “innovation grants” to encourage exploration. Organizations foster adaptability, unlock creative AI applications, and inspire confidence in new technologies by reframing failures as learning opportunities.
These strategies highlight how organizations can thrive by blending human ingenuity with AI’s potential.
Data and AI Learning Resources
Here are a few resources you can leverage to continue to expand your expertise in data and AI capabilities:
Data and AI Resources Google Sheet is a spreadsheet I curate to catalog popular learning resources and tools. Note that it does not cover the universe but resources that I find valuable, and I hope to save you time.
Hands-On AI for High-Performance Leaders is a new Maven course that focuses on helping leaders and professionals strengthen their technical acumen through the hands-on use of popular tools and frameworks. The first cohort will be limited to 20 people, and Graymatter subscribers can save 20% using promo code “graymatter” through December 15th.