Choosing Your First AI Application

Key Concepts

• Generative AI Workload Selection
• Employee Skilling & Data Readiness
• Content Safety & Evaluation
• Responsible AI & Governance
• Tangible Business Value
• Data Readiness & Technical Feasibility
• Ethical Implications & Scalability
• Return on Investment (ROI)
• Human-in-the-Loop Approach

Preparing the Organization for AI

Before selecting an AI workload, the organization must be prepared across several key areas:

• Employee Skilling: Comprehensive training is needed for developers (model selection, integration), governance, operations, monitoring, security, legal teams, and end-users. Learning paths and resources should be provided for all involved.
• Data Readiness: Data permissions must be strictly enforced. AI indexes can expose previously obscure data, requiring a thorough review and correction of data access policies.
• Content Safety: Generative models are non-deterministic and creative, necessitating robust content safety mechanisms to prevent inappropriate outputs or actions.
• Evaluation Capabilities: Traditional testing methods are insufficient. New quality assurance processes are needed to ensure the generated content is accurate, appropriate, and meets requirements.
• Responsible AI: Bias must be identified and mitigated to ensure fairness and equal treatment.
• Governance: Existing company and regulatory requirements still apply to AI solutions and must be validated.

Common Generative AI Use Cases

The video outlines several common use cases for generative AI:

• Text Summarization: Summarizing large documents, creating bullet points, or generating PowerPoint presentations.
• Conversational Chatbots: Internal HR bots or external customer service bots.
• Coding Assistants: Writing, debugging, and understanding code.
• Report Drafting: Generating reports from datasets and performing deep research.

Key Factors for Choosing Your First AI Workload

The video emphasizes several key factors to consider when selecting your first AI workload:

• Tangible Business Value: The most critical factor. The AI workload must address a real business need and have a measurable impact on the organization.
  • Examples of business needs: increasing revenue, reducing costs, improving customer or employee experience.
  • Measurement is crucial: "My customer satisfaction score has risen 10%." or "sentiment about our organization has risen 20%".
  • Aim for a quick win with measurable benefits in a short timeframe.
• Data Readiness: Assess the data required for the project and the organization's readiness to provide it. High-quality, clean, and properly permissioned data is essential.
• Technical Feasibility: Avoid overly complex projects. Consider infrastructure, skilling, and integration with existing tech stacks.
  • Leverage cloud providers for GPU resources.
  • Start with simpler implementations and ready-made APIs.
  • Ensure the AI solution integrates naturally into existing user experiences (e.g., Teams, website).
• Ethical Implications: Choose a use case with minimal ethical concerns. Avoid projects that directly impact human lives or involve sensitive data.
• Scalability and ROI: Evaluate the potential for scaling the solution and the return on investment.
  • Clear success metrics are essential to demonstrate positive ROI.
  • Capture lessons learned to share knowledge across teams.

Detailed Breakdown of Key Factors

• Business Value:
  • Align with business objectives (revenue, cost, experience).
  • Measure the impact (customer satisfaction, sentiment analysis).
  • Consider urgency and impact for a quick win.
  • Manage risk (customer-facing vs. internal use).
  • Ensure the solution will be used by the target audience.
• Data:
  • High data quality is crucial (garbage in, garbage out).
  • Address bias in the data.
  • Fix data permissioning issues (over-permissioning, oversharing).
  • Ensure efficient data access and proximity to AI capabilities.
• Technical Feasibility:
  • Consider infrastructure (cloud, GPU resources).
  • Assess skilling (data scientists, machine learning engineers, etc.).
  • Start with simpler implementations and ready-made APIs.
  • Integrate with existing tech stacks and user experiences.
  • Allocate sufficient time and budget.
• Ethics:
  • Avoid sensitive data and life-impacting decisions.
  • Minimize the need for extensive governance and compliance.
• Scalability and ROI:
  • Establish clear success metrics.
  • Quantify the cost benefit of the solution.
  • Automate routine tasks and processes.
  • Address error-prone processes.

Examples of Good First Workloads

• Automating RFP Responses: Using AI to generate draft responses to RFPs based on a database of previous successful proposals. This saves time, reduces errors, and has a clear ROI.
• Help Desk Assistance: Using AI for voice-to-text transcription and knowledge base lookup to assist human agents, reducing call times and improving customer satisfaction.

Minimizing Risk and Building Trust

• Start with AI as an assistant to a human, keeping the human in the loop for validation.
• Gradually increase the AI's autonomy as confidence grows.
• Implement strong governance and evaluation processes.
• Continuously re-evaluate the AI's performance after model upgrades.

Conclusion

The key to choosing your first AI workload is to start with something safe, tangible, and valuable to the business. Focus on a real need, ensure data readiness, prioritize technical feasibility, minimize ethical concerns, and demonstrate a clear return on investment. By following these guidelines and starting with a human-in-the-loop approach, organizations can build confidence and successfully integrate AI into their operations.