Beyond the digital exhaust
Summary
LLMs promise to make sense of a company’s unstructured information and surface meaningful information to each user. But a KM strategy that relies only on LLMs is incomplete. KM leaders and CTOs must go beyond point-to-point collaboration tools, and decentralise knowledge curation. Having a skilful knowledge management organisation, is also critical.
A version of this article also appeared in Reworked magazine.
Over the last few months, ever since Open AI and ChatGPT took the world by storm, I’ve had a few chats with leaders about how they see large language models (LLMs) revolutionizing knowledge management. The promise is alluring, and relies on a fundamental concept — that of “digital exhaust.” Digital exhaust, a term that’s been around for a while, describes all the data that individuals generate through all their digital interactions and activities.
Now here’s the idea. Hold your breath. No, don’t. You’ve probably heard this before. What if we could take all the digital exhaust of a company and train an LLM on it? We could turn that exhaust into shiny diamonds of knowledge! It’d be amazing, right?
The promise of LLMs for knowledge management
While a company’s digital exhaust is no doubt valuable, I’m less optimistic about such promises. They ignore what has made GPT and Gemini successful in the first place and also, some fundamentals of knowledge creation. So here, I want to challenge this idea of mining the digital exhaust using LLMs and suggest an alternative approach.
How we create knowledge
Experienced knowledge management professionals will be familiar with the SECI model of knowledge conversion. The abbreviation expands to four modes in which we transform any knowledge we have access to.
Socialization (tacit to tacit). In this first stage, we share knowledge with peers. In companies, such knowledge sharing happens through mentorship, coaching, apprenticeship or communities of practice. On the internet, we use social networks to facilitate such knowledge sharing. In the last decade or so, many companies have employed enterprise social networks to mimic how social networks behave.
Externalization (explicit to tacit). Tacit knowledge becomes explicit when someone curates it into a form that others can consume. On the internet, this is a decentralized, permissionless process. Someone like me can put together a website like asyncagile.org with little or no friction. Thousands of people can band together, to create a resource others can consume like Wikipedia. Each of the 1.1 billion websites out there is an act of decentralized curation. Indeed, Andrew McAfee’s Enterprise 2.0 model sought to model this same permissionless model of knowledge curation within corporations.
Combination (explicit to explicit). Knowledge is rarely static. We piece together many pieces of knowledge to create something new. This is what I’m doing right now, as I write this article. It’s what Google does in the background when its featured snippets show you information from its Knowledge Graph. It’s what ChatGPT or Gemini do in response to your questions. But guess what feeds this intelligence? The big, broad internet serves as the knowledge corpus for LLMs or search engines. We can’t ignore this fact.
Internalization (explicit to tacit). No knowledge exists for its own sake. People use explicit knowledge for real-world objectives and learn new things from experience. When they reflect on their experiences, they can make connections across various tasks they’ve performed. This experiential learning and reflection again transforms explicit knowledge into tacit. Some of that tacit knowledge could sit in people’s heads. Other bits can make their way into informal presentations, documents, chat messages and other content that teams share amongst themselves.
The SECI model of knowledge conversion
This process of knowledge conversion repeats itself over and over like a spiral. For sure, search, artificial intelligence and now LLMs play a big part in the “combination” stage, but they depend on a network of good old, structured websites to work effectively. Otherwise, it becomes a case of “garbage in, garbage out.” Within corporations, no number of LLMs will save your cheese, if you don’t focus on the unglamorous work of creating structured knowledge. And if you imagine your corporation being a micro-model of the internet, you must create frictionless, permissionless ways for your employees to curate and structure knowledge. This is what the cheerleaders of “LLMs for knowledge management” conveniently ignore.
A realistic model for knowledge management
However alluring LLMs may seem, we still need the conventional foundations of enterprise knowledge management to fire. I’ve written about this topic extensively, but here’s a three-point, prescriptive summary.
Go beyond team collaboration tools. The COVID-19 pandemic induced a remote work revolution and ensured that almost all teams have access to modern collaboration tools like Teams, Zoom, Workspace by Google or Office 365. But, many companies still ignore platforms that help individuals network across teams, much like they’d do on LinkedIn and Facebook. In companies where such platforms exist, they often suffer the poor usability and high friction that enterprise software is notorious for. Companies need usable enterprise social networks that provide a delightful user experience to facilitate the socialization phase in the SECI model.
Decentralize knowledge curation. The more permissions people have to seek, the harder it is for individuals and groups in a company to curate knowledge. Not only does it make it hard for these groups to build something as fundamental as a team handbook, but it’s also harder to represent collective knowledge for the benefit of others. Invest in a knowledge platform where individuals and teams can organize what they know, in an information hierarchy that makes sense to them.
Employ machine intelligence, but don’t forget the humans. If you succeed at helping people prop up these decentralized “sites” or spaces on your knowledge platform, you’ll surely benefit from machine intelligence. Be it more conventional enterprise search or cutting-edge LLMs, you’ll benefit from their ability to combine different sources of knowledge to offer you contextual information. This said you can’t do away with humans altogether. You’ll need community managers or practice leads to keep fine-tuning their knowledge bases. Curators at the company level are essential to fine-tune search engines and LLMs, and to ensure that your knowledge platform is “browsable and navigable.” Again, creating this support structure with real people isn’t seductive, but it’s essential!
I understand if you think I’m still stuck in a time warp as I recommend such conventional solutions. But let me be honest. I’m happy to be wrong. If LLMs can be a shortcut where after some initial training, companies can manufacture knowledge with zero effort, then what’s not to like? The way things stand today though, I don’t see such solutions outside the realms of scratchy proofs-of-concept that look great in labs, but have zero utility in global, distributed firms. So unless you’re holding out for a science-fiction future, you may need to sort out the fundamentals of knowledge management first.
In summary, I’d urge CTOs and directors of knowledge management at corporations to think about each stage of knowledge conversion in the SECI model. Even if you read this article, a decade after I’ve written it, the model will probably hold good. If machine intelligence plays a part in each stage, that’ll be amazing, but I’m certain it’ll result from conscious design and not lazy accidents. And while machine intelligence matures, what does your company’s knowledge and collaboration platform look like? That’s a good question to ponder over.