Camelot AI Power Hour: Practical Lessons on AI Agents, Copilot and Model Choice

Camelot AI Power Hour: 13 October 2026

Hosted by: David Melvin, Camelot Expert

Guest Speaker: Cien Solon, CEO & Founder, LaunchLemonade

Session Overview

October’s AI Power Hour welcomed Cien Solon, CEO and Founder of LaunchLemonade, for a fast-paced and highly interactive session exploring how individuals and organisations can get more value from AI.

Building on the practical themes of previous sessions, Cien addressed three questions:

• Are you getting enough from Microsoft Copilot? Making better use of existing Microsoft 365 capabilities and understanding when to introduce AI agents.

• Is the AI model pecking order being challenged? Why organisations should look beyond the best-known models and consider capability, cost, compliance and resilience.

• What makes an AI agent actually work? Understanding the components of an AI agent and how to deploy them effectively and responsibly.

  
  

The session combined live demonstrations, practical examples and lively audience discussion, with particular emphasis on making AI accessible to business leaders rather than just technical specialists.

1. What Makes an AI Agent Actually Work?

One of Cien’s central messages was that an AI agent is much more than a large language model (LLM).

Cien introduced the concept of an AI office, drawing an analogy between employing human colleagues and deploying AI agents.

Just as employees need appropriate skills, knowledge, tools, instructions, responsibilities and oversight, AI agents require a similar operating environment.

Cien identified several important components:

• LLMs: The underlying models that provide reasoning and language capabilities.

• System prompts: Instructions defining an agent’s role, responsibilities, objectives and limitations.

• Retrieval-Augmented Generation (RAG): Connecting agents to relevant organisational knowledge and information.

• Model Context Protocol (MCP): A standard enabling AI applications to connect with external tools and data sources.

• Memory: Retaining relevant information and context to support future interactions.

• Policies and guardrails: Controlling access, permissions, security and acceptable behaviour.

• Tokens: Units of information processed by models, which can influence usage costs.

  
  

Together, these components allow AI agents to undertake specific tasks, interact with other systems and potentially execute multi-step workflows.

Key takeaway

Think of an AI agent as a digital team member, not simply a chatbot. Before deploying one, establish what it needs to know, which tools it can access, what it is authorised to do and how its work will be supervised.

2. Getting More from Microsoft Copilot

Cien explored how organisations can make better use of their existing Microsoft 365 Copilot investment, particularly by moving beyond basic prompting.

She demonstrated the different environments available within Microsoft’s AI ecosystem, including Microsoft 365 Copilot, Copilot Studio and Azure AI Foundry.

These offer different levels of capability, from everyday assistance and configurable agents through to more sophisticated, custom-built AI solutions.

Practical recommendations:

• Check your Copilot environment and licensing Understand which capabilities are included in your organisation’s licence, whether Copilot has access to relevant Microsoft 365 information, and what additional functionality may be available.

• Personalise Copilot Provide relevant information about your role, responsibilities, organisation,preferred working style and recurring requirements. This can reduce repetitive prompting and help produce more relevant outputs.

• Explore different models and response modes Where available, experiment with different models and reasoning settings, particularly for more complex tasks.

• Use existing agents Explore Microsoft’s prebuilt agents, including Researcher and Analyst, before developing your own. These can support activities such as market research, analysis and the preparation of business information.

• Build agents for repetitive tasks Rather than repeatedly entering lengthy prompts, create agents with predefined instructions, appropriate knowledge and access to relevant tools.

• Consider more autonomous workflows Cien also discussed Copilot’s more autonomous capabilities, where AI can use connected applications and tools to undertake multi-step tasks. However, greater autonomy introduces additional considerations around permissions, oversight and cost.

  
  

Key takeaway

Getting more value from Copilot is not necessarily about buying more licences, rather it is about initially understanding your existing licencing, configuration and the capabilities already available to you.  

3. Is the AI Model Pecking Order Being Challenged?

Cien challenged the assumption that organisations should automatically use the largest, newest or most familiar AI model.

Different models have different strengths, and the most appropriate choice depends on the task being undertaken. Refer to www.lmarena.com.

She identified four important considerations:

• Capability: How well does the model perform the particular task? A model that excels at coding may not be the best choice for research, analysis or following detailed instructions.

• Cost: More powerful models can be significantly more expensive to operate. Smaller models may provide perfectly adequate results for routine or narrowly defined tasks.

• Compliance: Organisations, particularly those operating in regulated sectors, must consider where data is processed, how it is protected and whether the chosen service meets internal policies and regulatory obligations.

• Resilience: Reliance on a single model provider can introduce operational and strategic risks, including changes in availability, pricing, performance or geopolitical restrictions. Cien demonstrated an AI model comparison leaderboard and discussed the emergence of alternative models, including lower-cost challengers. She also showed how different models can produce noticeably different outputs from similar instructions.

  
  

Key takeaway

There is no single best AI model for every task. Organisations should select models according to their specific requirements and consider whether access to multiple models would provide additional benefits. 

4. Key Questions and Insights

The interactive discussion raised several particularly interesting points.

Why use an agent rather than simply prompting an AI tool?

• Agents can retain predefined instructions, access relevant knowledge and tools, and support repeatable workflows.

• For frequently repeated tasks, this can improve consistency and reduce the need to recreate lengthy prompts.

  
  

How do we prevent an uncontrolled proliferation of AI agents?

• One participant raised the risk of organisations encouraging employees to create and share agents without sufficient oversight.

• This could result in duplicated functionality, unnecessary costs, inconsistent outputs and poorly governed access to information.

• Cien described the use of agent review committees to evaluate, test and approve agents before wider deployment.

  
  

  How can we trust AI-generated research?

  
  

  • A particularly interesting discussion explored the limitations of AI research tools.

  • Participants emphasised the importance of:

    • Verifying factual claims against original sources.

    • Requiring clear references and evidence.

    • Distinguishing established facts from AI-generated interpretation.

    • Testing alternative or opposing hypotheses.

    • Assessing the reliability and independence of sources.

  • Mark, one of the call participants had developed a custom research agent that grades evidence, explicitly identifies unsupported AI-generated content and tests counter arguments before accepting a research conclusion. Cien also had her own research agent which she had developed.

  • The discussion reinforced that AI can dramatically accelerate research, but it does not remove the need for professional judgement and verification.

    
    

  Should organisations rely on a single AI model?

  
  

  • The discussion explored whether depending on one model provider could create longer-term operational risks.

  • A diversified approach may offer advantages in cost, capability and resilience, although this needs to be balanced against the complexity of managing multiple providers.

    
    

  5. Final Thought

  
  

  One of the key messages from the session was that the real opportunity with AI lies not simply in choosing the latest model, but in understanding how to combine models, knowledge, tools and appropriate governance to deliver practical business value.

  
  

  Cien’s demonstration showed how AI is evolving from a tool that responds to individual prompts into something capable of supporting more sophisticated, repeatable and increasingly autonomous activities.

  
  

  For independent consultants and business leaders, the challenge is to understand where these capabilities can genuinely improve productivity and quality, while retaining the judgement, accountability and oversight that remain essential.

  
  

  Our thanks to Cien Solon for an engaging and thought-provoking session, David Melvin for hosting and to everyone who contributed to the discussion.