Knowledge will determine the winners and losers of the AI race. While computing power, microprocessors, data centres and sophisticated models are necessary, they are less and less critical for successful AI deployment. This pivot towards knowledge (not just data!!!) is triggered by three main developments. First, knowledge inputs for AI are gaining relevance as technology per se becomes less scarce, commoditised through widely available open-source and low-cost LLMs. Second, it is knowledge that can make a difference as LLMs hit the ceiling with more processing power, yielding slight improvements, not proportional to the investment. And, third, knowledge can help resolve the current AI disillusionment, as technology, per se, despite massive investment, cannot deliver on many promises, including increasing productivity and reducing costs in businesses and organisations. It is a backdrop for the tech companies’ pivot towards tapping into individual institutional knowledge as critical AI resources. Companies made their opening acts for the AI scramble for knowledge. Instead of just providing AI platforms, OpenAI and Anthropic are developing consulting and deployment services for the AI transformation of businesses and governments. They have also launched a wide range of philanthropic initiatives involving universities, research institutions, and civil society organisations. In this way, OpenAI and Anthropic can get closer to the explicit and tacit knowledge of businesses and organisations: their workflows, professional practices, internal terminology, decision-making processes, and accumulated experience. Two companies are encroaching to institutional systems, dominated by Microsoft and Palantir. Microsoft is deeply embedded in institutional information systems. Palantir established itself early as a provider of data analysis and decision-support infrastructure. Both now face competitors that want to move beyond providing general-purpose AI models and become integrated into the daily operations of companies and governments. The reaction came fast.
Palantir CEO Alex Karp has warned that dominant AI companies could capture too much value from the organisations using their systems, and even shift decision-making power outside those institutions. The warning may be valid, but it is also somewhat cynical: Palantir has long built its business around embedding its technology deeply within the institutional decision-making of governments worldwide. Microsoft CEO Satya Nadella warned that organisations may pay twice for AI: first with money when purchasing the service, and then with knowledge when contributing prompts, corrections, workflows, and insights that increase the system’s value.
He has described this dynamic as a ‘reverse information paradox‘. Traditionally, a seller risks revealing valuable information before a buyer has paid for it. With AI, the relationship can be reversed: the customer pays for the service while simultaneously contributing knowledge to the provider. This is an interesting shift. Until recently, both companies benefited from a widespread lack of understanding about how knowledge moves from individuals and institutions into AI platforms. By warning customers about AI companies, Microsoft and Palantir are sending the message: remain within our technological environment, where your institutional knowledge will supposedly be safer. The emerging AI scramble is therefore not a battle between benevolent and exploitative companies. It is a competition over who will control the technological gateways through which institutional knowledge is stored, accessed, and used. To understand what is at stake in AI race, we must distinguish between data and knowledge. In contemporary discourse, data is the dominant buzzword. Yet, as the familiar data–information–knowledge–wisdom pyramid illustrates, data is only the raw material. Value emerges at higher levels when data is organised, interpreted, and placed in context. When we ask ChatGPT or another AI platform a question, we are not merely submitting text data. We are communicating tacit knowledge, a critical asset for the AI era. Our questions reveal our interests, uncertainties, fears, plans, assumptions, and gaps in understanding. Questions often disclose more about us than answers do. The same applies to the use of AI by businesses and institutions. Their knowledge does not reside only in databases, documents, and archives. It also exists in routines, professional judgement, informal practices, institutional memory and the experience of employees. Much of this knowledge is tacit. It has never been fully documented. Yet, it can gradually become visible through repeated interactions with AI systems: the questions people ask, the answers they correct, the emails they draft, the exceptions they explain, and the choices they make. Beyond operational relevance, knowledge also defines us as human beings. It carries both a legacy of thinking accumulated over thousands of years in books, traditions, and institutions, and a living resource created through daily encounters, conversations, and acts of judgement. Thus, the current AI scramble for knowledge has far-reaching effects on our society, beyond conventional debates about data ownership or copyright protection. The answer to this question should be the first step in designing AI transformation strategies, especially when it is provided by external companies and consultants. The answer and the full grasp of our AI changes can be gained by answering the following basic questions: External AI platforms create additional client dependency on AI agents, interfaces, proprietary tools, and accumulated workflows. The AI provider may become the indispensable gateway through which institutional knowledge is accessed and applied. Thus, instead of just asking the typical broad question: Is our data being used to train the model? We must also ask: Do we retain control over taxonomies, ontologies and learning processes that make our knowledge accessible and useful? This is not a theoretical question. It concerns system architecture, contracts, data flows, access rights, retention policies, interoperability, and institutional strategy. No. There are feasible options! For example, bottom-up AI, designed around the needs and knowledge of a particular institution, is technically feasible, increasingly affordable, and ethically desirable. The first common-sense question is whether we need super-big models for our daily and routine tasks. Much of our daily work relies on a relatively limited vocabulary of approximately 2000 words, and 20 thinking tools such as induction, deduction, association, and reasoning patterns. Using huge LLMs for such tasks is like trying to kill a mosquito with a cannon. When institutional knowledge is properly organised through taxonomies, ontologies, metadata and clearly defined relationships, smaller language models can compete and, often, outperform much larger general-purpose models. In-house infrastructure can combine language models with databases, knowledge graphs, retrieval systems, and rule-based components. In such a context, the role of LLMs, which can be small, is more to formulate the syntax of textual answers than to conduct profound reasoning that requires big models. The AI relevance shifts from focus on LLMs to, sometimes, non-technical, organisation of our knowledge, with the following additional gains: Overall, smaller and more focused AI solutions can typically provide answers that are more accurate, grounded, and contextualised than big LLMs covering all domains of human knowledge. The fact that we have to ask this question is even more absurd, knowing that affordable alternatives exist. Open-source models are becoming increasingly capable. Software development is more accessible. AI-assisted coding has reduced many of the technical barriers to building tailored applications. For numerous institutions, the main obstacle is no longer technology, and perhaps not even funding. The deeper obstacles are institutional and psychological. We expect AI to provide a quick, almost magical solution. We imagine that we can upload thousands of documents, connect a chatbot and instantly acquire institutional intelligence. Real knowledge work is a slower, but more rewarding final ‘refuge’ in our race with machines. It requires institutions to examine what they know, how that knowledge is organised, and which parts remain useful. It requires taxonomies, ontologies, metadata and editorial judgement. Furthermore, it requires conversations with experts and attention to tacit knowledge. It also forces institutions to confront uncomfortable realities: duplicated documents, outdated procedures, contradictory terminology, forgotten experience and knowledge trapped in organisational silos. Think of building useful AI systems inside an organisation like looking in a mirror. The organisation needs to understand what it knows and what it doesn’t. If we don’t know what knowledge we have, AI can’t take over our thinking. Instead, AI serves as a helpful tool to help us find and organise our knowledge. Purchasing access to a powerful external platform appears easier than undertaking this internal work. But today’s convenience can easily slide into tomorrow’s risks and dependencies. At Diplo, we are trying to walk the talk by demonstrating that bottom-up AI is possible even for a small organisation. Having no dedicated AI budget has, paradoxically, become a kind of collateral advantage. Without funds, we have not been tempted to seek solutions by acquiring cutting-edge platforms or purchasing computing power. We have therefore had to focus on what has been available to us: our knowledge, experience, and creativity. Concretely, we have invested time in organising material on diplomacy and digital policy that has been gathered over several decades. It includes development of taxonomies, ontologies, metadata, and knowledge structures that allow our institutional knowledge to be found, connected, and reused. With a solid knowledge structure, we are less dependent on external LLMs. The scramble for knowledge has already begun. We should avoid both technological isolation and the naive denunciation of major AI platforms. ChatGPT and similar systems have made extraordinary contributions to the accessibility and use of knowledge. But admiration should not block scrutiny and search for greater business, technical, and institutional clarity. Concretely speaking, our preparations for the intensification of AI scramble involve two immediate steps. First, we must understand what happens whenever we interact with an AI system: what is stored, what may be reused and for what purposes, what controls are available and who benefits from the resulting learning. Second, companies, governments, universities, and civil society organisations must begin treating their knowledge as a critical strategic resource. This does not mean that uploading our knowledge to big AI platforms is unacceptable. Some individuals and institutions may be perfectly comfortable exchanging parts of their knowledge for access to convenient and powerful services. But this should be an informed and conscious choice, not the accidental consequence of technological enthusiasm, unclear contracts, or organisational sleepwalking. Chips will become faster. Models will become larger. Platforms will rise and fall. Institutional knowledge – accumulated through years of experience, mistakes, conversations, and human judgement – will remain constant despite technological changes. In the AI era, the decisive question will not simply be who develops the most powerful model. It will be who manages and controls the knowledge that makes AI useful and impactful.Entering the institutional knowledge business
AI is about knowledge beyond data
What happens to our knowledge?
Do we have to slide into knowledge dependency?
Why are we sleepwalking into AI dependency?
DiploAI and the collateral advantage of limited resources
Preparing for the AI scramble
In the end, knowledge will prevail
