<?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 Praful Krishna on Medium]]></title>
        <description><![CDATA[Stories by Praful Krishna on Medium]]></description>
        <link>https://medium.com/@praful-krishna?source=rss-2ae1c32a010------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*CCtz-5GwOhuE8fahPvg4FQ.jpeg</url>
            <title>Stories by Praful Krishna on Medium</title>
            <link>https://medium.com/@praful-krishna?source=rss-2ae1c32a010------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 20:01:17 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@praful-krishna/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[The Architecture of Reliable Enterprise Intelligence]]></title>
            <link>https://praful-krishna.medium.com/the-architecture-of-reliable-enterprise-intelligence-ffa2f77347e5?source=rss-2ae1c32a010------2</link>
            <guid isPermaLink="false">https://medium.com/p/ffa2f77347e5</guid>
            <category><![CDATA[enterprise-architecture]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[pharmaceutical]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[finance]]></category>
            <dc:creator><![CDATA[Praful Krishna]]></dc:creator>
            <pubDate>Wed, 21 Oct 2026 22:55:03 GMT</pubDate>
            <atom:updated>2026-10-21T22:55:03.292Z</atom:updated>
            <content:encoded><![CDATA[<h4>Models provide the raw horsepower. An enterprise harness gives that intelligence somewhere useful to operate — and the discipline to know when to stop.</h4><p>Consider a hypothetical but common enterprise problem: a pharmaceutical manufacturing team asks an AI system to assess the regulatory and operational impact of a process change. A capable foundation model can review batch records, scan standard operating procedures, extract relevant chemistry and controls data, and draft an articulate assessment. It may even cite its sources, producing a document that looks convincing in a demonstration. That is the easy part.</p><p>The difficult part begins when the records do not support a single, complete account of what should happen next. Two internal documents describe conflicting versions of the same process, and a critical stability study is missing. One unit conversion calls for a tested calculation, while an unexpected impurity profile demands scientific judgment. Some workstreams can proceed immediately; another depends on an authorized Quality reviewer, requiring coordination of evidence, execution, and authority. A better model can help identify these problems, provided it has access to authoritative records, appropriate tools, and explicit decision rights.</p><p>An agent harness earns its place by organizing those conditions into a working system. Its purpose is <strong>to design the operational environment in which non-deterministic intelligence can perform reliable, governed work.</strong> Domain expertise must shape the plan, the execution methods, and the criteria for accepting results. This is where a harness can turn an articulate assessment into work that supports a defensible decision.</p><h3>1. Intelligence Needs an Operating Environment</h3><p>At its simplest, a harness provides an execution loop: call the model, route its tool requests, return the results, and continue. That loop is a useful foundation, but enterprise work also involves persistent state, changing assumptions, permissions, and explicit human sign-offs. If a process stops halfway through, the system must know where it left off and which earlier conclusions remain valid. It must also establish whether a requested action actually succeeded, rather than treating an attempted operation as a completed one.</p><p>For enterprise work, the architecture considered here organizes that environment around four stages:</p><ol><li><strong>Context Hydration</strong> establishes the relevant situation, evidence, and assumptions.</li><li><strong>Goal Decomposition</strong> identifies the tasks needed to satisfy the objective.</li><li><strong>Governance Specification</strong> defines three parts: <strong>guardrails</strong> that set permitted boundaries; a <strong>verification framework</strong> that checks individual outputs and actions; and <strong>evals</strong> that assess system behavior across representative cases.</li><li><strong>Iterative Orchestration</strong> determines the next action from the latest results.</li></ol><p>Together, these stages connect the original request to an executable plan whose progress can be inspected.</p><p>The four stages establish an initial basis for work while allowing that basis to be revised. When a missing study becomes apparent, the plan may need to change; when a reviewer narrows the scope, the applicable constraints may change with it. Those changes should remain traceable to the evidence or decision that prompted them. A harness that preserves the original plan at all costs can execute the wrong work very efficiently.</p><h3>2. Domain Expertise Must Change the Plan, Not Just the Prose</h3><p>A system may retrieve the correct documents through RAG and still investigate the wrong questions. A domain-informed harness uses that knowledge to shape the structure of the work. The distinction becomes visible in the tasks the system creates and the evidence it requires.</p><p>Consider our pharmaceutical example through the lens of chemistry, manufacturing, and controls, commonly called CMC. A generic planner might interpret the assignment as a side-by-side comparison of documents followed by an executive summary. A domain-informed planner would consider which approved baseline applies, which dossier dependencies may be affected, whether analytical comparisons are appropriate, and which scientific questions remain unresolved. The tasks depend on the change and its scope, but the difference is visible before either system writes its assessment.</p><p>In this architecture, Domain Depth feeds three stages: Context Hydration, Goal Decomposition, and Governance Specification. It helps establish which records matter, which tasks belong in the plan, and which constraints and acceptance criteria apply. Retrieval can support each stage, provided the retrieved knowledge influences the corresponding decisions. A document introduced only during final drafting may improve the explanation; introduced during planning, it may reveal that an entire workstream is missing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WKd4O_mEwWBzEn3k3kq3gQ.png" /><figcaption>Overview of the Building Blocks of a Reliable Enteprise Harness.</figcaption></figure><p>This yields a useful question for an enterprise AI implementation: <strong>when relevant domain knowledge changes, does the system reconsider its plan?</strong> The answer need not always be yes, but the system should assess the implications. If domain knowledge changes only the wording, its contribution to execution remains limited. An expert vocabulary can make an incomplete investigation sound remarkably complete.</p><p>Deeper integration also creates a maintenance obligation. Regulations change, local standards diverge, and internal procedures are revised. The harness therefore needs to preserve the source, version, and applicability of knowledge wherever those distinctions affect the work. Otherwise, domain depth can distribute the same stale assumption across context, planning, and governance.</p><h3>3. Match the Execution Method to the Required Accuracy</h3><p>Once the system has a plan, it must choose how to perform each task reliably. Sending every task to a large language model overlooks the different sources of error in calculation, rule application, extraction, and interpretation. A dependable harness selects an execution method whose behavior can be checked against the task’s requirements. The organizing principle is accuracy and repeatability, with cost and latency as additional considerations.</p><p>The architecture supports five execution mechanisms:</p><ol><li><strong>Deterministic scripts:</strong> For explicit calculations, transformations, and repeatable operations with defined inputs and checks.</li><li><strong>Decision models:</strong> For applying codified decision logic to a specified set of conditions.</li><li><strong>Prescriptive recipes:</strong> For carrying out established procedures with known steps and dependencies.</li><li><strong>Small language models (SLMs):</strong> For bounded language tasks within a tested scope.</li><li><strong>Large language models (LLMs):</strong> For broader interpretation, synthesis, and contextual reasoning.</li></ol><p>These mechanisms can work together: a recipe may invoke a script or a language model, and the list represents neither exclusive alternatives nor a fixed escalation ladder. In the pharmaceutical assessment, a unit conversion calls for a tested calculation with explicit handling of precision and rounding. Comparing technical justifications may call for a language model, while conflicting analytical evidence may require further investigation and expert review. The choice depends on each task’s failure modes and the evidence required to accept its result.</p><p>Determinism does not guarantee correctness: a script can apply the wrong conversion factor with perfect consistency. A decision model can encode an outdated rule, and a prescriptive recipe can omit an important dependency. Language models also need evaluation on the tasks assigned to them, rather than being trusted on the strength of general capability. Reliability comes from matching the method, its scope, and its verification requirements, then checking the actual result.</p><h3>4. Work Delegation Demands an Unambiguous Owner of Control</h3><p>The orchestrator owns the workflow state and subsequent action. It dispatches work, inspects results, and decides whether to continue, retry, switch methods, request information, or involve a human. An execution method can make decisions within its assignment without acquiring authority over the entire objective.</p><p>Every execution method returns one of three states: <strong>Success, Failure, or Needs Input</strong>. The result carries the evidence, artifacts, or diagnostic information needed to interpret its status. All results return to Iterative Orchestration, which decides how to handle them; the communication pathways simply carry the information.</p><p>The payload behind each state gives the contract its operational meaning:</p><ul><li><strong>Success</strong> includes artifacts and evidence that the assigned acceptance criteria have been met.</li><li><strong>Failure</strong> includes diagnostic information sufficient to distinguish, for example, a temporary service outage from an invalid operation.</li><li><strong>Needs Input</strong> identifies the missing evidence, clarification, or authorization and explains which work depends on it.</li></ul><p>A temporary outage may justify a bounded retry, while a missing validation report requires a different response. Repeatedly asking a model to finish the assessment will not supply the absent evidence. The orchestrator should preserve completed work, hold the affected conclusions or actions, and continue independent tasks where appropriate. It can then seek the missing input through another authorized source or present a focused request to a human.</p><p>Success applies to the assigned task; it does not by itself establish the validity of a scientific conclusion or authorize a downstream action. A calculation can succeed while the overall assessment remains blocked by missing evidence. The orchestrator must preserve that distinction when interpreting results, and its own routing and recovery decisions require evaluation.</p><h3>5. Design Human Involvement Around Intent and Attention</h3><p>“Human in the loop” leaves a central question unanswered: what judgment is the person being asked to provide? The <strong>Design Center</strong> establishes intent, scope, relevant sources, and authorization thresholds, including the governance requirements defined earlier. The <strong>Control Center</strong> supports supervision during execution, presenting consequential questions to people with the appropriate authority. Defining the work and resolving an exception require different information, so each interface should serve its particular decision.</p><p>The Control Center also performs <strong>attention triage</strong>. Even a diligent, deeply invested professional has limited attention to distribute across a process. The harness must identify where that attention is most critical, considering the consequences of error, unresolved uncertainty, dependencies, and decisions reserved for human authority. An undifferentiated review queue makes the most consequential decisions harder to identify.</p><p>Attention triage preserves the standards for acceptable work, and mandatory review remains mandatory regardless of queue priority. If an output lacks required evidence or fails its quality checks, the affected work stays on hold even when a reviewer is unavailable. Limited attention cannot authorize an unsupported conclusion. Any authorized change to scope or requirements must be explicit and recorded.</p><p>For the missing stability study, the Control Center should identify the affected conclusions and present the available decisions: provide the study, commission further work, or narrow the scope. The reviewer needs the issue, its consequences, and supporting evidence, with access to the underlying record. A long transcript may support investigation, but it is a poor default decision interface. The harness should make human attention more effective while keeping the unresolved dependency visible.</p><p>***</p><p>Return to the pharma manufacturing team considering the process change. The assessment becomes consequential when someone uses it to authorize further work, request a study, or make an operational decision. At that point, scientific discipline must extend through the action itself: the evidence must support the conclusion, required checks must have passed, and the decision must remain within the available authority. <strong>The real test of enterprise intelligence is whether its conclusions remain defensible when someone has to act on them.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ffa2f77347e5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Seven Ways Generative AI is Quietly Revolutionizing Data Analytics]]></title>
            <link>https://praful-krishna.medium.com/seven-ways-generative-ai-is-quietly-revolutionizing-data-analytics-05705b4cec12?source=rss-2ae1c32a010------2</link>
            <guid isPermaLink="false">https://medium.com/p/05705b4cec12</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Praful Krishna]]></dc:creator>
            <pubDate>Fri, 27 Jun 2025 03:47:09 GMT</pubDate>
            <atom:updated>2025-06-27T03:47:09.935Z</atom:updated>
            <content:encoded><![CDATA[<h4>Pragmatic CEOs and business leaders are leveraging generative AI to power non-generative AI use cases as well. They have had to rely on traditional ML so far.</h4><p><em>This is the fifth in a </em><a href="https://praful-krishna.medium.com/pragmatic-playbook-for-enterprise-ai-b54419af9c98"><em>series</em></a><em> of posts meant to help CEOs and business leaders adopt Artificial Intelligence at enterprises. The series tries to cut through the hype and focuses on ROI. These are the author’s personal views (</em><a href="https://medium.com/@praful-krishna/best-practices-for-building-an-enterprise-ai-platform-62c6fc0758a6"><em>Previous post</em></a><em>).</em></p><p>When we hear “Generative AI,” our minds often conjure images of chatbots writing poetry, algorithms composing music, or AI creating stunning digital art from a simple text prompt. The focus has largely been on its power to create unstructured, creative content. But to see generative AI solely through this lens is to miss one of the most significant technological shifts on the horizon — one that is set to fundamentally redefine the highly structured world of data analytics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fX03Nbj311hD47xVBkxL2w.png" /><figcaption>Generated by Gemini using Author’s prompts</figcaption></figure><p>For decades, the analytics space has been on a steady, incremental journey. We’ve moved from asking “What happened?” (descriptive analytics) to “Why did it happen?” (diagnostic), and more recently, “What is likely to happen?” (predictive). The holy grail has always been prescriptive analytics: “What is the best course of action?” While we’ve made strides, achieving this has often required legions of data scientists, complex software, and a deep understanding of statistical modeling. The process has been powerful, but rarely intuitive or accessible to the average business user.</p><p>Generative AI is about to change all of that. It’s not just another step in the journey; it’s a series of massive leaps that will reshape how we interact with, understand, and leverage data. By bridging the gap between human language and machine logic, generative AI will democratize data insights and automate complex processes that were once the exclusive domain of specialists.</p><h4>1. The End of Dashboards as We Know Them: The Rise of Conversational UX</h4><p>For years, the dashboard has been the primary window into our data. While useful, dashboards are static. They are pre-configured to answer a specific set of questions. If you have a follow-up query or want to explore a different angle, you often have to go back to an analyst to modify the report.</p><p>Generative AI is replacing the rigid dashboard with a dynamic, conversational interface. Imagine being able to simply ask your questions in plain English: “Show me our sales in the Southeast region for the last quarter, but exclude returns and compare our top three products’ performance against the same period last year.” The AI doesn’t just spit out a table of numbers; it would generate a comprehensive, multi-faceted report with charts, summaries, and key insights, all in a matter of seconds. This intuitive user experience for querying, reporting, and analytics means that anyone, from a CEO to a marketing intern, can have a direct conversation with their data.</p><h4>2. From Prediction to Simulation: Advanced Reasoning and Scenario Planning</h4><p>Predictive analytics has been a game-changer, but it often tells you what is <em>likely</em> to happen based on historical data. Generative AI takes this a step further by enabling advanced reasoning and scenario planning. Instead of just one prediction, you can model multiple futures.</p><p>A supply chain manager could ask, “What would be the impact on our delivery times and costs if our primary shipping partner in Asia experiences a two-week delay? Model three potential responses and recommend the optimal one.” The AI could then run thousands of simulations, factoring in countless variables — alternative routes, secondary logistics partners, inventory levels — and present a prescriptive plan. This moves analytics from a reactive tool to a strategic, proactive partner in decision-making.</p><h4>3. Automated Integrations: Speaking the Language of Data</h4><p>One of the most time-consuming and technically challenging aspects of data analytics is integration. Getting different systems, with their unique data formats and specifications, to talk to each other is a major hurdle.</p><p>Generative AI can act as a universal translator. By understanding the specifications of different data sources and the desired outcome, it can automatically generate the code and logic required to build integration pipelines. An analyst could simply provide the documentation for two different APIs and state, “I need to combine customer data from our CRM with purchase history from our e-commerce platform.” The AI would handle the complex mapping, transformation, and integration, turning a multi-week engineering project into a task of a few minutes.</p><h4>4. Taming the Chaos: Automated Normalization of Data Streams</h4><p>Data rarely arrives in a clean, perfectly structured format. Real-world data streams are often messy, inconsistent, and full of anomalies. Normalizing this data — getting it into a standardized, usable format — is a foundational step that consumes a significant portion of an analyst’s time.</p><p>Generative AI excels at pattern recognition and can learn the “ideal” structure of your data. It can automatically identify and correct inconsistencies, standardize formats (like dates or addresses), and flag outliers that require human attention. This automated normalization ensures that the data feeding into your analytics engine is clean and reliable, leading to more accurate and trustworthy insights.</p><h4>5. The Analyst’s Co-pilot: Accelerated Modeling</h4><p>Building operational models, such as digital twins of a supply chain or a manufacturing plant, is both an art and a science. It requires deep domain expertise in physical processes and simulation. Generative AI can act as a “co-pilot” for engineers and operations analysts, dramatically accelerating this process.</p><p>An operations manager could describe a physical system they want to model — “I need a digital twin of our main distribution center, including inventory flow, staffing levels, and vehicle movements.” The AI could then suggest appropriate simulation frameworks, generate the initial code for the model, and even create preliminary visualizations based on historical operational data. This partnership allows the human expert to focus on validating and refining the digital twin against real-world behavior, rather than getting bogged down in the mechanics of coding the simulation from scratch.</p><h4>6. Uncovering Connections: Accelerated Detection of Data Dependencies</h4><p>In complex organizations, data is intricately interconnected in ways that are not always obvious. Understanding these dependencies — how a change in one dataset can ripple through and affect others — is crucial for maintaining data integrity and building accurate models. Generative AI can analyze the entire data landscape of an organization, from structural asset data to day-to-day transactional data, and automatically map these dependencies. This creates a “digital twin” of the operational data model, providing a clear and comprehensive view of how everything is connected. This is invaluable for impact analysis, change management, and building more robust and resilient data systems.</p><h4>7. Black Box No More: Complete Transparency and Human Control</h4><p>A common fear with AI is that it becomes a “black box” where decisions are made without clear justification. True enterprise-grade generative AI for analytics must be built on a foundation of transparency and human control. Because the AI is generating the code, reports, and models, it can also explain its work. Every step, from data integration to the final recommendation, can be audited. A user can ask, “Why did you recommend this course of action?” and the AI can trace its logic back through the entire analytical process. This ensures that humans remain firmly in the driver’s seat, able to understand, question, and ultimately override the AI’s suggestions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-C_nnjvETDldZ7huA9IPTQ.png" /></figure><p>The future of analytics is not about replacing humans with machines. It’s about empowering humans with more intelligent, intuitive, and powerful tools. By taking on the heavy lifting of data integration, normalization, and modeling, and by providing a simple, conversational interface for exploration and analysis, generative AI will unlock the true potential of data for everyone. The era of wrestling with rigid dashboards and complex software is coming to an end. The era of the data conversation is just beginning.</p><p><em>Complete Series:</em></p><p><a href="https://praful-krishna.medium.com/pragmatic-playbook-for-enterprise-ai-b54419af9c98">Pragmatic Playbook for Enterprise AI</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=05705b4cec12" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Best Practices for Building an Enterprise AI Platform]]></title>
            <link>https://praful-krishna.medium.com/best-practices-for-building-an-enterprise-ai-platform-62c6fc0758a6?source=rss-2ae1c32a010------2</link>
            <guid isPermaLink="false">https://medium.com/p/62c6fc0758a6</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[platform]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[engineering]]></category>
            <dc:creator><![CDATA[Praful Krishna]]></dc:creator>
            <pubDate>Sat, 03 May 2025 21:23:15 GMT</pubDate>
            <atom:updated>2025-05-03T21:23:15.990Z</atom:updated>
            <content:encoded><![CDATA[<h4>Pragmatic CEOs and business leaders work with their engineering leaders to adopt a platform approach that is scalable, outsourceable, geared for innovation and future-proof.</h4><p><em>This is the fourth in a </em><a href="https://praful-krishna.medium.com/pragmatic-playbook-for-enterprise-ai-b54419af9c98"><em>series</em></a><em> of posts meant to help CEOs and business leaders adopt Artificial Intelligence at enterprises. The series tries to cut through the hype and focuses on ROI. These are the author’s personal views (</em><a href="https://praful-krishna.medium.com/a-systematic-approach-to-generating-enterprise-ai-ideas-c5c4d10cba3d"><em>Previous post</em></a><em>).</em></p><p>Earlier posts in the series talk about how to generate various AI related ideas and the best way to prioritize them. Pragmatic CEOs and business leaders partner with their CTOs to then take a platform approach to implementing these ideas and preparing for future ones. This post suggests a platform architecutre and discusses the best practices.</p><h4>Understanding the Layered Architecture</h4><p>The platform architecture diagram presents a comprehensive framework for enterprise AI transformation, structured in layers.</p><p>The <strong>Resource Managers</strong> layer serves as the fundamental orchestration tier, providing transparent, scalable, and configurable access to underlying resources. Underpinning these Resource Managers is the <strong>APIs layer</strong>, which provides the actual integration points that Resource Managers coordinate and abstract. These APIs represent the raw capabilities that the Resource Managers make accessible and manageable for the rest of the platform. Above this sits the <strong>Common Components layer</strong> — logically developed components used by most applications for consistency and scale. At the top is the <strong>Applications layer</strong>, where individual apps function through the framework.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*cFVmIc6jx2Y5TMOqrNfW4Q.jpeg" /><figcaption>Reach out to the author for the complete presentation.</figcaption></figure><p>This application layer provides for critical capabilities that enable enterprise-grade AI solutions. Applications coordinate complex workflows across various services and components, managing the entire lifecycle from user input to final output, ensuring cohesive experiences regardless of underlying complexity. They also define the prompts or chains of prompts that define the agentic patterns, bring various components together, call on various tools and implement the business logic.</p><p>Some applications employ strategic caching at various levels to balance freshness of information with performance requirements, reducing latency and costs while maintaining appropriate responsiveness.</p><p>Each application leverages these capabilities while remaining modular participants in the broader ecosystem, drawing from common components while addressing specific business needs. This platformized approach accelerates AI transformation by creating a cohesive ecosystem where components are reusable, consistent, and scalable.</p><p>If you are thinking about such a platform for your own company/ team (highly recommended), please keep the following in mind as well.</p><h4>1. Treat Every API as an Agentic Tool</h4><p>The landscape of AI capabilities is evolving rapidly, with increasingly sophisticated agentic abilities becoming standard in modern applications. By managing every API as a potential tool for your agentic core, you future-proof your entire development.</p><p>The Model Context Protocol represents a promising approach here — it allows your applications to interact with other tools in a consistent way regardless of the underlying implementation. When your own internal applications are developed with modularity and API-driven architecture in mind, they seamlessly become tools within the broader platform.</p><p>This approach creates a virtuous cycle: as LLMs grow more capable of orchestrating complex workflows, your application components can be composed in increasingly sophisticated ways without requiring architectural overhauls. This also means that the speed with which your team can produce new apps gets accelerated.</p><h4>2. Embrace Transparent Complexity with Strong Defaults</h4><p>Enterprise AI platforms must balance two competing needs: providing complete control for power users while remaining accessible to everyday users. The solution lies in transparency of tools while establishing strong, intelligent defaults.</p><p>Consider LLM selection — rather than forcing users to choose which model is appropriate for summarization versus code generation versus data analysis, implement smart routing that automatically selects the optimal model based on the detected task. The platform should handle the complexity while presenting a simple interface. At the same time, if a particular app has a specific need, it should be able to configure for it.</p><p>This “complexity behind glass” approach extends to prompt management, memory systems, and evaluation frameworks. Users of the platform should be able to focus on their business problems rather than AI implementation details, while still having the option to customize when necessary.</p><h4>3. Design for Ecosystem Extension and Third-Party Integration</h4><p>No enterprise can — or should — build every component of their AI in-house. A successful platform acknowledges this reality through deliberate architectural choices that facilitate integration with third-party solutions. A successful platform also seamlessly integrates with third-party apps.</p><p>This requires careful consideration of your organization’s “right to win” — those areas where your unique expertise and capabilities give you a competitive advantage. Focus your development resources on these components while designing clean interfaces for external tools and services to handle the rest. See this link below for a suggested prioritization framework.</p><p><a href="https://praful-krishna.medium.com/the-five-lanes-in-a-ceos-ai-roadmap-1d116a30dfb6">The Five Lanes on a CEO’s AI Roadmap</a></p><p>For example, if customer support automation isn’t your core competency, consider integrating established customer service platforms rather than building your own ticketing system, knowledge base, and conversation history management tools. This not only accelerates your time-to-market but also ensures you benefit from specialized expertise in customer experience management while allowing you to focus on your industry-specific AI capabilities.</p><h4>4. Integrate Evaluation Throughout the Stack</h4><p>AI systems require continuous evaluation to ensure quality, safety, and improvement over time. A robust enterprise platform doesn’t treat evaluation as an afterthought but integrates it throughout every layer.</p><p>At the lower levels, this might include automated unit tests for individual API integrations and components. As you move up the stack, evaluation becomes more sophisticated — measuring hallucination rates, answer relevance, and business impact metrics. Additionally, establish clear patterns for application-specific evaluations, allowing teams to extend the platform’s evaluation framework with domain-specific metrics and test cases.</p><p>By providing evaluation capabilities as platform services, you democratize quality assurance and enable teams to implement appropriate testing without reinventing methodologies. You create levers of higher control on the applications, and get the data necessary to celebrate early successes.</p><h4>5. Create a Unified Experience Layer for Future-Proof Interactions</h4><p>User experience channels are multiplying and evolving — from traditional web interfaces to conversational AI, generative experiences, and emerging technologies like AR/VR. Building separate integration paths for each channel creates unsustainable complexity and fragmentation.</p><p>Instead, implement a pass-through but integrated layer that standardizes how applications connect to different interaction modalities. This architecture ensures that as new interfaces emerge, your applications can adapt without fundamental changes to their core logic.</p><p>This approach acknowledges that the future of AI interaction will likely involve blended experiences that combine multiple modalities — for instance, a customer service application might switch seamlessly between chat, voice, and visual interfaces depending on the context and user preference.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B-HoTJEFO56mYCZX8ftbqw.jpeg" /></figure><p>Enterprise AI platforms face unique challenges compared to consumer applications. Security requirements, governance concerns, legacy system integration, and organizational change management all influence architectural decisions. The platformized approach described above addresses these challenges by creating a structured, consistent framework that still allows for innovation and experimentation.</p><p>As we continue to see AI capabilities evolve at breakneck speed, these architectural principles provide a foundation that can adapt to new models, techniques, and interaction paradigms without requiring constant reinvention.</p><p><em>Complete Series:</em></p><p><a href="https://praful-krishna.medium.com/pragmatic-playbook-for-enterprise-ai-b54419af9c98">Pragmatic Playbook for Enterprise AI</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=62c6fc0758a6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pragmatic Playbook for Enterprise AI]]></title>
            <link>https://praful-krishna.medium.com/pragmatic-playbook-for-enterprise-ai-b54419af9c98?source=rss-2ae1c32a010------2</link>
            <guid isPermaLink="false">https://medium.com/p/b54419af9c98</guid>
            <category><![CDATA[strategy]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[careers]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <dc:creator><![CDATA[Praful Krishna]]></dc:creator>
            <pubDate>Sun, 06 Apr 2025 17:58:03 GMT</pubDate>
            <atom:updated>2025-06-27T03:49:20.761Z</atom:updated>
            <content:encoded><![CDATA[<p>This playbook is being written for pragmatic CEOs and business leaders who want to use AI for transformative step up in their enterprises.</p><p>The playbook is being written as a series of posts. I keep writing these posts based on feedback and requests I get, and based on top of the mind issue in any week. As a result the series is not necessarily sequential. You may want to read the posts in the following order:</p><p>First, this post talks about how to systematically start thinking of enterprise AI ideas, and not just cave in to the headline of the day.</p><p><a href="https://praful-krishna.medium.com/a-systematic-approach-to-generating-enterprise-ai-ideas-c5c4d10cba3d">A Systematic Approach to Generating Enterprise AI Ideas</a></p><p>Next, this post offers a framework (Five Lanes) for prioritizing those ideas. As pragmatic CEOs and business leaders know very well, sometimes it’s best not to act.</p><p><a href="https://praful-krishna.medium.com/the-five-lanes-in-a-ceos-ai-roadmap-1d116a30dfb6">The Five Lanes on a CEO’s AI Roadmap</a></p><p>Next, this post offers a platform approach to developing and integrating various AI apps. This is especially relevant for the first two lanes of the Five Lanes framework (Initiate and Insource).</p><p><a href="https://praful-krishna.medium.com/best-practices-for-building-an-enterprise-ai-platform-62c6fc0758a6">Best Practices for Building an Enterprise AI Platform</a></p><p>Any transformation is about people at the end of the day. This post double clicks on the Inspire lane of the Five Lanes framework, and lays out a proven path to deal with the people dimension in an AI transformation.</p><p><a href="https://praful-krishna.medium.com/a-blueprint-for-inspiration-ai-change-management-at-enterprises-e4d750615ff5">A Blueprint for Inspiration — AI Change Management at Enterprises</a></p><p>It’s also important to keep in mind that Generative AI is not just about language or image dominant usecases. It can help transformatively in areas which are considered domains of traditional AI or machine learning.</p><p><a href="https://praful-krishna.medium.com/seven-ways-generative-ai-is-quietly-revolutionizing-data-analytics-05705b4cec12">Seven Ways Generative AI is Quietly Revolutionizing Data Analytics</a></p><p>I will keep adding more posts in the series over here as I write them. If you have something on your mind and would like to discuss, please reach out to me here or on <a href="https://www.linkedin.com/in/prafulkrishna/">LinkedIn</a>. I will be very happy to hear from you.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HDX6bXOgnGHolhH9snJkYA.png" /><figcaption>Image generated by author using Gemini.</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b54419af9c98" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Systematic Approach to Generating Enterprise AI Ideas]]></title>
            <link>https://praful-krishna.medium.com/a-systematic-approach-to-generating-enterprise-ai-ideas-c5c4d10cba3d?source=rss-2ae1c32a010------2</link>
            <guid isPermaLink="false">https://medium.com/p/c5c4d10cba3d</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[ideas]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[strategy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Praful Krishna]]></dc:creator>
            <pubDate>Sun, 06 Apr 2025 17:12:16 GMT</pubDate>
            <atom:updated>2025-04-06T17:58:58.530Z</atom:updated>
            <content:encoded><![CDATA[<h4>Pragmatic CEOs and business leaders know how to systematically drive AI-transformation by 1. Understanding core strengths, 2. Thinking about revenue if there was no cost, 3. Methodically identifying AI opportunities, and 4. Pursuing what matters.</h4><p><em>This is the third in a </em><a href="https://praful-krishna.medium.com/pragmatic-playbook-for-enterprise-ai-b54419af9c98"><em>series</em></a><em> of posts meant to help CEOs and business leaders adopt Artificial Intelligence at enterprises. The series tries to cut through the hype and focuses on ROI. These are the author’s personal views (</em><a href="https://praful-krishna.medium.com/a-blueprint-for-inspiration-ai-change-management-at-enterprises-e4d750615ff5"><em>Previous post</em></a><em>).</em></p><p>In the rapidly evolving landscape of artificial intelligence, many organizations find themselves caught in a cycle of chasing popular use cases without a coherent strategy. While generative AI chatbots and image generation tools may capture headlines, pragmatic CEOs and business leaders recognize that meaningful implementation requires a more systematic approach. Rather than merely following trends, or relying solely on creative thinking and serendipity, successful organizations are methodically and systematically mining three areas for transformative AI-led value: cost disruption, speed enhancement, and breakthrough innovation.</p><h3>The Cost Disruption Imperative</h3><p>Perhaps the most significant yet underappreciated power of AI lies in its potential to dramatically reshape cost structures. Most importantly, cost disruption represents the single most powerful mechanism for unlocking new revenue streams and business models.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ue5D8NrIy-LsZZqS1OBl8A.png" /><figcaption>A simplified graph illustrating the argument that a lot of value remains unlocked because the incremental cost of delivering an incremental feature remains prohibitive.</figcaption></figure><p>Let us consider the universe of current and potential features for a product, and rank them by importance. The incremental willingness to pay for each feature will generally follow a downward sloping curve. The incremental cost for delivering each feature can vary considerably. Basic economics suggests that sustainable businesses can only deliver those features where willingness to pay exceeds costs.</p><p>However, if costs were to change drastically, the value unlocked can be immense. Features that were previously shelved due to prohibitive development or operational costs can now be key differentiators. Business models that couldn’t scale profitably can potentially generate monopolistic returns. This cost disruption fundamentally redefines what’s possible. In today’s world, such disruptive cost curves also mean the ability to scale by orders of magnitude.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vv0gqzgumTzYuJmCoLdKUw.png" /><figcaption>A simplified graph illustrating the argument that disruptive reduction in cost unlocks significant revenue. Adding to speed of delivery and quality will enhance the willingness to pay as well.</figcaption></figure><p>Most importantly, focusing on costs is a scientific process, where ideation can be methodical and comprehensive.</p><p>History provides us with compelling parallels. The Internet made the cost of sharing information disruptively low, which led to transformations far beyond simple cost savings. At a macro level, this cost disruption enabled entirely new categories: social media platforms connecting billions, globally distributed teams collaborating seamlessly, and eventually the training data abundance that powers today’s generative AI systems. Similarly, AI has already begun disrupting costs in domains ranging from hyper-personalization in marketing to drug discovery.</p><p>For business leaders, the strategic implications are clear. When approaching AI implementation, the primary question should be: <strong><em>“Where can we most meaningfully disrupt our cost structures in ways that unlock new revenue potential?”</em></strong> This requires an honest assessment of your organization’s competitive strengths and core value propositions. Pragmatic leaders focus AI investments on disrupting costs in their areas of differentiation, while leveraging ecosystem solutions for non-core functions.</p><p>Beyond the revenue unlock — which remains the most important aspect of disrupting delivery costs — many leaders are also leveraging AI for meaningful improvements to their bottom line. Automating routine tasks, optimizing resource allocation, and reducing error rates all contribute to healthier margins. While these efficiency gains may not be as strategically transformative as unlocking new revenue streams, they provide the financial flexibility to invest in more ambitious AI initiatives.</p><h3>The Speed Advantage</h3><p>Apart from reducing the cost of delivering features or services, AI can make their delivery remarkably fast in most cases. Customers often value faster delivery, quicker responses, or more frequent updates. By leveraging AI to accelerate these processes, you can actually shift the willingness-to-pay curve upward for many features. A medical diagnosis delivered in minutes rather than days carries substantially higher value. Financial insights provided in real-time rather than at month-end enable better decision-making. Customer service responses received instantly rather than after hours in a queue dramatically improve satisfaction.</p><p>More profoundly, speed enables iterations between product nodes and users that unlock entirely new patterns of interaction and value creation. Again, historical parallels are instructive. The Internet’s ability to enable near-instantaneous information sharing led to transformative business models like Amazon’s marketplace, Uber’s ride-hailing service, and instant messaging platforms. Similarly, AI has already accelerated workflows for knowledge workers across industries. Product development cycles that previously took months now require weeks. Customer insights that demanded quarters of analysis can now be generated in days. Marketing campaigns that once required extensive pre-planning can now adapt in near-real-time to market conditions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gE97MYfOVmHOWHr483EEZg.png" /><figcaption>Image generated by author using Gemini.</figcaption></figure><p>For business leaders, the strategic question becomes: <strong><em>“Where would dramatic improvements in speed create disproportionate value for our customers and operations?”</em></strong> This requires examining your entire value chain to identify bottlenecks and friction points where AI-driven acceleration could transform the experience or economics. Again, this is a methodical and systematic process.</p><p>Companies successfully implementing this approach are focusing on speed at key interaction points. Healthcare providers are using AI to reduce diagnostic waiting times from weeks to hours, leading to better patient outcomes. Financial services firms are accelerating loan approval processes from days to minutes, leading to larger total addressable markets. Manufacturing companies are enabling real-time production line optimization rather than periodic adjustments, boosting productivity. In each case, the speed enhancement transforms not just operational efficiency but customer experience and business capabilities.</p><h3>Breakthrough Innovation</h3><p>While cost disruption and speed enhancement represent powerful and immediately applicable approaches to AI implementation, the most transformative potential lies in breakthrough innovations — capabilities that were previously inconceivable. These aren’t simply cheaper or faster versions of existing products and services, but entirely new paradigms made possible by AI’s unique capabilities.</p><p>The parallel with previous technological revolutions is again instructive. The Internet didn’t just make information sharing cheaper and faster — it enabled entirely new business models like Software-as-a-Service (SaaS). This pay-per-use, multi-tenant approach to software delivery fundamentally transformed both the economics and accessibility of business applications. Similarly, AI has the potential to enable business models and offerings that simply couldn’t exist in a pre-AI world.</p><p>We’re seeing early examples of these breakthroughs across industries. In healthcare, AI systems can now predict health events before physical symptoms appear, enabling a shift from reactive to preventative care. In manufacturing, generative design approaches produce structures and components with properties impossible to achieve through traditional design methods. Some experts believe the emergence of superintelligence in the next 5–10 years will push our collective imaginations to new boundaries.</p><p>For business leaders, identifying and pursuing these breakthrough opportunities requires a different mindset from cost disruption or speed enhancement initiatives. Rather than starting with existing products and processes, breakthrough innovation begins with the question: <strong><em>“What would be possible if the constraints we’ve always accepted simply didn’t exist?”</em></strong> This requires deliberate imagination exercises and cross-disciplinary collaboration.</p><h3>A Systematic Approach to Implementation</h3><p>For CEOs and business leaders convinced by this framework of cost, speed, and breakthroughs, the path forward becomes clear. A systematic approach to AI implementation follows these essential steps:</p><ol><li><strong>Understand your core strengths: </strong>First, understand what your business is really about, AI or no AI. This means gaining clarity on your core value proposition, competitive advantages, and customer needs. AI implementation must serve these fundamental business realities rather than becoming an end in itself. Organizations that pursue AI without this clarity often find themselves with impressive technical capabilities that deliver minimal business impact.</li><li><strong>Think about revenue if there was no cost: </strong>Second, imagine your business in a world without constraints on cost or time. If delivering your core products and services cost a fraction of what it does today and could happen instantaneously, how would your offerings evolve? What new customer segments would become accessible? What adjacent markets might you enter? What entirely new business models might emerge? This unconstrained thinking exercise helps identify the most valuable targets for AI implementation.</li><li><strong>Methodically think through AI opportunities: </strong>Third, follow a systematic approach to disrupting delivery costs and timeframes using AI. This means mapping your value chain to identify the highest-impact opportunities for cost disruption and speed enhancement. It means establishing clear metrics for success that connect technical capabilities to business outcomes. It means developing an implementation roadmap that balances quick wins with longer-term transformational initiatives.</li><li><strong>Pursue what matters, and only what matters: </strong>Finally, focus relentlessly on these strategic priorities when it comes to AI implementation. The landscape of AI capabilities and vendors grows more complex by the day, and organizations without clear strategic focus can easily find themselves pursuing dozens of disconnected initiatives with minimal aggregate impact. You may consider the <a href="https://praful-krishna.medium.com/the-five-lanes-in-a-ceos-ai-roadmap-1d116a30dfb6">Five Lanes</a> framework outlined in a previous post.</li></ol><p><a href="https://praful-krishna.medium.com/the-five-lanes-in-a-ceos-ai-roadmap-1d116a30dfb6">The Five Lanes on a CEO’s AI Roadmap</a></p><p>As AI continues its rapid evolution, the gap between organizations pursuing systematic implementation and those chasing popular use cases will widen. The former will achieve transformative business impact through strategically aligned applications of AI capabilities. The latter will accumulate an impressive collection of pilot projects with minimal aggregate effect on their competitive position or financial performance. The organizations that approach this technology with strategic clarity and systematic implementation will not merely adapt to change — they will drive it.</p><p><em>Complete series:</em></p><p><a href="https://praful-krishna.medium.com/pragmatic-playbook-for-enterprise-ai-b54419af9c98">Pragmatic Playbook for Enterprise AI</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c5c4d10cba3d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Blueprint for Inspiration — AI Change Management at Enterprises]]></title>
            <link>https://praful-krishna.medium.com/a-blueprint-for-inspiration-ai-change-management-at-enterprises-e4d750615ff5?source=rss-2ae1c32a010------2</link>
            <guid isPermaLink="false">https://medium.com/p/e4d750615ff5</guid>
            <category><![CDATA[human-resources]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Praful Krishna]]></dc:creator>
            <pubDate>Sat, 29 Mar 2025 17:24:08 GMT</pubDate>
            <atom:updated>2025-04-06T17:59:37.696Z</atom:updated>
            <content:encoded><![CDATA[<h3>A Blueprint for Inspiration — AI Change Management at Enterprises</h3><h4>Pragmatic CEOs and business leaders know how to inspire the humans around them to be partners in a bold AI vision, rather than detractors.</h4><p><em>This is the second in a </em><a href="https://praful-krishna.medium.com/pragmatic-playbook-for-enterprise-ai-b54419af9c98"><em>series</em></a><em> of posts meant to help CEOs and business leaders adopt Artificial Intelligence at enterprises. The series tries to cut through the hype and focuses on ROI. These are the author’s personal views. (</em><a href="https://medium.com/@praful-krishna/the-five-lanes-in-a-ceos-ai-roadmap-1d116a30dfb6"><em>Previous post</em></a><em>).</em></p><p>The integration of artificial intelligence into enterprise operations represents both a profound opportunity and a significant leadership challenge. While the technological dimension of AI implementation demands careful consideration, the human aspects of this transformation often determine success or failure. As you lead your organization through this evolution, you must acknowledge that resistance to AI rarely stems from the technology itself, but rather from deeply human concerns about what these changes mean for careers, expertise, and agency.</p><p>While the rest of the article focuses on inspiring your team, other human stakeholders may face similar challenges, and you must inspire them similarly. These could include your customers, partners, vendors or others.</p><h4>Understanding Resistance to Change</h4><p>When you introduce AI initiatives, you will frequently encounter employee apprehension that manifests in several forms. The most prevalent concern centers on <strong>job security</strong> — the understandable anxiety about whether roles will become obsolete as automation expands. This concern often operates beneath the surface, expressing itself through subtle resistance rather than direct opposition.</p><p>Many professionals also experience a profound <strong>discomfort with shifting expertise</strong> requirements. After investing years developing specialized knowledge, the prospect of requiring entirely new skills creates uncertainty about professional identity and value. This anxiety particularly affects mid-career employees who have established themselves through domain expertise that AI may appear to challenge.</p><p>Perhaps most fundamentally, many resist <strong>relinquishing control</strong> to algorithmic systems. This resistance reflects not mere technophobia but legitimate questions about accountability and judgment. When professionals have developed nuanced decision-making capabilities through years of experience, algorithmic alternatives may feel like a devaluation of human wisdom rather than an enhancement of capabilities. A lack of transparency about how AI systems operate only compounds these concerns, raising questions about unintended consequences and potential biases.</p><p>As shared in a previous post in the series, each of these, and every other human perspective, is valid even if it seems unfounded.</p><p><a href="https://praful-krishna.medium.com/the-five-lanes-in-a-ceos-ai-roadmap-1d116a30dfb6">The Five Lanes on a CEO’s AI Roadmap</a></p><h3>1. Align the AI Strategy with Organizational Purpose</h3><p>Effective AI transformation begins with strategic alignment. Rather than pursuing implementation as a technical exercise, you must position these initiatives within your broader business vision — how will this help the company grow faster, compete better or become more profitable and valuable. When AI clearly advances organizational mission, the narrative shifts from disruption to strategic evolution.</p><p>This alignment must permeate all leadership levels. When middle managers receive mixed messages about priorities, they inevitably transmit this ambivalence to their teams. Conversely, when leaders at every level consistently articulate how AI advances core objectives, they create coherence that facilitates adoption. This requires deliberate effort to ensure every leader from the C-suite to front-line management understands and communicates a consistent vision.</p><p>AI adoption cannot be confined to IT teams or innovation departments; it must be ubiquitous, championed at the highest levels of the organization. When executives demonstrate their commitment not just in rhetoric but in action, it signals to employees that this transformation is both necessary and inevitable, accelerating the pace of change throughout the enterprise.</p><h3>2. Communicate Positively, Empathetically and Honestly</h3><p>Perhaps no element of change management proves more crucial than thoughtful communication that begins well before implementation. Forward-thinking organizations <strong>initiate AI awareness programs</strong> through channels like internal forums and educational events that familiarize employees with concepts and applications before specific projects commence. This early exposure demystifies AI and establishes shared understanding for subsequent conversations.</p><p>How you frame these communications significantly influences reception. Messaging must <strong>emphasize enhanced capabilities, new opportunities, and improved outcomes</strong>. This resonates more effectively than narratives focused primarily on efficiency or cost reduction. While operational benefits matter to performance, they often trigger defensive responses that impede adoption.</p><p>You must approach organizational transformation with honesty. Rather than making unrealistic promises about AI’s impact, effective leaders <strong>acknowledge uncertainty while expressing confidence</strong> in the collective ability to adapt. This approach balances transparency with reassurance, addressing concerns without fostering undue anxiety.</p><p>The communication also needs to two-way — <strong>create structured opportunities for feedback</strong> to demonstrate commitment to inclusive transformation. Town halls, anonymous input mechanisms, and one-on-one conversations all provide valuable channels for surfacing concerns and identifying implementation challenges before they derail progress. When employees feel heard and valued, their perspective shifts from resistance to participation.</p><h3>3. Build Trust Through Thoughtful Product Strategy</h3><p>The sequencing of AI solutions significantly impacts organizational receptivity. When you begin with applications that address clear pain points and deliver immediate value, you build positive momentum for subsequent, more complex implementations. Early successes create organizational ambassadors who can testify to AI’s benefits from personal experience. Smaller use cases also flatten the learning curve before big ideas become more acceptable.</p><p>When you incorporate principles of transparency, verifiability, and human control into solution design, you address many prevalent concerns about adoption. Systems that explain their recommendations, allow for verification of outputs, and maintain human oversight over consequential decisions build trust through their fundamental architecture. These features should not be afterthoughts but core design principles from initial conception.</p><p>Rigorous measurement and transparent sharing of results reinforce the empirical foundations of your AI implementation. When you objectively track both successes and limitations, you demonstrate commitment to evidence-based advancement rather than technological faith. Employees need to see tangible outcomes — efficiency gains, improved decision-making, enhanced customer satisfaction — to appreciate AI’s value and overcome skepticism.</p><h3>4. Help People Catch up with AI</h3><p>Comprehensive training represents an essential component of successful adoption. This extends beyond technical instruction to include conceptual understanding of how AI systems work, their capabilities, and their limitations. When your colleagues understand the principles underlying AI solutions, they engage more confidently and apply these tools more effectively to novel situations.</p><p>Providing safe environments for experimentation can be an effective part of such training. Sandbox environments allow employees to explore AI capabilities without fear of consequential errors or professional exposure. These low-risk opportunities for familiarization help bridge the gap between theoretical understanding and practical application, while addressing legitimate privacy and security concerns through appropriate safeguards.</p><p>Your educational initiatives must also reinforce the organization’s commitment to ethical AI use and data privacy, addressing any lingering concerns about security or misuse. By emphasizing your ethical framework, you further alleviate anxieties about how AI will be deployed within the organization.</p><h3>5. Ensure Successes Precede Changes</h3><p>Recognition of early adopters serves both practical and symbolic functions. It identifies internal experts who can support broader implementation efforts while signaling organizational commitment to valuing those who embrace new capabilities. These individuals’ success stories serve as inspiration, demonstrating real-world benefits of AI and encouraging broader participation.</p><p>If hard decisions become necessary, the successes start creating an atmosphere of understanding and empathy. Patience regarding organizational restructuring represents the final critical element. of an effective change management approach.</p><p>Premature reorganization based on anticipated AI impacts often creates unnecessary anxiety and resistance. By allowing time for implementation results to materialize before making structural changes, you demonstrate respect for human adaptation and enable more informed organizational design decisions. Leaders must resist the temptation to impose sweeping structural changes before AI’s impact is fully realized, allowing AI-driven efficiencies to emerge organically before making significant workforce adjustments.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WmKvooIKVUERfVMx1Q2EPw.jpeg" /><figcaption>Image generated by author using Gemini.</figcaption></figure><p>The most successful AI transformations reject the false dichotomy between human and artificial intelligence. Instead, they embrace a partnership model that leverages complementary strengths. By implementing thoughtful change management strategies that address human concerns while highlighting this complementary relationship, you can accelerate AI adoption while strengthening organizational culture. In this approach, AI becomes not a threat to human contribution but an amplifier of human potential — expanding capabilities, creating new opportunities, and enabling your teams to focus on the most meaningful aspects of their work.</p><p><em>Next post in the series:</em></p><p><a href="https://praful-krishna.medium.com/a-systematic-approach-to-generating-enterprise-ai-ideas-c5c4d10cba3d">A Systematic Approach to Generating Enterprise AI Ideas</a></p><p><em>Complete series:</em></p><p><a href="https://praful-krishna.medium.com/pragmatic-playbook-for-enterprise-ai-b54419af9c98">Pragmatic Playbook for Enterprise AI</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e4d750615ff5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Five Lanes on a CEO’s AI Roadmap]]></title>
            <link>https://praful-krishna.medium.com/the-five-lanes-in-a-ceos-ai-roadmap-1d116a30dfb6?source=rss-2ae1c32a010------2</link>
            <guid isPermaLink="false">https://medium.com/p/1d116a30dfb6</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Praful Krishna]]></dc:creator>
            <pubDate>Wed, 26 Mar 2025 18:39:13 GMT</pubDate>
            <atom:updated>2025-04-10T12:05:23.607Z</atom:updated>
            <content:encoded><![CDATA[<h4>Pragmatic CEOs and business leaders know how to 1. Initiate, 2. Insource, 3. Innovate (or wait), 4. Inspire, and/or 5. Insulate when it comes to incremental or transformative AI initiatives in their enterprises.</h4><p><em>This is the first in a </em><a href="https://praful-krishna.medium.com/pragmatic-playbook-for-enterprise-ai-b54419af9c98"><em>series</em></a><em> of posts meant to help CEOs and business leaders adopt Artificial Intelligence at enterprises. The series tries to cut through the hype and focuses on ROI. These are the author’s personal views.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7f-LatceElxgO9nUDWAKlg.png" /></figure><p>Most CEOs and business leaders understand that AI is not an end goal nor a strategy unto itself. It’s merely a tool that we must use in service of our business priorities. Any approach to AI must start with a clear prioritization of the overall business objectives, AI or no AI.</p><p>At the same time, AI can be a potentially transformative tool. If applied correctly, it can hyper-charge your business. As a simple thought experiment, let’s start by assuming that everything is possible with AI. We all know it’s not true, but if it were, what could your business priorities look like? In reality, this makes a good starting point for a bold AI vision. Ironically, it actually let’s you focus even deeper on your core competitive strengths.</p><p>AI is also evolving super fast. Open AI launched ChatGPT less than 2.5 years ago, and everyone is already experimenting with advanced concepts like Multimodality, Agentic AI with reasoning, and Large Action/ Concept Models. This pace of evolution makes the thought experiment less experimental. It’s reasonable to assume multiple breakthroughs in the next 3–5 years. Ergo, it’s reasonable to assume that any AI-fueled strategy can target 10x more growth/ change in top-level metrics than a strategy without AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LuzDbv7k_q_yNp5gEafhTA.png" /><figcaption>A bold AI vision aims to 10x the achievement of major business priorities over 3–5 years.</figcaption></figure><p>The as-is Business Priorities and the AI-Fueled Business Priorities may not match 1:1, and that’s fine. However, once you have a list of AI Levers that can hyper-charge your business priorities, it’s time to pick a lane on your roadmap for each of them. You can jump ahead to a complete decision tree towards the end if you like.</p><h3>Lane 1 — Initiate with 3rd Party Solutions</h3><p><em>Key Question: Is the current state of the art reliable enough for your use case?</em></p><p>The AI ecosystem is innovating at two levels: First, the headline level — the things that work in the labs — and the second, the workhorse level — the things that perform reliably and at scale. There may be many vendors that are already providing the AI solutions that you need for your priorities. An easy win would be to start incorporating these products in your systems. In general buying is far quicker, more effective, more future proof and cheaper than building on your own in this category.</p><p>This lane is about super fast testing and rolling out. In this lane your team needs to worry about quality, governance, privacy, integration, and change management to drive higher adoption of these tools. In this lane you need to worry about articulating and enforcing a high bar around helpfulness, latency, compliance and measurement related to the solutions, including definitively measuring the ROI.</p><h3>Lane 2 — Insource and Build your own Solutions</h3><p><em>Key Question: Do you need a bespoke solution based on the state of the art AI technology?</em></p><p>There may be certain processes or insights that are core to your team’s competitive advantage. This is the most important reason you would consider insourcing the development of AI solutions — a vendor may never be able to understand all the nuances. This is perhaps the only acceptable reason as well. Most other reasons are typically made up and have more to do with team dynamics.</p><p>In this lane it may still be a good idea to initiate the solutions in Lane 1 with third party vendors to assess, to learn and to put in place the necessary quality, governance and privacy frameworks. In this lane your team needs to focus on the key differentiation for your business. In this lane, you must make sure that your team is spending majority of its time bolstering the differentiation rather than building commoditized products/ features that should really be in Lane 1 — Initiate.</p><h3>Lane 3 — Innovate with the overall AI Ecosystem</h3><p><em>Key Question: Is the current state of the art insufficient for your strategic needs?</em></p><p>A big part of your list of AI levers/ projects will be things that the AI is not yet ready for. Most likely you will find headlines on how this can be done, but no real testimonials from your peers. In a vast majority of these cases, it makes sense to just wait for the ecosystem to catch up. Given the pace of evolution of AI, it may be only six months away. However, there are situations where it may make sense to start innovating ahead of this ecosystem.</p><p>This lane has the potential to be the most meaningful for your business, and yet the most frustrating and/or expensive. Before you commit to any use case in this lane, put it through this four part test:</p><ol><li>If the state of the art was ready today, would this use case fall in the Lane 2 — Insource (as against the Lane 1 — Initiate)?</li><li>Is the time to market for this use case critical for your team’s success? For example, is this a blocker for adoption of other AI initiatives? Or perhaps there is a key first-mover advantage?</li><li>If it were feasible, is this use case the highest priority application of your resources?</li><li>Is the innovation that the use case needs something that a future, better version of LLMs cannot do out of the box?</li></ol><p>I would reemphasize the fourth test. All major players like Open AI and Gemini are continuously introducing new versions of the large language models, and each version is that much more powerful than the previous one. Then there are players like Deepseek who upend the status quo. More are likely to emerge in the future. The innovation that you need to invest in should be parallel to the LLMs themselves. For example, providing the right connectors, building the right memory or context, or investing in a unique, bespoke need that most LLMs are unlikely to cater to. A red flag in this lane would be if your team tells you that they need to improve the quality of an LLM to get the results. While there are exceptions, to me the word ‘fine tuning’ is also a major red flag.</p><p>If all four of these tests come back with a Yes, it’s a good sign that you must invest behind innovation in these use cases. Otherwise just wait a few months and reassess. In this lane you and your team need to unlearn and learn product development skills again, this time for AI. Pragmatic CEOs and business leaders know that the journey of innovation is rarely straight-forward. In this lane managing the distinction and the duality of innovating and waiting for the ecosystem can become a critical success factor.</p><h3>Lane 4 — Inspire your Team to be more Open to AI</h3><p><em>Key Question: Even if everything is good to go, does your team feel ready?</em></p><p>CEOs and business leaders know that the definition of a right-thing-to-do spans more than dollars and technology. Adoption of AI at scale can lead to unique human challenges with your team if not handled appropriately. For example, there may be anxieties around job security. Or there may be a sense of obsolescence because some people feel they are unable to use AI. There may be others who struggle with a machine-first approach.</p><p>Each of these, and every other human perspective, is valid even if seemingly unfounded. You may want to consider similar perspectives from other stakeholders as well, esp. your customers.</p><p>At the same time as more and more people use AI and realize its powers as well as limitations, they get more comfortable delegating to it. As a leader your job is to inspire no matter what. From a roadmapping perspective, you may find that some use cases are not the first ones you should take up. They need more discussion, higher familiarity with AI/ automation and a more supportive environment to be successful, even if every other indicator is green.</p><p>This is an often ignored, and yet super important, gestation lane for your bold ideas. In this lane you need to bring everyone together and make sure that your team (and other relevant stakeholders) are ready for your vision before you roll it out. An effective way to do this would be to focus on gateway use cases first, while emphasizing on human control at all times.</p><p>It’s also important to recognize that any approach to change management or inspiration should be comprehensive and not just for the AI projects in this lane.</p><p><a href="https://praful-krishna.medium.com/a-blueprint-for-inspiration-ai-change-management-at-enterprises-e4d750615ff5">A Blueprint for Inspiration — AI Change Management at Enterprises</a></p><h3>Lane 5 — Insulate the Legal and Policy No-Nos</h3><p><em>Key Question: Are your use cases limited by legal or policy compliance?</em></p><p>The legal and compliance boundaries that good CEOs and business leaders set for their scope of operations make for an overarching consideration.</p><p>There are multiple pieces of legislation related to AI that can be pertinent to your team. For example the EU AI Act or those being proposed by the California Civil Rights Council. In addition each team or company typically makes its own policy choices based on its own context. For example, a financial organization with fiduciary responsibility to consumers will think about this very differently than a B2B tech player. Over time, the regulatory burden is likely to increase while the policy burden, like human acceptance of AI, is likely to evolve favorably.</p><p>Irrespective, at any point it’s super important to insulate the use cases that are a no go from any kind of AI development. This is critical to keep your overall AI vision free from these types of distractions, which can bubble up unexpectedly. The larger the team you manage, the larger is the potential for legal/ compliance distractions to derail your AI agenda.</p><p>In this lane you and your team must focus on a good governance framework for AI, one that can proactively alert for any kind of non-compliant behavior. It can be a great way to push the legal/ compliance limits without actually inadvertently crossing the line. In this lane you should also think about pushing the envelope creatively, and use humans in the loop as a key lever to derisk.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d4b6qazUvCynPhwZ0Lo85g.png" /><figcaption>Picking the right lane for each AI initiative on the roadmap can help CEOs and business leaders get a strong head-start by focusing on what truly matters and compounding a series of small and big wins.</figcaption></figure><p>There is simply too much in this playbook for a single article to even scratch its surface. However, framing your problem in these five parts will give you a strong head-start in adopting AI in an effective way for your enterprise. This will also ensure that your mistakes are small, and wins, even if hard-fought, transformational.</p><p><em>Next post in the series:</em></p><p><a href="https://praful-krishna.medium.com/a-blueprint-for-inspiration-ai-change-management-at-enterprises-e4d750615ff5">A Blueprint for Inspiration — AI Change Management at Enterprises</a></p><p><em>Complete series:</em></p><p><a href="https://praful-krishna.medium.com/pragmatic-playbook-for-enterprise-ai-b54419af9c98">Pragmatic Playbook for Enterprise AI</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1d116a30dfb6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Enterprise RAG: Beyond the Basics]]></title>
            <link>https://praful-krishna.medium.com/enterprise-rag-beyond-the-basics-b0ff874a3915?source=rss-2ae1c32a010------2</link>
            <guid isPermaLink="false">https://medium.com/p/b0ff874a3915</guid>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[knowledge-management]]></category>
            <category><![CDATA[retrieval-augmented-gen]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Praful Krishna]]></dc:creator>
            <pubDate>Mon, 24 Mar 2025 02:43:36 GMT</pubDate>
            <atom:updated>2025-03-24T02:43:36.090Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>While RAG offers a straightforward approach to leveraging LLMs with organizational knowledge, enterprise implementations face four unique challenges that require thoughtful consideration.</em></h4><p>Retrieval-Augmented Generation combines the reasoning capabilities of large language models with targeted information retrieval from an organization’s knowledge repositories. This integration grounds AI responses in authoritative sources, enhancing accuracy and relevance in ways that standalone models cannot achieve. For enterprises, RAG represents a compelling entry point into the generative AI landscape. It builds upon existing knowledge, reduces the need for extensive model fine-tuning, requires relatively modest technical expertise compared to alternative approaches, and delivers immediate value through improved information access and knowledge work automation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pqIbc72RODXF3zxr" /><figcaption>Photo by <a href="https://unsplash.com/@flyd2069?utm_source=medium&amp;utm_medium=referral">FlyD</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Yet we must recognize that enterprise environments introduce complexities that demand a more sophisticated implementation strategy. The stakes are simply higher in business contexts, where information quality directly impacts decision-making, compliance, and ultimately, organizational success.</p><h4>1. Verifiability: The Foundation of Trust</h4><p>In enterprise settings, the cost of inaccurate information extends far beyond minor inconvenience. Incorrect outputs can lead to flawed business decisions, compliance violations, legal exposure, reputational damage, and wasted productivity as employees validate AI-generated content. To address these concerns, we must prioritize verifiability in our RAG implementations. This means developing systems that allow users to trace assertions back to their source snippets and documents effortlessly. Effective verifiability features should include direct citation of source materials with links to originals, confidence indicators for different content segments, clear differentiation between retrieved information and model-generated content, and comprehensive audit trails.</p><p>Transparent access to the underlying documents that inform AI responses builds essential trust with users while providing necessary protection against potential misinformation. When we implement these verification mechanisms, we transform RAG from a convenient tool into a trusted advisor.</p><p>Effective RAG solutions should also integrate seamlessly with existing systems for knowledge management, document storage, and information retrieval. This integration reduces adoption friction and leverages previous investments in information architecture.</p><h4>2. Hallucination Warning: Acknowledging Limitations</h4><p>Despite our best efforts, we must accept that no RAG system can guarantee perfect accuracy. Large language models occasionally generate content that appears plausible but lacks factual foundation — a phenomenon commonly known as hallucination. In enterprise contexts, undetected hallucinations represent a significant risk that we cannot afford to ignore.</p><p>Responsible RAG implementations require robust hallucination detection mechanisms. These include automated fact-checking against retrieved documents, clear indication when generated content lacks supporting evidence, real-time confidence indicators, and thoughtful fallback procedures that defer to human judgment when confidence is low.</p><p>These warning systems serve as essential guardrails, helping users distinguish between verified information and content that requires additional scrutiny. By acknowledging and flagging potential hallucinations, we reduce the risk of decisions based on fabricated information while maintaining the system’s overall utility.</p><h4>3. Granular Access Control: Respecting Information Boundaries</h4><p>Enterprise knowledge is rarely intended for universal access. Different information types naturally require different permission levels based on security classifications, departmental boundaries, role-based requirements, and regulatory mandates.</p><p>Unlike consumer implementations, enterprise RAG systems must integrate seamlessly with existing access control frameworks to ensure that users retrieve only information they’re authorized to access. System responses must respect document-level and content-level permissions, protect sensitive information even when relevant to a query, maintain comprehensive access logs for security audits, and adapt dynamically to changing roles and responsibilities.</p><p>Without these controls, RAG systems risk becoming security vulnerabilities that could expose confidential information. We must ensure that our implementations maintain existing information boundaries while delivering value across the organization. This balance between accessibility and security represents one of the primary challenges in enterprise RAG deployment.</p><h4>4. Dynamic Content Adaptation: Maintaining Relevance</h4><p>Enterprise knowledge bases are living entities that evolve continuously. New documents are created daily, existing information receives updates, and outdated content gradually loses relevance. An effective RAG system must keep pace with these natural changes to maintain its value over time.</p><p>This capability requires real-time or near-real-time indexing of new content, automatic detection and processing of document updates, thoughtful deprecation of outdated information, and intelligent prioritization systems that favor current information when appropriate. Integration with document management systems becomes essential to track these changes effectively.</p><p>Without these adaptation mechanisms, RAG systems quickly become outdated, potentially leading to decisions based on obsolete information. We should implement comprehensive monitoring to track document freshness and ensure our RAG implementations remain current and relevant.</p><p>Enterprise RAG implementations must also scale to handle enterprise volumes — potentially millions of documents across multiple repositories — while maintaining consistent performance and accuracy. This scalability challenge requires careful attention to system architecture and resource allocation.</p><p>Lastly, we need clear governance mechanisms to determine which information sources to include in the RAG knowledge base, how frequently to update them, and who oversees system performance and accuracy. This governance framework ensures alignment with organizational objectives and compliance requirements.</p><h4>A Path Forward</h4><p>RAG represents a promising entry point for organizations seeking to leverage generative AI, but successful implementation requires more than basic functionality. By addressing verifiability, hallucination detection, access control, and dynamic content adaptation, we can build systems that deliver substantial value while managing the unique risks of enterprise environments. These foundational RAG capabilities will serve as building blocks for increasingly sophisticated implementations.</p><p>Perhaps most importantly, we must remember that technical capabilities alone are insufficient. Enterprise RAG systems require intuitive interfaces that fit naturally into existing workflows, supported by appropriate training and ongoing support. The most sophisticated system will fail if users find it difficult to incorporate into their daily activities.</p><p>RAG can transform how enterprises leverage their collective knowledge. The question is not whether organizations should implement RAG, but rather how they can do so responsibly and effectively.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b0ff874a3915" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Product Design with Generative AI — The HiClass Framework]]></title>
            <link>https://praful-krishna.medium.com/product-design-with-generative-ai-the-hiclass-framework-4affeccce95e?source=rss-2ae1c32a010------2</link>
            <guid isPermaLink="false">https://medium.com/p/4affeccce95e</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Praful Krishna]]></dc:creator>
            <pubDate>Wed, 14 Feb 2024 04:21:54 GMT</pubDate>
            <atom:updated>2024-02-14T04:21:54.396Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Product Design with Generative AI — The HiClass Framework</strong></h3><h4>How to Define Products in a World where 2+2 Can Be Anything between 0 and 100</h4><p>As Generative AI goes through multiple micro-cycles of hype and disillusionment, Product Managers across the globe and across types of companies — Main Street enterprises and tech startups alike — have been asked to build great LLM-based products in the most efficient way. As someone who has been doing this for a bit now, I have found traditional product design frameworks to be helpful but not completely extensible to the new context.</p><p><strong>Identify the Design Factors</strong></p><p>The HiClass framework describes these seven design factors, some of them new, as important to developing products with LLMs in addition to various factors you use in any other product design. Each of these may mean something different for your context.</p><ol><li><strong>H</strong>allucination — LLMs will make things up. What is your strategy to minimize hallucinations or the damage they can cause?</li><li><strong>I</strong>naccuracy — No AI system is perfect. How do you handle inaccurate answers? How do you know if an answer is inaccurate?</li><li><strong>C</strong>ost — What is the cost per transaction of your product: Is your product viable? At least is there a path to viability? Equally important, what is the cost (and time) to get the product up and running?</li><li><strong>L</strong>atency — Esp if you are using chain of prompts or a more advanced prompting scheme, it may take seconds for LLMs to generate output. What is your user doing in the meanwhile?</li><li><strong>A</strong>ttributability — If you need to, can you explain the source and/or reasoning the LLM used to take the action it took? Can you produce the evidence to assure your user that your product is not hallucinating?</li><li><strong>S</strong>ecurity — What are the possible points of leakage of your and your users’ data? What are the various risk scenarios?</li><li><strong>S</strong>tyle — For generative applications, how are you making sure that the generated content is in sync with the tone of your brand, and is generally positive about your brand (believe it or not, LLMs may end up advocating for your competitors if left alone)?</li></ol><p><em>Note 1</em>: It’s important to differentiate between hallucination and accuracy. If a system hallucinates, i.e. makes things up that are not true, then it’s definitely inaccurate, however there could be many other sources of inaccuracy for an LLM system. As a PM actually you do want a small level of hallucination because its control is the same as creativity, and creativity is what gives your bots personality.</p><p><em>Note 2</em>: This list excludes Access to Talent as a design factor, which is probably the most important factor as of Feb 2024, but hopefully this is only a temporary situation.</p><p><strong>Prioritize</strong></p><p>Not all factors may be important to your user story. For example if you are building an app to generate automated marketing emails, latency, attributability and security may not be as high priority as the style. If this product involves a human in the loop to edit the AI output, then accuracy and hallucination are not mission critical either. Here is an indicative table to start thinking about this.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jJWnh1CeK1HvLQew9cuVPg.png" /><figcaption>Source: Author. <a href="https://docs.google.com/spreadsheets/d/1UBkXm2Bn7RjQmlZXQXTaUmKPZ0qS2L5QayuEDa4K5-8/edit?usp=sharing">Original Sheet</a>.</figcaption></figure><p>The more factors survive this stage, the harder you will have to work as a Product Manager before you can launch something. It’s a good idea to rank every surviving factor, identify ideal goals and draw red-lines below which the product will not be acceptable.</p><p><strong>Innovate and Trade</strong></p><p>Now is when brilliant PMs start to shine. I had written earlier that in general <a href="https://praful-krishna.medium.com/creativity-is-overrated-and-misapplied-797fd107aa5c">creativity is overrated and misapplied</a> with a couple of exceptions. This scenario is a big one in the exception category. I always ask myself two questions:</p><p><strong>One</strong> — can I innovate on the product to remove one of the constraints completely. For example, if latency is a big concern due to a chain of prompts, can we design the system not as a bot, where the expectation is for instant messaging, but as a multi-page questionnaire. Now, you can still have dynamic questions leading up to the outcome but each transition gives you breathing space in terms of user expectations. Perhaps you can add a progress bar and a gif showing how hard the system is working in the back end. Or you can keep going with an anticipated workflow, and correct as and when an async signal tells you to. In any case, latency is no longer a huge concern and you can pay some of it to reduce cost.</p><p><strong>Two</strong> — what are the tradeoffs between the various dimensions of HiClass, and where is a good equilibrium. This is going to be iterative. Given the complexity, it’s also important to do a lot of research upfront because the iterations may be expensive. Here is a chart to remind you of how the different factors of HiClass affect the others (if you improve a row, the cells indicate what happens to the columns).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/798/1*InrW-6m1SR367aQ0aWnyUA.png" /><figcaption>Source: Author. <a href="https://docs.google.com/spreadsheets/d/1UBkXm2Bn7RjQmlZXQXTaUmKPZ0qS2L5QayuEDa4K5-8/edit#gid=1146781353">Original Sheet</a>.</figcaption></figure><p>Hope this is helpful. As a parting comment — please remember that the architecture of the rest of the system is as important for an AI product as the AI itself. Please don’t ignore it!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4affeccce95e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Creativity is Overrated and Misapplied]]></title>
            <link>https://praful-krishna.medium.com/creativity-is-overrated-and-misapplied-797fd107aa5c?source=rss-2ae1c32a010------2</link>
            <guid isPermaLink="false">https://medium.com/p/797fd107aa5c</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product]]></category>
            <category><![CDATA[saas]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Praful Krishna]]></dc:creator>
            <pubDate>Sat, 13 Jan 2024 19:20:57 GMT</pubDate>
            <atom:updated>2024-01-13T19:20:57.513Z</atom:updated>
            <content:encoded><![CDATA[<p>There seems to be a notion that creative people make better product managers. That they can just think of a wonderful product and build it.</p><p>It’s partly true. Creativity is a wonderful, very-human, trait that can help anything and everything. Any task can be done better by using some out-of- the box thinking and continuously innovating to see if better solutions are available.</p><p>However, I think creativity is <em>overestimated</em> and <em>misapplied</em> in the product management process. Let me explain both.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HkSC4Mrh4aRRhfngwmRG-g.jpeg" /><figcaption>Image Courtesy: <a href="https://www.pexels.com/photo/top-view-photo-of-people-near-wooden-table-3183150/">Pexels</a></figcaption></figure><h4><strong>Overestimated Creativity</strong></h4><p>Consider the problem of what to build? Let’s assume that the seemingly mundane process of prioritization of ideas and their orchestration is robust enough. If so, all you have to do is to fill the funnel with the universe of all possible ideas related to that feature or product. The process itself will direct you to the best possible outcome. So how to you generate these hypothesis to take through the process? Some will say that the most efficient way is to think of the things that would make for a great product and focus on those things.</p><p>I humbly submit that user interviews, customer feedback, feedback from sales and support teams, competitive analysis, a rigorous analysis of user modes, user persona and user stories, and other commonly used techniques are better sources of these hypotheses. Creative ideation definitely has a role to play, but I have seen people come up with ideas, get attached to them, assume that these are the only or the best ideas available, ignore the proper analysis and end up with subpar products. I call this Smart Person Fallacy (see the second point in the<a href="https://praful-krishna.medium.com/mistakes-i-have-made-in-product-management-325d02e3727d"> list of mistakes I have made</a>).</p><p>On the other hand if you are speaking to enough people and are doing enough research then some pesky notions will refuse to die, and they may end up being the critical thing that makes your product successful.</p><p><strong><em>Long Tail Ideas</em></strong></p><p>A big source of such hypotheses that many PMs, yours truly included, largely ignore is managing the long-tail ideas. How many times have you met a colleague who has some random suggestion about how you could do your job better? Maybe you’re focused on something else at that time, or maybe you are at the stage where that particular idea may not be useful immediately, but there’s always something to those ideas. Similarly, how many times has a customer told you what they would like to see in the product? Again, those things may not be on your roadmap in the short term, but they can be very valuable. Then there is the reviews your customers leave for your product — they are literally telling you what you need. Alas, most teams I have been part of are not equipped to handle qualitative reviews.</p><p>In this era of generative AI, the long-tail nature of such feedback and its fragmentation can no longer be an excuse. In fact it is my personal growth target for 2024. Even without access to fancy AI tools, a simple Google Form often does wonders.</p><h4>Misapplied Creativity</h4><p>So is there any room for creativity? Oh, of course. A huge one. While most people think of creativity as the force behind product features or UX designs there are two areas where it can be applied much better.</p><p>The most important and the most frequent decisions a product manager makes are choices, mostly binary. Which of the two features is a higher priority? Which solution should they try first? Which persona should they build for? In many cases the choices are really obvious e.g. let’s go after $1 billion TAM vs. a $10 million one, but often the choices are hard. People have written tomes on how to make these hard choices and this is truly where a good product manager’s skill outshines.</p><p>With experience , though, PM’s learn that hidden somewhere between the two hard choices is always a third one that is better than either on the table. All it takes is some deliberate thought and, you got it, creativity. A team with consistent commitment to creatively expanding the scope of available solutions will always outrun others in discovery, innovation and execution.</p><p><a href="https://praful-krishna.medium.com/mistakes-i-have-made-in-product-management-325d02e3727d">Mistakes I Have Made in Product Management</a></p><p>There is another way creativity can be useful: in prioritization. Here’s how. Old school product management says, very rightly so, that plot every task on an Effort-Impact axis and then knock off Low-High ones. Next, think about relative importance of all the Low-Low and High-High tasks before you look at Low -High ones. This is a time tested way to do it.</p><p>Great PMs, however, follow a principle — there should be no High Effort tasks in your list, period. (see <a href="https://praful-krishna.medium.com/mistakes-i-have-made-in-product-management-325d02e3727d">2x2 Prioritization is All Wrong</a>, the eighth point). The ideal scenario is to break down all High Effort tasks to a series of Low Effort ones, so that you can deliver something every sprint. This simple idea takes a lot of innovative thinking in real life. To come up testable deliverables for every sprint for a monolithic task is deceptively complex, and this is where creativity helps a lot.</p><p>If you got anything from this article, it’s that don’t trust single opinions, not yours, definitely not others’, and instead rely on data and a wide range of inputs. In my opinion, though, creativity has a big role to play in the product process, just not in the way many would assume.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=797fd107aa5c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>