Rethinking Knowledge Management Principles for an AI-Powered World
by Faisal Hoque, Thomas H. Davenport, and Paul Scade
Column Editor’s Note: The renewed interest in knowledge management reflects an important shift in how organizations think about the role of generative AI in creating value from information. Rather than assuming that broader access to more content will automatically improve performance, this article argues that successful knowledge management begins with clearly defined business problems and carefully designed knowledge flows. In this HDSR Active Industrial Learning column article, the authors revisit the long-standing distinction between knowledge stocks and knowledge flows and show why generative AI, despite its powerful capabilities, may amplify old access-first mistakes if it is not deployed with greater discipline. Through examples from SafeSide Prevention, CMS, PepsiCo, and Robinhood Markets, the article illustrates how targeted use cases, curated knowledge bases, retrieval-augmented generation, workflow integration, and human governance can turn organizational knowledge into measurable outcomes. At the same time, the article offers a balanced view of generative AI’s contribution, recognizing its ability to search fragmented repositories, summarize content at scale, work across languages, and support more natural interaction with enterprise knowledge. By combining practical cases with five success patterns for AI-enabled knowledge management, the piece provides a timely and useful contribution for leaders seeking to move from information access to business value.
Introduction
A decade ago, one of this article’s authors—Tom Davenport (2015), an early advocate of knowledge management—declared that the movement was “gasping for breath.” After 20 years of ambitious initiatives and major investments in software and knowledge managers, organizations had proven largely unable to solve a seemingly simple challenge: getting the right information to the right people at the right time. The field that once promised to transform how organizations capture and deploy their collective intelligence had become, in Davenport’s assessment, largely dormant. Instead, executives had shifted their focus to the newly fashionable fields of data analytics and automated decision tools, while the maturing of internet search engines made external search easier than trawling through organizational knowledge repositories. The need to master internal knowledge remained, but the technology and methods used in traditional knowledge management were not up to the task.
Today, generative artificial intelligence (AI) appears to have raised knowledge management (KM) from its deathbed. Large language models (LLMs) can search unstructured data, connect fragmented repositories, and deliver answers in plain language—solving many of the technical challenges that plagued earlier KM initiatives. Microsoft’s Copilot claims that with its software “knowledge flows freely across the organization” (Spataro, 2023) while vendors everywhere tout AI-powered systems that promise to finally realize the dream of democratizing institutional knowledge.
Many organizations are augmenting the general knowledge within LLMs with their own internal content (Davenport & Alavi, 2023), typically using a technique called retrieval-augmented generation, or RAG. The tools that KM practitioners could only dream of 20 years ago now exist, and for many purposes they work remarkably well. In addition to providing a conversational interface for searching knowledge stores, LLMs can also assemble content across fragmented repositories, make recommendations about how to address issues raised in search, work across many languages, summarize content at scale, and even generate code to analyze structured data.
Yet this technical revolution masks a deeper problem that AI by itself cannot solve. The failure of knowledge management was never really about technology; it involved a fundamental misunderstanding about how knowledge creates value. Organizations have long assumed that maximizing access to information automatically improves decision-making and performance. But access without context creates noise, not knowledge. Making everything available to everyone, whether through SharePoint repositories or AI-powered chat interfaces, simply accelerates the delivery of purposeless data. As the late Larry Prusak, one of Davenport’s coauthors (Davenport & Prusak, 1998) and an influential thinker on knowledge management, often warned, separating knowledge from its uses and emphasizing knowledge stocks over flows are among the deadliest sins of the field (Fahey & Prusak, 1998).
Generative AI, for all its capabilities, threatens to amplify rather than resolve this fundamental sin. The organizations succeeding with modern knowledge management – whether AI-powered or not—have abandoned the access-maximization paradigm entirely. Instead, the key to delivering value is to focus on solving specific business problems through targeted knowledge flows that are integrated with users’ workflows.
Purpose-Driven Knowledge Management
AI, and LLMs in particular, have breathed new life into knowledge management by resolving many of the technical challenges the field has faced in the past. But solving the technical challenges is not enough; KM initiatives will still fail to deliver unless they are tightly focused on solving specific problems.
The fundamental flaw in most knowledge management initiatives is that they begin with asking, ‘How can we maximize access to our organizational knowledge?’ Instead, leaders should ask ‘What specific enterprise problems need solving, and could better knowledge flows help solve them?’ When KM projects start from this strategic understanding, they naturally align with organizational priorities: a pharmaceutical company might focus on accelerating drug discovery by connecting research teams with historical trial data; a consulting firm might prioritize capturing and deploying client insights to win new business; a manufacturing company might concentrate on preserving technical expertise as senior engineers retire. Each of these initiatives emerges from specific strategic needs rather than from a generic desire to ‘manage knowledge better.’
Rather than implementing universal repositories that promise to make all information accessible to everyone, strategically aligned KM initiatives work backward from desired outcomes. They map the specific knowledge flows required to achieve strategic objectives, identify where current processes break down, and design targeted interventions that integrate seamlessly with existing workflows. This means asking pointed questions: Which decisions would improve with better knowledge access? What knowledge is being recreated wastefully across teams? Where do delays in finding expertise create competitive disadvantages? The answers determine not just what knowledge to capture and share, but also its structure, access, and interfaces.
Success metrics flow naturally from this approach—not adoption rates or documents uploaded, but business outcomes like reduced project cycle times, improved RFP win rates, or faster product development. When KM initiatives are designed to meet strategic needs rather than to use technological capabilities, they stop being expensive experiments in information management and become engines of competitive advantage.
Access First Approaches
Many technology vendors lean into the maximal knowledge access vision. When Microsoft’s Copilot launched, it promised to create “a new knowledge model” by working “across all your business data and apps to surface the information and insights you need from a sea of data” (Spataro, 2023). Atlassian, meanwhile, claimed that its Rovo product would give team members “superpowers” by connecting them to vast repositories of institutional data (Atlassian, 2024). The suggestion underlying such claims is that, with KM now enabled by generative AI, organizations can simply point an LLM at their existing data repositories and expect productivity gains as a result. AI will handle the rest, from indexing the collected information to using natural language processing to understand queries and surface answers.
While AI-powered approaches to KM promise efficiency and work quality increases, an Australian whole-of-government study of a Copilot deployment found a more nuanced picture (Digital Transformation Agency, 2024). Though framed as a broad productivity evaluation, the trial assessed many of the knowledge flows that KM systems exist to support. In some cases, the results were clearly positive. Use of Copilot led to significant perceived improvements in summarizing existing information, preparing first drafts of documents, and searching for information. However, its lack of predictability, accuracy, and fit to the users’ specific context impacted overall efficiency gains. A parallel qualitative study conducted by researchers at Australia’s Commonwealth Scientific and Industrial Research Organisation (CSIRO), based on interviews with 27 trial participants, found that users reported problems in areas requiring “domain-specific knowledge, creative problem-solving, and nuanced decision-making” (Bano et al., 2025; cf. Bano et al. 2024). Findings included inconsistent results in “research-related tasks requiring critical thinking and domain-specific expertise” and “[misalignment] with the specialised demands of the user’s role.” These shortcomings contributed to additional workloads, requiring staff to check and correct the suggestions they received. The ultimate result was no clear overall reduction in workload for these more complex tasks, just a shift in areas of effort. This “productivity paradox” (Bano et al., 2025)—where time saved through automation was countered by time spent verifying and correcting outputs—highlights the importance of ensuring that knowledge flows are tailored for specific users and tasks.
Purpose-Driven Knowledge Management in Practice
Rather than starting with technology and searching for applications, some companies began with specific business challenges and worked backward to design knowledge systems that directly address those needs. Their implementations demonstrate that when KM initiatives are anchored in clear strategic objectives—winning more contracts, reducing duplicate work, or accelerating product development—they deliver measurable value. The following examples from SafeSide Prevention, the law firm CMS, and PepsiCo illustrate how this problem-first approach transforms knowledge management from an expensive IT experiment into a driver of competitive advantage.
SafeSide’s InPlace® Learning Assistant
SafeSide Prevention founder Professor Anthony Pisani shared with us how his organization transitioned from an access-first approach to KM to one that focused on meeting specific challenges. SafeSide delivers suicide prevention education at scale, including for very large organizations such as health systems and national militaries. Its core trainings are delivered through a blended learning model known as InPlace Learning in which teams complete video modules together, combining the scalability of online training with the accountability and interaction of group work (Donovan et al., 2023). Because there is no expert instructor present to answer questions during group sessions, SafeSide deployed a traditional KM system to give participants access to its expert knowledge. This included a curated library of searchable text and multimedia files supplemented by twice-monthly live office hours at which participants could directly ask questions of experts. While well received by its users, this approach required participants to remember their questions and access resources outside training sessions.
To address this problem, SafeSide initiated an AI-powered InPlace Learning assistant that sits alongside its group-based training. During and after a video-based InPlace workshop, team members can now ask natural-language questions and receive tailored answers grounded in SafeSide’s expert practice guidance. The data set behind the assistant is highly curated to ensure that the information delivered is narrowly focused, with guardrails in place to discourage conversations outside the scope of the training. At the same time, the LLM’s persona is designed to ensure that answers are relevant to the questioner. The domain-specific corpus includes transcripts from 4 years of twice-monthly Q&A sessions led by clinical experts and trained lived experience advisors, the core teaching modules, and selected components of the previous resource library, labeled and structured to support question-answering in the context of group training. The system uses a RAG approach on top of commercial LLM models to achieve its targeted goals. SafeSide uses an iterative process of testing with nonexpert users followed by expert assessment of answers. Human-revised answers are then used to support reindexing of the knowledge base, leading to increasingly precise outputs. LLM outputs are evaluated using the DeepEval open source evaluation framework, which scores responses on faithfulness, answer relevancy, contextual relevancy, and contextual precision.
Bottom-Up Approaches to Knowledge Management at CMS
John Craske, chief innovation and knowledge officer at CMS UK, and Colin Fisher, head of knowledge, described the approach to AI-powered KM taken by Europe’s largest law firm. Law firms must maintain strict controls on knowledge availability, since many documents belong to clients rather than the firm and much content is privileged rather than freely accessible. These restrictions set inherent limits on a ‘maximal access’ approach to KM. Even in areas where access to broad data is permissible, CMS emphasizes the importance of working with carefully selected data sets. This bottom-up approach ensures domain-specific relevance through careful curation of the data with which an LLM works, rather than seeking to impose order from above with custom indexing tools or extensive fine-tuning of language models. CMS’s approach also shows how knowledge management can solve specific business problems without relying on information retrieval solutions. The firm uses a specialized document review tool to generate draft reviews of lease agreements. CMS’s knowledge base is deployed to create a ‘blueprint’ for use with the tool: a structured set of prompts that directs the copilot to extract defined types of data and metadata from the lease, such as whether the terms are industry standard, whether certain key information is included, and whether the lease is assignable. The role played by the knowledge base here is structural: the firm’s knowledge is encoded into the blueprint prompts rather than being the target for a search tool.
PepsiCo’s Ask Ada Platform
One of PepsiCo’s focused knowledge management efforts involved customer and market insights. In an interview with one of the authors of this article, and also in a book (Phillips et al., 2024) describing the transformation of customer insights at PepsiCo, Stephan Gans, chief customer insights and analytics officer, described a knowledge platform called “Ask Ada” that incorporates results from advertising, influencer, and other types of campaigns, as well as present and past customer insight lessons. Generative AI provides a conversational interface providing access to all Ask Ada content. However, Ask Ada is not just a knowledge retrieval system, but a deep platform for creating and managing customer insights, including capabilities for testing ad-oriented creative content with a customer panel, social listening, predictive modelling of campaign results, and identification of meta-learning across all present and past customer insights. Generative AI can now be the primary interface to all of these features.
Gans credits Ask Ada and associated interventions with helping to create more independence from external agencies and consultants. According to Gans, the platform has generated savings of millions of dollars per year in customer insight spending while also changing the culture of customer insights professionals from ‘order taking’ to consultation with marketing decision-makers.
Many other organizations have created special-purpose knowledge systems using AI, some of which are tied to mission-critical decisions. Robinhood Markets, for example, uses knowledge of financial crimes, external events, and customer trading patterns to streamline the process of investigating potential crimes (Amazon Web Services, n.d.). It uses generative AI to synthesize customer and transactional data across multiple accounts into detailed investigative summaries. Identification of relevant knowledge across systems and data sources is aided by a detailed knowledge graph. Although the final decision on customer fraud issues is still made by humans, an Amazon Web Services case study claims the new system has resulted in cumulative efficiency gains of up to 20%.
Common Success Patterns
KM systems that deliver value share several distinguishing features compared to generic information retrieval tools:
- Strategic problem definition: Successful KM initiatives begin with clear business objectives—winning more contracts, reducing project delivery time, preventing knowledge loss from retirements—and work backward to determine what knowledge flows would address these challenges. This reverses the traditional KM approach of building repositories first and hoping use cases emerge. The breadth of generative AI capabilities means that organizations can accomplish a higher percentage of strategic goals than was possible with previous AI. This breadth will only increase with the widespread adoption of agentic AI.
- Concrete success metrics: Rather than vague aspirations about ‘improving knowledge sharing,’ effective systems define concrete outcomes: hours saved per proposal, reduction in duplicate research, percentage improvement in bid win rates. These metrics directly tie to business value rather than using system usage statistics to assess success. With generative AI it is important not just to measure individual productivity gains, as we seldom know what the user is doing with the saved time. Generative AI is capable of enterprise-level metrics, including improved financial performance.
- Workflow integration: Knowledge tools must meet users where they work. Where possible, knowledge retrieval should not require leaving familiar applications or disrupting established work patterns. Workflow integration also keeps KM initiatives honest: knowing where integration is needed and where it is not helps ensure a tight focus on solving defined problems. Agentic AI—specifically, teams of agents with orchestration—is particularly well-suited to automating or closely augmenting knowledge workflows.
- Context-specific architecture: Instead of generic search across all organizational data, successful systems structure information for particular purposes. This includes domain-specific indexing, specialized semantic chunking that reflects how practitioners actually use information, and interfaces designed around specific tasks rather than general exploration. In most cases, the proprietary content used in RAG-based knowledge systems needs to be carefully curated to ensure quality (Davenport et al., 2025). It is rarely feasible within a large organization to curate all knowledge content anyway.
- Active human governance: Quality control cannot be outsourced entirely to algorithms. Monthly reviews of system outputs, domain expert validation, and continuous refinement based on user feedback ensure that the system remains aligned with evolving business needs and maintains accuracy over time. Close involvement in the management of the knowledge and the system that manages it by the people who work in the relevant area is usually desirable. Again, this success factor will only grow more important with the use of relatively autonomous AI agents.
Not Just Knowledge Management; Not Just AI
Whether organizations frame their efforts as knowledge management, AI implementation, digital transformation, or simply operational improvement initiatives matters less than their approach to the challenge. The principles that distinguish success from failure remain constant: start with specific business problems rather than generic access goals, design systems that integrate with workflows, maintain rigorous human oversight of outputs, and measure success through business outcomes rather than usage statistics.
Instead of maximizing information access, organizations must embrace a more disciplined approach: defining precise objectives, understanding exactly what knowledge flows will achieve those objectives, and building systems—whether AI-enhanced or not—specifically architected to deliver those flows. In an era in which both data volumes and AI capabilities are expanding exponentially, this focused approach to knowledge management will be increasingly critical for organizations that want to convert their information assets into competitive advantage.
Disclosure Statement
Faisal Hoque, Thomas H. Davenport, and Paul Scade have no financial or nonfinancial disclosures to share for this article.
References
Amazon Web Services. (n.d.). Robinhood transforms financial crimes investigations using Amazon Bedrock.https://aws.amazon.com/solutions/case-studies/robinhood-case-study/
Atlassian. (2024, May 1). Unlock enterprise knowledge with Atlassian Rovo.https://www.atlassian.com/blog/announcements/introducing-atlassian-rovo-ai
Bano, M., Zowghi, D., Whittle, J., Zhu, L., Reeson, A., Martin, R., & Parsons, J. (2024). Survey insights on M365 Copilot adoption. ArXiv. https://doi.org/10.48550/arXiv.2412.16162
Bano, M., Zowghi, D., Whittle, J., Zhu, L., Reeson, A., Martin, R., & Parsons, J. (2025). A qualitative study of user perception of M365 AI Copilot. ArXiv. https://doi.org/10.48550/arXiv.2503.17661
Davenport, T. H. (2015, June 24). Whatever happened to knowledge management? The Wall Street Journal. https://blogs.wsj.com/cio/2015/06/24/whatever-happened-to-knowledge-management/
Davenport, T., & Alavi, M. (2023, July 6). How to train generative AI using your company’s data. Harvard Business Review. https://hbr.org/2023/07/how-to-train-generative-ai-using-your-companys-data
Davenport, T. H., Hoerl, R. W., & Redman, T. C. (2025, May 28). To create value with AI, improve the quality of your unstructured data. Harvard Business Review. https://hbr.org/2025/05/to-create-value-with-ai-improve-the-quality-of-your-unstructured-data
Davenport, T. H., & Prusak, L. (1998). Working knowledge: How organizations manage what they know. Harvard Business School Press.
Digital Transformation Agency. (2024). Australian Government trial of Microsoft 365 Copilot. Australian Government. https://www.digital.gov.au/initiatives/copilot-trial/microsoft-365-copilot-evaluation-report-full
Donovan, S., Maggiulli, L., Aiello, J., Centeno, P., John, S., & Pisani, A. (2023). Evaluation of sustainable, blended learning workforce education for suicide prevention in youth services. Children and Youth Services Review, 148, Article 106852. https://doi.org/10.1016/j.childyouth.2023.106852
Fahey, L., & Prusak, L. (1998). The eleven deadliest sins of knowledge management. California Management Review, 40(3), 265–276.
Phillips, S., Barry, R., Gans, S., & Schardt, K. (2024). The consumer insights revolution: Transforming market research for competitive advantage. Rethink Press.
Spataro, J. (2023, March 16). Introducing Microsoft 365 Copilot – Your copilot for work. Official Microsoft Blog. https://blogs.microsoft.com/blog/2023/03/16/introducing-microsoft-365-copilot-your-copilot-for-work/
©2026 Faisal Hoque, Thomas H. Davenport, and Paul Scade. This article is licensed under a Creative Commons Attribution (CC BY 4.0) International license, except where otherwise indicated with respect to particular material included in the article.
[Photo: Adobe Stock]
Original article @ HDSR.
