<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Arbindo Chattopadhyay on Medium]]></title>
        <description><![CDATA[Stories by Arbindo Chattopadhyay on Medium]]></description>
        <link>https://medium.com/@avipioneer?source=rss-6fb87a57acaa------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*8bdir-gp-MaTTKXy</url>
            <title>Stories by Arbindo Chattopadhyay on Medium</title>
            <link>https://medium.com/@avipioneer?source=rss-6fb87a57acaa------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 14:38:59 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@avipioneer/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Build highly scalable, secured, and compliant AI agents]]></title>
            <link>https://medium.com/@avipioneer/build-highly-scalable-secured-and-compliant-ai-agents-8dcf5fb144fb?source=rss-6fb87a57acaa------2</link>
            <guid isPermaLink="false">https://medium.com/p/8dcf5fb144fb</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[sharepoint-online]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Arbindo Chattopadhyay]]></dc:creator>
            <pubDate>Wed, 03 Sep 2025 22:17:38 GMT</pubDate>
            <atom:updated>2025-09-03T22:17:38.183Z</atom:updated>
            <content:encoded><![CDATA[<p>Just like whole food is fuel for healthy living, content is fuel for AI agents.</p><p>Yet enterprises are missing out on this little known but powerful tech for storing, searching, and securing agent content.</p><p>It is the content that AI agents use to reason, analyze and generate new insights.</p><p>Think about AI agents such as claims and mortgage processing in financial services, or clinical studies and pharma research in healthcare, or legal, tax and audit processing in professional services.</p><p>Each of these agents needs content, and content needs storage.</p><p>The content storage in such agents needs to be secured, isolated, scalable with high throughput. In addition, architects need to ensure that the content is searchable, supports semantic index for AI, supports compliance controls such as eDiscovery, retention, and privacy.</p><p>Until now, there were 2 broad choices for such storage needs — <br> a. A blog storage such as Azure blob, Amazon EBS, Google Blobstore etc. or <br> b. Document management platform such as SharePoint.</p><p>A blob storage provides high throughput, isolated and secured storage, and SharePoint provides search, semantic index, document management capabilities, and governance controls for eDiscovery and retention.</p><p>Now, architects do not have to make a choice between a blob storage and document management platform.</p><p>Introducing SharePoint Embedded that brings together power of blob storage along with document management controls of SharePoint.</p><p>SharePoint Embedded provides isolated containers, high throughput, security, search, semantic index, Copilot, and integration with Microsoft Purview for compliance.</p><p>Microsoft Loop and Microsoft Designer are applications that use SharePoint Embedded behind the scenes.</p><p>And now enterprises can use SharePoint Embedded to build high performing, secured, compliant and AI ready applications.</p><p>Want to see a demo? check out this recorded webinar by <a href="https://lnkd.in/gikJ76Hp">https://lnkd.in/gikJ76Hp</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8dcf5fb144fb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why True AI Agents Will Reshape Work]]></title>
            <link>https://medium.com/@avipioneer/why-true-ai-agents-will-reshape-work-0382b76b2573?source=rss-6fb87a57acaa------2</link>
            <guid isPermaLink="false">https://medium.com/p/0382b76b2573</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <dc:creator><![CDATA[Arbindo Chattopadhyay]]></dc:creator>
            <pubDate>Wed, 27 Aug 2025 23:29:20 GMT</pubDate>
            <atom:updated>2025-08-27T23:29:20.037Z</atom:updated>
            <content:encoded><![CDATA[<p>Does incorporating a large language model into a workflow qualify it as an agent?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wA4Sk-3sIsNcldQSGl9Nzw.png" /></figure><p>I have found it challenging to fully comprehend the concept of an agent within the context of AI solutions. For instance, does incorporating a large language model into a workflow qualify it as an agent?</p><p>While some argue that it does, this raises questions about whether we are advancing towards genuinely agentic AI solutions capable of transforming firms into frontier firms.</p><p>It appears that we proceeded to discuss building, optimizing, managing, and securing agents without first establishing a clear understanding of what an agent is within the context of artificial intelligence. Consulting colleagues or experts in AI regarding the definition of an AI agent may reveal that there is no universally accepted interpretation. Even prominent technology firms in the field of artificial intelligence lack consensus on the definition of an AI agent or on what does not qualify as one.</p><p>An AI agent offers substantial potential for enhancing productivity. However, without a shared understanding among stakeholders, we risk missing out on the full advantages that this transformative technology can deliver.</p><p>This article seeks to provide a comprehensive overview of AI agents by examining their definitions, analyzing key distinctions, outlining shared characteristics, and highlighting several prominent examples within the industry.</p><p>The objective of this article isn’t just to arrive at a single unified definition, but to ensure that leaders remain grounded in the promises that AI Agent brings, of becoming the frontier firm.</p><h3>Definition of an Agent</h3><p>Let us begin by defining an agent, a general-purpose agent.</p><p>As per Investopedia, an agent is a person or entity authorized to act on behalf of another individual or organization, often in business, legal, or other formal contexts. The key concept is representation — a principal entrusts the agent to perform tasks, make decisions, or negotiate transactions, either broadly or with specific limitations <a href="https://www.investopedia.com/terms/a/agent.asp">Source</a></p><p>According to this definition, an agent undertakes decision-making and other responsibilities on behalf of another individual or organization. It is important to highlight these essential components for further discussion.</p><h3>Definition of AI Agent</h3><p>Next, let us review how leading technology and analysts define AI agents.</p><p><strong>Microsoft</strong></p><p>Systems that make decisions, invoke tools, and participate in workflows — sometimes independently, sometimes collaboratively. Agents differ from assistants by completing goals with less direct prompting and orchestrating tasks autonomously. <a href="https://news.microsoft.com/source/features/ai/ai-agents-what-they-are-and-how-theyll-change-the-way-we-work/">source</a></p><p><strong>OpenAI</strong></p><p>Systems that intelligently accomplish tasks, from simple workflows to complex objectives. Built from composable components (models, tools, knowledge/memory, guardrails, orchestration) to act independently, solve problems, use tools, and adapt behavior. <a href="https://platform.openai.com/docs/guides/agents">source</a></p><p><strong>Anthropic</strong></p><p>A single system capable of reasoning through any task independently, planning and iterating, directing its own tool usage, and controlling task execution with flexibility. <a href="https://www.anthropic.com/engineering/building-effective-agents">source</a></p><p><strong>Gartner</strong></p><p>An autonomous or semiautonomous software entity that uses AI to perceive, decide, act, and achieve goals in digital or physical environments. Supports or augments users by operating with autonomy to pursue goals. <a href="https://www.nojitter.com/ai-automation/conversations-in-collaboration-gartner-tom-coshow-on-ai-agents-and-agentic-ai">source</a></p><p><strong>Salesforce</strong></p><p>Goal-oriented, autonomous AI assistants that perform tasks and business interactions by initiating and completing sequences, handling conversations, adapting actions, and escalating as needed for security and ethics. <a href="https://help.salesforce.com/s/articleView?id=ai.copilot_overview.htm&amp;type=5">source</a></p><p>The following is my analysis of these definitions:</p><p>· <strong>Microsoft</strong> distinguishes between agents and assistants. This distinction is relevant because Microsoft provides M365 Copilot, which serves as an assistant, along with tools and platforms for developing, operating, and managing agents.</p><p>· <strong>OpenAI</strong> emphasizes that agents act autonomously and adapt their behavior.</p><p>· <strong>Anthropic</strong> is similar to OpenAI, but it emphasizes independent task execution. According to their blog post, Anthropic distinguishes between workflow and agent: workflows follow a predefined code path, while agents determine their own processes dynamically.</p><p>· <strong>Gartner</strong> notes that an agent may function in either digital or physical environments. For example, a self-driving car is considered an agent operating in a physical environment.</p><p>· <strong>Salesforce</strong> highlights that an agent could escalate, presumably to a human, for security and ethics.</p><p>A common theme across the definitions is that an agent operates autonomously or independently. Autonomy and independence are often considered key characteristics of an AI agent, distinguishing agents from other types of software such as workflows and microservices.</p><h3>Key characteristics — Autonomous</h3><p>This discussion will focus on autonomy. Autonomy means possessing the ability or right to make one’s own decisions, act independently, and govern oneself without external control or interference. Autonomous entities determine their actions and direction freely, based on internal reasoning and goals.</p><p>In the context of AI Agents, autonomy refers to an agent’s ability to make decisions and take actions without direct human intervention or constant oversight.</p><p>For instance, in a workflow where expense reports are routed to a human approver. An AI agent could autonomously make approval decisions or supply relevant context about the expense along with a recommendation for approval or rejection to the human reviewer.</p><p>Autonomous execution brings another factor for consideration when compared to traditional workflows — probabilistic vs deterministic output</p><h3>Probabilistic vs Deterministic output</h3><p>AI Agents are considered probabilistic because their outputs can vary for the same input, relying on models, statistical inference, and randomness to make decisions — unlike traditional deterministic workflows, which always produce the same result from the same input based on fixed rules.</p><p>This behavior of probabilistic output is like how a human would analyze certain inputs and decide. For instance, if multiple scientists are reviewing a research paper, each one would have a unique perspective based on their individual experiences and expertise.</p><p>Deterministic approaches excel in structured, critical contexts needing consistency and explainability, while probabilistic agents thrive in situations demanding adaptability, learning, and nuance amid uncertainty</p><h3>Examples of AI Agents</h3><p><strong>Microsoft Researcher Agent (within Microsoft 365 Copilot)</strong></p><p>Automates complex, multi-step research workflows by leveraging advanced AI reasoning models. It autonomously extracts key insights from massive internal and external data sources, integrates data from third-party platforms (like Salesforce, ServiceNow), and collaborates with other specialized agents to deliver comprehensive, context-aware insights. Demo <a href="https://youtu.be/lfruwkpqvk4?si=z43c8sJna7PXI-PN">video</a></p><p><strong>Siemens Industrial Automation Agent</strong></p><p>Siemens’ AI agents for industrial automation demonstrate autonomy by proactively executing complete workflows across the industrial value chain, moving beyond traditional query-based assistants to orchestrate multi-stage processes independently. These agents utilize a sophisticated orchestrator to coordinate specialized agents that understand user intent, adapt through continuous learning, and collaborate with both internal and third-party tools as needed. <a href="https://press.siemens.com/global/en/pressrelease/siemens-introduces-ai-agents-industrial-automation">Source</a></p><p><strong>DHL Enhanced Customer Solutions AI Agent</strong></p><p>DHL Supply Chain recently announced its partnership with Boston Consulting Group to deploy two generative AI applications aimed at enhancing data management and analytics. This shift is intended to improve how the company assesses proposals and develops customer-specific <a href="https://supplychain360.io/uberization-warehouses-game-changer-supply-chains-tech-driven-shift/">logistics</a> solutions <a href="https://supplychain360.io/generative-ai-benefits-dhl-supply-chain/">Source</a></p><p><strong>Microsoft Discovery (Enterprise Agentic R&amp;D Platform)</strong></p><p>A platform enabling collaboration with a team of specialized AI agents performing research, hypothesis testing, and iterative learning in R&amp;D processes. Researchers configure agents with domain knowledge and processes, allowing them to collaborate autonomously to accelerate scientific discovery. <a href="https://azure.microsoft.com/en-us/blog/transforming-rd-with-agentic-ai-introducing-microsoft-discovery/">Source</a></p><p><strong>SharePoint agents:</strong></p><p>SharePoint agents can analyze documents and answer questions about any site or library where users have permission, using natural language. Scenario <a href="https://youtu.be/stvjeYqW2wA">video 1</a>, <a href="https://youtu.be/Hzz7I9t--Bg">video 2</a></p><p><strong>Anthropic: Multi-agent research system</strong></p><p>Anthropic’s multi-agent system (powered by Claude) is deployed in research, software development, and customer support. For example, enterprises use it for literature review, knowledge discovery, coding agents for GitHub issue resolution, and handling open-ended customer support scenarios that require both conversational flow and programmatic actions. <a href="https://www.anthropic.com/engineering/multi-agent-research-system">Source</a></p><p>These cases illustrate autonomous agent capabilities like independent task planning, decision-making, tool use, and adaptive learning within complex enterprises and SaaS environments, exemplifying what defines agentic AI solutions today.</p><h3>Summary</h3><p>Agentic AI solutions show autonomy by independently perceiving environments, reasoning through complex tasks, and making decisions to achieve goals with minimal human oversight. They act flexibly and learn continuously, enabling examples such as autonomous vehicles, customer service bots, intelligent document handlers, and complex research assistants to operate effectively in dynamic real-world contexts.</p><p>This autonomy differentiates agentic AI from traditional AI tools by providing decision making and task completion capabilities, augmenting human behavior.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0382b76b2573" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Socratic approach to prompting will change the way you learn with ChatGPT or any other LLMs]]></title>
            <link>https://medium.com/@avipioneer/socratic-approach-to-prompting-will-change-the-way-you-learn-with-chatgpt-or-any-other-llms-a3af42d55f23?source=rss-6fb87a57acaa------2</link>
            <guid isPermaLink="false">https://medium.com/p/a3af42d55f23</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Arbindo Chattopadhyay]]></dc:creator>
            <pubDate>Fri, 22 Aug 2025 12:24:38 GMT</pubDate>
            <atom:updated>2025-08-22T12:24:38.222Z</atom:updated>
            <content:encoded><![CDATA[<p>Today, my 12 yo daughter showed me a cool way of using ChatGPT. This applies to every Generative AI application that uses LLMs including M365 Copilot</p><p>We are used to instant gratification where we seek quick response to our queries. In this process, we lose out on digging deeper and learning broadly about topics. We lose out on critical thinking, self-discovery, and engagement.</p><p>As my daughter pointed out, we can change that with the use of Gen AI chatbots using the Socratic learning method. It is a teaching and learning approach based on the techniques used by the ancient Greek philosopher Socrates. It emphasizes dialogue, questioning, and critical thinking rather than direct instruction</p><p>All of us can use this in our Gen AI application of choice. Socratic tutoring approach is especially powerful for humans because it mirrors how we naturally learn best: through dialogue, reflection, and discovery. Instead of passively receiving information, learners are actively engaged, challenged to think critically, and guided to uncover insights on their own. This leads to deeper understanding, better retention, and the development of essential cognitive skills like reasoning and problem-solving.</p><p>I used this approach with M365 Copilot to learn about Microsoft’s Responsible AI initiative. See the conversation using the Socratic learning platform method.</p><p>Socratic is a teaching and learning approach based on the techniques used by the ancient Greek philosopher Socrates. It emphasizes dialogue, questioning, and critical thinking rather than direct instruction. This method is especially powerful for humans because it mirrors how we naturally learn best: through dialogue, reflection, and discovery.</p><p><strong>The following is my conversation with M365 Copilot, using the Socratic method, to learn about Responsible AI initiative at Microsoft</strong></p><p><strong>Prompt </strong><a href="https://m365.cloud.microsoft/chat/entity1-d870f6cd-4aa5-4d42-9626-ab690c041429/eyJpZCI6IlZYTmxjbFl4ZkdoMGRIQnpPaTh2YzNWaWMzUnlZWFJsTFdsdWRDNXZabVpwWTJVdVkyOXRMM3hQU1VRNlpqQXlNMlE1TmprdE0yTmhNaTAwT0RSaExXSmtOR010WWpreFlXVTNZVGc1TldWaWZEVm1aVGxqTmpKbUxXSmxObUV0TkRRd09TMDVNbU0yTFRCaFltSXpNalF5TlRVMlkzd3lNREkxTFRBNExUSXhWREU1T2pNek9qRXlMams0TlRBMk1qRmEiLCJzY2VuYXJpbyI6InNoYXJlTGlua1ZpYVBhZ2UiLCJwcm9wZXJ0aWVzIjp7InByb21wdFNvdXJjZSI6InVzZXIiLCJjbGlja1RpbWVzdGFtcCI6IjIwMjUtMDgtMjFUMTk6MzM6MTMuNzYyWiJ9LCJjaGF0VHlwZSI6IndlYiIsInZlcnNpb24iOjEuMX0"><strong>Let’s use the Socratic method: You are a Socratic tutor. Please guide me through understanding the Responsible AI initiative at Microsoft by asking me a series of thought-provoking questions, and only provide explanations if I get stuck</strong></a></p><p><strong>Great approach! Let’s dive in Socratically. I’ll start with a foundational question:</strong></p><p><strong>Question 1: Why do you think a company like Microsoft would need a Responsible AI initiative in the first place?</strong></p><p>(Take your time — there’s no single “right” answer, but think about risks, opportunities, and societal impact.)</p><p><strong>Answer: </strong><a href="https://m365.cloud.microsoft/chat/entity1-d870f6cd-4aa5-4d42-9626-ab690c041429/eyJpZCI6IlZYTmxjbFl4ZkdoMGRIQnpPaTh2YzNWaWMzUnlZWFJsTFdsdWRDNXZabVpwWTJVdVkyOXRMM3hQU1VRNlpqQXlNMlE1TmprdE0yTmhNaTAwT0RSaExXSmtOR010WWpreFlXVTNZVGc1TldWaWZERmxZVEkzWW1FeExXVmtNVFV0TkRBNFpDMDRNR1JoTFRNMU0yWXpNR0kxTlRFeE5ud3lNREkxTFRBNExUSXhWREU1T2pNek9qUXhMakk1T1RreE16bGEiLCJzY2VuYXJpbyI6InNoYXJlTGlua1ZpYVBhZ2UiLCJwcm9wZXJ0aWVzIjp7InByb21wdFNvdXJjZSI6InVzZXIiLCJjbGlja1RpbWVzdGFtcCI6IjIwMjUtMDgtMjFUMTk6MzM6NDEuNjQ5WiJ9LCJjaGF0VHlwZSI6IndlYiIsInZlcnNpb24iOjEuMX0"><strong>To ensure that consumers of Microsoft’s services feel confident and safe about using AI services</strong></a></p><p><strong>That’s a strong starting point! Building trust and safety is definitely central.</strong></p><p><strong>Follow-up Question 2: If trust and safety are the goals, what kinds of risks or harms do you think Microsoft is trying to prevent through Responsible AI practices?</strong></p><p>(Hint: Think about both technical risks and societal impacts.)</p><p><strong>Answer: </strong><a href="https://m365.cloud.microsoft/chat/entity1-d870f6cd-4aa5-4d42-9626-ab690c041429/eyJpZCI6IlZYTmxjbFl4ZkdoMGRIQnpPaTh2YzNWaWMzUnlZWFJsTFdsdWRDNXZabVpwWTJVdVkyOXRMM3hQU1VRNlpqQXlNMlE1TmprdE0yTmhNaTAwT0RSaExXSmtOR010WWpreFlXVTNZVGc1TldWaWZHWXlOemRrWlRKbUxXRTVZemt0TkdabU15MWhOekU0TFRFeE5tUm1NMk5qTjJSa1pud3lNREkxTFRBNExUSXhWREU1T2pNek9qUTJMamM0T0RrMU16VmEiLCJzY2VuYXJpbyI6InNoYXJlTGlua1ZpYVBhZ2UiLCJwcm9wZXJ0aWVzIjp7InByb21wdFNvdXJjZSI6InVzZXIiLCJjbGlja1RpbWVzdGFtcCI6IjIwMjUtMDgtMjFUMTk6MzM6NDcuMTUwWiJ9LCJjaGF0VHlwZSI6IndlYiIsInZlcnNpb24iOjEuMX0"><strong>One of the risks I see is where the AI goes rogue and tries to suggest actions that could be illegal or harmful to the user</strong></a></p><p><strong>Exactly, that’s a critical risk — AI suggesting harmful or illegal actions could have severe consequences.</strong></p><p><strong>Follow-up Question 3: If preventing harmful or illegal suggestions is one goal, what kinds of principles or guidelines do you think Microsoft might adopt to ensure AI systems behave responsibly?</strong></p><p>(Hint: Think in terms of values like fairness, accountability, transparency, etc.)</p><p><strong>Answer: </strong><a href="https://m365.cloud.microsoft/chat/entity1-d870f6cd-4aa5-4d42-9626-ab690c041429/eyJpZCI6IlZYTmxjbFl4ZkdoMGRIQnpPaTh2YzNWaWMzUnlZWFJsTFdsdWRDNXZabVpwWTJVdVkyOXRMM3hQU1VRNlpqQXlNMlE1TmprdE0yTmhNaTAwT0RSaExXSmtOR010WWpreFlXVTNZVGc1TldWaWZEQmlPVEZsTmpOaUxUZzVaV0V0TkRsaU5TMWlOVEU1TFRVd1l6ZzBZbU0yWkdVMk1ud3lNREkxTFRBNExUSXhWREU1T2pNek9qVXpMall4TXpFNE9EWmEiLCJzY2VuYXJpbyI6InNoYXJlTGlua1ZpYVBhZ2UiLCJwcm9wZXJ0aWVzIjp7InByb21wdFNvdXJjZSI6InVzZXIiLCJjbGlja1RpbWVzdGFtcCI6IjIwMjUtMDgtMjFUMTk6MzM6NTQuMDc3WiJ9LCJjaGF0VHlwZSI6IndlYiIsInZlcnNpb24iOjEuMX0"><strong>One of the guidelines could include checks in the technology such that AI can identify when it is suggesting content that is not in line with the law, and stopping it from being responded</strong></a></p><p><strong>That’s a very practical guideline — technical safeguards are essential.</strong></p><p><strong>Follow-up Question 4: Beyond technical checks, what ethical principles do you think should guide AI development at Microsoft to ensure fairness and accountability?</strong></p><p>(Hint: Think about values like inclusiveness, reliability, transparency, and privacy.)</p><p><strong>Answer </strong><a href="https://m365.cloud.microsoft/chat/entity1-d870f6cd-4aa5-4d42-9626-ab690c041429/eyJpZCI6IlZYTmxjbFl4ZkdoMGRIQnpPaTh2YzNWaWMzUnlZWFJsTFdsdWRDNXZabVpwWTJVdVkyOXRMM3hQU1VRNlpqQXlNMlE1TmprdE0yTmhNaTAwT0RSaExXSmtOR010WWpreFlXVTNZVGc1TldWaWZEbG1NMk0zWVRJMUxUSmlaalV0TkRBM1lTMDVOR1prTFdNd01qUmlNbUppTURGaFpYd3lNREkxTFRBNExUSXhWREU1T2pNek9qVTVMamMzTXpRNU16UmEiLCJzY2VuYXJpbyI6InNoYXJlTGlua1ZpYVBhZ2UiLCJwcm9wZXJ0aWVzIjp7InByb21wdFNvdXJjZSI6InVzZXIiLCJjbGlja1RpbWVzdGFtcCI6IjIwMjUtMDgtMjFUMTk6MzQ6MDAuMjc2WiJ9LCJjaGF0VHlwZSI6IndlYiIsInZlcnNpb24iOjEuMX0"><strong>Ethical principles should include respecting user’s privacy, being transparent on how the LLMs operate with the data, and being inclusive in the data that is used for training the LLM</strong></a></p><p><strong>Excellent! You’ve identified three of the core principles Microsoft emphasizes: privacy, transparency, and inclusiveness.</strong></p><h3>Additional examples of Socratic prompts</h3><p><strong>Structural Prompt Example</strong></p><p>· You are a Socratic study coach. Your goal is to help me deeply understand Microsoft’s Responsible AI initiative using the role-task-format approach: Ask me probing questions, encourage critical thinking, and only provide explanations if necessary. Let’s converse until I can explain the Responsible AI principles in my own words.</p><p><strong>Opening the Dialogue</strong></p><ul><li>“Let’s use the Socratic method: You are a Socratic tutor. Please guide me through understanding the Responsible AI initiative at Microsoft by asking me a series of thought-provoking questions, and only provide explanations if I get stuck.”</li><li>“Act as a Socratic tutor helping me to explore Microsoft’s Responsible AI principles. Begin with an open-ended question and continue the dialogue based on my responses, offering hints or follow-ups instead of direct answers.”</li></ul><p><strong>Example Socratic-Style Questions</strong></p><ul><li>“What are the core values outlined in Microsoft’s Responsible AI framework, and why might each be important for AI systems?”</li><li>“How might fairness and inclusiveness be ensured when designing AI according to Microsoft’s policies?”</li><li>“Why do you think transparency and accountability are included as guiding principles for Responsible AI at Microsoft?”</li><li>“What are some real-world situations where reliability and safety could conflict with other Responsible AI principles, and how should such conflicts be addressed?”</li><li>“How does Microsoft approach privacy and security in their Responsible AI practices, and why are these critical?”</li></ul><p><strong>Probing for Deeper Reflection</strong></p><ul><li>“Could you ask me follow-up questions to challenge my understanding of how Microsoft balances ethical considerations in AI?”</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a3af42d55f23" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond basics — Using powerful GPT-5 specific prompts in M365 Copilot to analyze contracts]]></title>
            <link>https://medium.com/@avipioneer/beyond-basics-using-powerful-gpt-5-specific-prompts-in-m365-copilot-to-analyze-contracts-636e9bf06e54?source=rss-6fb87a57acaa------2</link>
            <guid isPermaLink="false">https://medium.com/p/636e9bf06e54</guid>
            <category><![CDATA[microsoft-copilot]]></category>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Arbindo Chattopadhyay]]></dc:creator>
            <pubDate>Mon, 18 Aug 2025 00:03:27 GMT</pubDate>
            <atom:updated>2025-08-18T00:03:27.118Z</atom:updated>
            <content:encoded><![CDATA[<h3>Beyond basics — Using powerful GPT-5 specific prompts in M365 Copilot to analyze contracts</h3><p>Now that GPT-5 is rolling out to M365 Copilot, learn how you can take advantage of its powerful capabilities. Below you will learn how we use M365 Copilot with GPT-5 to analyze contract</p><p>Let’s look at 4 powerful prompts that take advantage of GPT-5 features in M365 Copilot, such as extended context window, advanced semantic understanding, structured output generation and scenario simulation &amp; role play</p><h3>1. Clause by Clause risk assessment with extended context + structured-output capability</h3><p><strong>Prompt</strong></p><blockquote><em>Analyze the following contract clause-by-clause and identify potential legal, financial, and operational risks. For each risk, explain the impact, likelihood, and possible mitigation strategies, citing relevant legal principles or case law where applicable.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GrRqnm2PtXi_SG1m6vd8Rg.png" /></figure><p>GPT-5 ingests an entire contract and produces a clause-by-clause structured risk report (e.g., JSON objects for clause, risk, impact, likelihood, mitigation).</p><h3>2. Compliance and regulatory check with retrieval augmented grounding / citation aware reasoning</h3><p><strong>Prompt</strong></p><blockquote><em>Review this contract for compliance with California laws and industry regulations. Highlight clauses that may require revision to align with statutory requirements, licensing rules, or common contractual standards</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KmXztJR-mGEMABMpNM8SDQ.png" /></figure><p>When paired with statute/regulation text or a retrieval tool, GPT-5 can flag non-compliant language and link suggested edits to the controlling rule or statute</p><h3>3. Hidden obligations and ambiguities detection with advanced semantic parsing</h3><p><strong>Prompt</strong></p><blockquote>Scan the contract for vague, ambiguous, or overly broad terms that could lead to disputes. For each, suggest precise rewording and explain the legal implications of leaving the language as is</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yW282oPleQ4lUj9PmkA6Hw.png" /></figure><p>GPT-5’s deeper semantic understanding helps it detect subtle ambiguity, identify implied obligations, and propose clearer, legally precise rewordings</p><h3>4. Negotiation Strategy Brief with strategic role play and scenario simulation</h3><p><strong>Prompt</strong></p><blockquote><em>Based on this contract, generate a negotiation brief listing the top five clauses to renegotiate, the reasoning for each, suggested alternative wording, and the likely counterparty pushback points with counter-arguments</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NrafiMgWxPyS46IxToaVdQ.png" /></figure><p>Enables generating a negotiation playbook, simulating likely counterparty responses, and producing prioritized fallback positions / BATNA reasoning.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=636e9bf06e54" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Part III: Grounding Intelligent Content with Copilot Agent Builder]]></title>
            <link>https://medium.com/@avipioneer/part-iii-grounding-intelligent-content-with-copilot-agent-builder-5d9b58e30cbf?source=rss-6fb87a57acaa------2</link>
            <guid isPermaLink="false">https://medium.com/p/5d9b58e30cbf</guid>
            <category><![CDATA[writing-prompts]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Arbindo Chattopadhyay]]></dc:creator>
            <pubDate>Tue, 12 Aug 2025 13:54:33 GMT</pubDate>
            <atom:updated>2025-08-12T13:54:33.109Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Empowering AI Agents with Contextual Awareness for Accurate and Relevant Outputs</em></p><p>This article constitutes the third installment in the prompt engineering series. The <a href="https://medium.com/@avipioneer/prompt-engineering-with-m365-copilot-zero-one-and-multi-shot-prompting-94c2dafdcd6a">first part</a> outlines zero, one, and multi-shot prompting techniques, while the <a href="https://medium.com/@avipioneer/prompt-engineering-part-ii-unlocking-the-power-of-system-prompts-9c6829c95c7e">second examines</a> how system prompts can enhance output by providing structure and tone. In this article, we discuss methods for increasing output relevance by grounding the agent in specific knowledge sources.</p><p>In the era of intelligent automation and AI-driven productivity, the ability to harness organizational knowledge and contextual information has never been more critical. As businesses strive to optimize workflows, deliver personalized customer experiences, and make informed decisions, the challenge often lies not in the lack of data, but in making sense of vast streams of unstructured information scattered across diverse communication channels.</p><p>In the context of preparing a Monthly Business Review that includes customer and field seller feedback on specific products, this article discusses strategies to enhance the relevance of findings by directing agents to consult prioritized content sources rather than the entire suite of M365 resources available to them. By focusing on selected sources, we ensure that the feedback collected is both current and pertinent to our team’s objectives.</p><h3>Understanding Copilot Agent Builder: The Foundation</h3><p>At its core, Copilot Agent Builder is a tool for designing AI agents that blend conversational intelligence with organizational expertise. Unlike generic AI models, a Copilot agent can be configured to “see” and “reference” selected knowledge bases, including chat histories, emails, and other documentation. This ability to ground responses in verifiable content is what differentiates a Copilot agent from a traditional chatbot or digital assistant.</p><p>But what does “grounding” mean in this context? Grounding refers to the process of connecting an AI agent’s output — its answers, recommendations, summaries, or actions — to authentic, real-world data. When the agent is asked a question or given a task, it doesn’t just rely on its pre-trained language model; it actively consults the knowledge sources you specify, ensuring that every response can be traced back to factual content.</p><h3>Why Grounding Matters: Accuracy, Trust, and Compliance</h3><p>In enterprise settings, the stakes are high. A misplaced answer, a hallucinated fact, or a missed nuance can lead to operational errors and erode trust. By grounding agent responses to real exchanges — such as chat transcripts or archived email threads — organizations can:</p><ul><li><strong>Ensure Factual Accuracy:</strong> The agent’s output is verifiable and based on existing knowledge, not speculation.</li><li><strong>Support Regulatory Compliance:</strong> By referencing original communication, organizations can demonstrate due diligence and adherence to industry regulations.</li><li><strong>Enhance Transparency:</strong> Users can audit the source behind each response, fostering confidence in the agent’s recommendations.</li><li><strong>Reduce Risks of Misinformation:</strong> Grounding limits the risk that the agent “hallucinates” or invents facts not present in the knowledge base.</li></ul><h3>Scenario: Grounding content from Monthly Business Review</h3><p>In our last article, we created an agent named Yoda. We made it smarter by providing instructions, which are system prompts in LLM. Purpose of Yoda is to gather feedback provided by customers and field sellers on specific topic, provided in the user prompt.</p><p>Yoda currently scans all M365 content sources where the user has at least read permissions. These sources include SharePoint sites, OneDrive, documents accessible to the user in other OneDrive accounts, Teams channels and chats, and emails. Due to the substantial size of the content source, the feedback collected may not always be precise or closely related.</p><p>Our team has specific group chats and channels in Teams where we discuss feedback within our team members. What if we can direct the agent to look at these sources first?</p><h3>Connecting Knowledge Sources</h3><p>The first step in using Copilot Agent Builder effectively is to connect it to the relevant knowledge sources.</p><p>We start by editing our agent, Yoda. In the Configure tab, find Knowledge section, and add My emails, and specific Teams chats. We can now select specific Teams chats and channels.</p><p>You will notice that we have enabled “<strong><em>Prioritize the knowledge sources you added for agent knowledge-based queries”</em></strong>. This switch tells the agent to prioritize the selected sources first.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/1*W1XnhOms4wdmKHIM-l0tfg.png" /></figure><p>Next, click on the Update button on the top right corner to save the changes to the agent.</p><h3>Yoda in action</h3><p>Now that we have an intelligent and shall I say, a very smart agent Yoda, lets put it to work.</p><p>Yoda can be invoked from the M365 Copilot app. It will appear under the Agents section in the app. Once selected, you will be presented with the familiar chat window where you can write your prompts. The prompt I wrote to test it out is “<strong><em>Please provide feedback for &lt;&lt;product / service&gt;&gt;.”. </em></strong>In this case, the prompt is short because we have provided much richer instructions in the agent.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/1*sovh_ndLH8TFflPHI145Bg.png" /></figure><p>The other option would be invoking Yoda from Microsoft Word or PowerPoint. This is useful because the output from Yoda is directly inserted into the Word document or presentation.</p><h3>Other Real-World Use Cases</h3><h3>1. Customer Support and Case Resolution</h3><p>When handling customer inquiries, agents grounded in chat and email histories can instantly surface prior conversations, warranty claims, or troubleshooting steps. Instead of asking customers to repeat themselves, the agent provides context-aware responses, reducing frustration and accelerating issue resolution.</p><h3>2. Project Management and Team Collaboration</h3><p>Complex projects generate countless threads across chats and emails. A Copilot agent can act as a knowledge concierge — summarizing key decisions, tracking commitments, and even reminding team members of agreed-upon actions. Because every summary or recommendation is grounded in actual discussions, teams trust the agent’s guidance.</p><h3>3. Sales Enablement and Client Onboarding</h3><p>Sales and client success teams rely on accurate, up-to-date information. A grounded agent can reference past proposals, negotiation points, and feedback captured in email or chat, ensuring new offers are consistent and personalized — and that onboarding checklists never miss a step.</p><h3>4. Compliance and Audit Readiness</h3><p>For industries like finance, health, or legal services, audit trails are non-negotiable. Copilot agents can be configured to document the precise source of any information shared, making it easy to respond to compliance reviews or regulatory inquiries.</p><h3>Conclusion: Building Trust Through Grounded AI</h3><p>As organizations look to AI to streamline operations and enhance decision-making, trust remains the cornerstone of successful adoption. Copilot Agent Builder’s grounding capabilities empower teams to build intelligent agents that not only “know” but can “show their work” by referencing the exact conversations and communications that shaped their responses.</p><p>By rooting AI outputs in real, traceable knowledge — from chats and emails to broader documentation — businesses unlock new levels of efficiency, transparency, and reliability. Whether you’re supporting customers, collaborating across teams, or managing compliance, Copilot Agent Builder puts trustworthy, contextually-aware AI at your fingertips.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5d9b58e30cbf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Prompt Engineering Part II: Unlocking the Power of System Prompts]]></title>
            <link>https://medium.com/@avipioneer/prompt-engineering-part-ii-unlocking-the-power-of-system-prompts-9c6829c95c7e?source=rss-6fb87a57acaa------2</link>
            <guid isPermaLink="false">https://medium.com/p/9c6829c95c7e</guid>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[microsoft-copilot]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[writing-prompts]]></category>
            <dc:creator><![CDATA[Arbindo Chattopadhyay]]></dc:creator>
            <pubDate>Fri, 08 Aug 2025 17:03:42 GMT</pubDate>
            <atom:updated>2025-08-08T17:04:13.508Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A Deeper Dive in Advancing Your Prompting Strategy for M365 Copilot</em></p><p>As we journey deeper into the world of effective prompting, having explored <a href="https://medium.com/@avipioneer/prompt-engineering-with-m365-copilot-zero-one-and-multi-shot-prompting-94c2dafdcd6a">zero, one, and multi-shot strategies</a>, it’s time to unveil one of the most powerful — yet often invisible — tools in the AI practitioner’s toolkit: the system prompt. Understanding and harnessing system prompts can be the key to extracting more consistent, reliable, and context-aware responses from M365 Copilot and other generative AI models.</p><h3>What is System Prompt?</h3><p>A system prompt is an instruction or set of instructions embedded within the initial context given to a generative AI model — prior to any user interaction. Unlike user prompts, which are explicit requests or questions posed by the end-user, system prompts work behind the scenes to define the AI’s overarching behavior, role, and boundaries throughout interaction. You can think of them as the “stage directions” given to an actor before the curtain rises, guiding the performance in subtle but significant ways.</p><p>System prompts might specify tone, persona, or ethical guidelines; they may restrict outputs to certain formats or provide background knowledge that will persist throughout a session. In many large language model (LLM) deployments — including M365 Copilot — system prompts are fundamental in shaping the AI’s responses to meet business requirements, compliance standards, or user expectations.</p><h3>System Prompts vs. User Prompts: How Are They Different?</h3><p>While both system and user prompts influence model output, their intent, timing, and visibility are distinct:</p><p>· Intent: System prompts establish foundational parameters, like instructing Copilot to always be helpful, concise, and professional. User prompts are the operational queries or tasks issued during a session — such as asking Copilot to draft a report or summarize an email thread.</p><p>· Timing: System prompts are set at the initiation of an AI session or embedded within the application’s backend. User prompts occur dynamically, whenever a user interacts with the model.</p><p>· Visibility: System prompts are generally invisible to the end-user, while user prompts are the direct interface between the user and the AI.</p><p>To illustrate, imagine M365 Copilot has a system prompt guiding it to “act as an expert business assistant, always respond in a formal tone, cite data when available, and never provide personal opinions.” Regardless of the user prompt — say, “Summarize the latest sales report” — these guiding principles silently inform the AI’s approach, ensuring consistency and compliance.</p><h3>Scenario — Monthly Business Review</h3><p>The <a href="https://medium.com/@avipioneer/prompt-engineering-with-m365-copilot-zero-one-and-multi-shot-prompting-94c2dafdcd6a">previous article</a> discussed the use of multi-shot prompting to structure insights for a monthly business review. In summary, the scenario involves creating a draft of the monthly business review with product engineering and marketing leaders. The document includes feedback from customers and field sellers intended to inform the product roadmap and business models.</p><p>As demonstrated, multi-shot prompting enables structured Copilot responses by incorporating examples directly within the prompt. While this approach is effective, a key consideration is how to automate the process so that detailed instructions with examples are not required each time a prompt is composed in Copilot.</p><p>Welcome, Agents. Agents assist in configuring Copilot to serve a specific purpose. In this context, the objective is to efficiently organize the output for the Monthly Business Review.</p><p>We will start by creating an agent, Yoda, in the agent builder in M365 Copilot. After naming and providing description for the agent, we will do the most important task — which is to provide instructions to the agent. The instructions in the agent is essentially the system prompt in LLM.</p><p>Following will be our system prompt aka instructions for the Monthly Business Review:</p><blockquote><strong><em>You are an agent that acts like a program manager in a large technology company. Your task is to synthesize information and put together a list of feedback from customers and field sellers to help influence the engineering roadmap and marketing priorities. Information should be recent, impactful and include customer name or the field team that provided that feedback.</em></strong></blockquote><blockquote><strong><em>Group the output into 2 categories — Customer insights and Field insights. Consolidate similar feedback into one. For each feedback, format the items into bulleted list as follows:</em></strong></blockquote><blockquote><strong><em>&lt;&lt;Feedback theme&gt;&gt;- &lt;&lt;Details on the feedback&gt;&gt;. Customers include Contoso, Woodgrove and Fabrikam</em></strong></blockquote><blockquote><strong><em>Always greet the user with ‘My name is Yoda. Please let me know the topic from which you would want to gather insights’</em></strong></blockquote><p>With this instruction, we are telling the agent</p><p>1. Act like a program manager</p><p>2. What is it’s purpose</p><p>3. What kind of information it should gather</p><p>4. How the information should be structured. We also provide example for the response structure</p><p>5. How the information is grouped</p><p>Because this instruction is a system prompt, user will not need to repeat these instructions every time they write a prompt.</p><h3>When and Why Use System Prompts?</h3><p>System prompts are invaluable in scenarios where AI reliability, safety, and brand alignment are non-negotiable. Here are some practical scenarios:</p><p>· Ensuring Consistent Persona and Tone: In enterprise environments, AI responses should reflect company voice. A system prompt can enforce formality or friendliness, as required.</p><p>· Maintaining Compliance and Guardrails: When legal, ethical, or regulatory considerations apply, system prompts can restrict sensitive topics, prevent disclosure of private information, or require disclaimers.</p><p>· Driving Task-specific Performance: For applications like M365 Copilot, a system prompt can instruct the model to prioritize actionable insights, defer to organizational data sources, or avoid unsupported speculation — all before user queries begin.</p><p>· Enhancing Context Retention: Unlike user prompts, which are ephemeral, system prompts persist. This helps maintain context across longer interactions, reducing the need for repeated instructions.</p><h3>Best Practices for System Prompt Design</h3><p>To maximize the effectiveness of system prompts within M365 Copilot or any LLM deployment, consider these principles:</p><p>· Clarity and Specificity: Clearly define the AI’s role, tone, and boundaries, using unambiguous language.</p><p>· Alignment with Business Goals: Tailor system prompts to reflect company values, compliance requirements, and user expectations.</p><p>· Iterative Optimization: Test, monitor, and refine system prompts based on user feedback and observed model behavior.</p><p>· Complement, Don’t Replace, User Prompts: System prompts set the stage, but user prompts drive the action. Design both to work in harmony for best results.</p><h3>Conclusion</h3><p>System prompts represent the silent backbone of successful AI deployments, playing a critical role alongside user-facing prompting strategies. By thoughtfully crafting system prompts for M365 Copilot and other generative models, you ensure your AI not only understands what to do — but how to do it, every single time.</p><p>As you build on your prompting expertise, don’t overlook the value of these unseen instructions. They are your first and last line of defense for consistency, compliance, and quality in every AI-powered interaction.</p><p>Continue experimenting, learning, and refining — because in the world of generative AI, a well-designed prompt is where power, precision, and possibility converge.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9c6829c95c7e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Prompt engineering with M365 Copilot -Zero, One and Multi shot prompting]]></title>
            <link>https://medium.com/@avipioneer/prompt-engineering-with-m365-copilot-zero-one-and-multi-shot-prompting-94c2dafdcd6a?source=rss-6fb87a57acaa------2</link>
            <guid isPermaLink="false">https://medium.com/p/94c2dafdcd6a</guid>
            <category><![CDATA[writing-prompts]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[m365-copilot]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <dc:creator><![CDATA[Arbindo Chattopadhyay]]></dc:creator>
            <pubDate>Thu, 07 Aug 2025 14:44:18 GMT</pubDate>
            <atom:updated>2025-08-07T14:44:18.244Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Unlocking the Power of Generative AI through Effective Prompting Strategies</em></p><h3>Introduction</h3><p>As artificial intelligence continues to evolve, so too do the ways in which humans interact with these advanced systems. For those beginning their AI journey — particularly with generative models such as language-based AI — understanding how to effectively communicate with these systems is critical. This communication takes shape in the form of <strong>prompting</strong>, a technique that guides the AI’s output by providing it with specific instructions or data. Three foundational prompting strategies dominate this landscape: <strong>zero-shot</strong>, <strong>one-shot</strong>, and <strong>multi-shot</strong> prompting.</p><p>In this blog post, we’ll demystify these concepts, provide practical examples, and outline their differences. By the end, you’ll be equipped with the knowledge to craft prompts that elicit precise, relevant, and contextually rich responses from generative AI models.</p><h3>What is Prompting?</h3><p>Prompting is the method of providing input or context to an AI model to guide its output. The quality, specificity, and structure of your prompt can significantly influence the AI’s response, making prompting both an art and a science. Beginners often overlook this step, but mastering prompting can dramatically enhance the accuracy and usefulness of AI-generated results.</p><h3>Zero, One and Multi shot prompting</h3><p>Zero, One and Multi shot prompting allows you to customize the output of the response from Copilot. It helps not only structure the response but also guide Copilot by providing examples.</p><p>One of the monthly activities for me and my team is to put together Monthly Business Review (MBR) that gets presented to our ENG and Marketing stakeholders. Prior to Copilot, this exercise used to take us few meetings to gather, refine and format the insights. Now, Copilot does the initial groundwork that saves us several hours. I will walk you through MBR drafting scenario with zero, one and multi shot prompting.</p><h3>Zero-Shot Prompting</h3><h3>Definition</h3><p>Zero-shot prompting involves asking the AI to perform a task without providing any explicit examples. The AI must rely solely on its pre-existing training and general knowledge to understand and complete the task.</p><h3>Scenario</h3><p>Imagine you want an AI to generate list of top 5 insights with regard to M365 Copilot business. In zero-shot prompting, you simply make the request without any examples:</p><p><em>Prompt:</em></p><p><em>Act as a program manager responsible for driving adoption of M365 Copilot. Summarize top 5 insights you can provide for engineering and marketing stakeholders to help improve M365 Copilot. Refer to my emails, notes, Teams chats and meeting transcripts for information. Format the content into a bullet list. Keep the tone professional. Your audience include executive decision makers</em></p><p>In the example prompt above, I have a prompt that tells Copilot its role, task, format, tone and audience. For most of your scenarios, you’d be using something like this prompt for your task.</p><h3>Benefits and Limitations</h3><ul><li><strong>Benefits:</strong> Fast, requires minimal setup, and useful for straightforward tasks.</li><li><strong>Limitations:</strong> May produce less accurate or inconsistent results for complex or nuanced queries.</li></ul><h3>One-Shot Prompting</h3><h3>Definition</h3><p>One-shot prompting provides the AI with a single example of the desired input-output pattern. This helps orient the model, especially when the task is slightly ambiguous or context-dependent.</p><h3>Scenario</h3><p>We will expand the same scenario as we have in zero shot prompting, and we will include an example of how Copilot should include additional insight in the outcome.</p><p><em>Prompt:</em></p><p><em>Act as a program manager responsible for driving adoption of M365 Copilot. Summarize the top 5 insights you can provide for engineering and marketing stakeholders to help improve M365 Copilot.</em></p><p><strong><em>For each insight, include list of customers impacted. For instance — “Prioritize role-based enablement to drive usage intensity — Contoso, Fabrikam”.</em></strong></p><p><em>Use Markdown for clarity. Keep the tone professional. Your audience includes executive decision makers. Ensure the insights are actionable and impactful, highlighting key areas for improvement and strategic focus.</em></p><p>In this prompt, I am asking Copilot to include references of customers who have provided this insight and also guiding Copilot to structure the output in specific format. This is one-shot prompting because there is one example.</p><h3>Benefits and Limitations</h3><ul><li><strong>Benefits:</strong> Clarifies the task, provides a reference point, and improves the relevance of responses with minimal effort.</li><li><strong>Limitations:</strong> If the example is unclear or poorly chosen, the AI may follow an incorrect pattern; limited scope for highly varied tasks.</li></ul><h3>Multi-Shot Prompting</h3><h3>Definition</h3><p>Multi-shot prompting involves providing multiple examples of the task at hand. This is particularly useful for more complex or nuanced requests, as it helps establish a clear pattern for the AI to follow.</p><h3>Scenario</h3><p>We will further expand our scenario by including not only customer, but also field and partner enablement insights. We will provide multiple examples of how the output needs to be structured.</p><p><em>Prompt:</em></p><blockquote><em>Act as a program manager responsible for driving adoption of M365 Copilot. Summarize the top 5 insights you can provide for engineering and marketing stakeholders to help improve M365 Copilot.</em></blockquote><blockquote><strong><em>Categorize each insight by Customer Insights, Field enablement and Partner enablement. For example –</em></strong></blockquote><blockquote><strong><em>Customer insights -”Need role based adoption guides to drive usage”</em></strong></blockquote><blockquote><strong><em>Field enablement — “Clarity on incentive structure and seller compensation”</em></strong></blockquote><blockquote><strong><em>Partner Enablement — “Ability to attach partner specific solutions”.</em></strong></blockquote><blockquote><em>Use Markdown for clarity. Keep the tone professional. Your audience includes executive decision makers. Ensure the insights are actionable and impactful, highlighting key areas for improvement and strategic focus.</em></blockquote><h3>Benefits and Limitations</h3><ul><li><strong>Benefits:</strong> Greatly enhances output accuracy for complex tasks; helps the AI better understand nuanced requirements.</li><li><strong>Limitations:</strong> Requires more effort to craft effective examples; lengthy prompts may exceed input limits in some systems.</li></ul><h3>Choosing the Right Prompting Strategy</h3><p>Selecting the appropriate prompting technique depends on the complexity of the task, the desired format, and the need for consistency. Here are some guidelines to help you choose:</p><ul><li>Use <strong>zero-shot prompting</strong> for simple, well-understood tasks where AI’s general knowledge suffices.</li><li>Choose <strong>one-shot prompting</strong> when your task requires a specific style, tone, or structure and a single example can clarify the requirement.</li><li>Opt for <strong>multi-shot prompting</strong> when the task is complex, nuanced, or varied, and you want to establish a clear and consistent pattern for the AI to follow.</li></ul><h3>Best Practices for Effective Prompting</h3><ul><li><strong>Be Clear and Concise:</strong> Ambiguous prompts may lead to off-target responses. Always state your intent plainly.</li><li><strong>Use Relevant Examples:</strong> Especially for one-shot and multi-shot prompts, ensure your examples are highly relevant to the desired output.</li><li><strong>Test and Iterate:</strong> Don’t hesitate to refine your prompts based on the AI’s responses. Iterative prompting often leads to better outcomes.</li><li><strong>Stay Within Limits:</strong> Be mindful of the model’s input restrictions — overly long prompts may be truncated.</li></ul><h3>Conclusion</h3><p>Prompting is a foundational skill when working with generative AI. Whether you’re a beginner experimenting with zero-shot queries or an advanced user crafting multi-shot prompts for highly specific tasks, understanding these strategies will empower you to extract more value from AI systems. Practice each approach, observe the differences in output, and continue refining your technique — the world of AI is vast, and your ability to communicate with it is the key to unlocking its true potential.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=94c2dafdcd6a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>