The Future of Knowledge Management

Companies are successful in part because of their earned knowledge and the way they deploy it in the world, yet our knowledge management systems are bad in the best case and catastrophically awful at large organisations and in governments.

Poor management of knowledge means that people repeat work already done by others, waste time searching in the dark across multiple platforms for what they're looking for, and use us valuable co-worker time asking the same questions again and again.

The volume of data inside organisations is increasing rapidly, and as we integrate agentic machine intelligence into our knowledge creation processes, this problem is only going to get worse.

Knowledge management is a process that includes the aggregation, parsing, semantic interpretation, locating of existing relevant knowledge, comparison, insertion and updating of knowledge, and finally presentation as required.

Today, language models are capable of automating much of this work, and even in cases where they struggle, it is still the case that the verification of knowledge is easier than its initial production, meaning machine reasoning integrated with human oversight can enable more effective knowledge management today.

This software doesn't exist. Traditional systems are based around the human production, management and consumption of knowledge. I think that the dynamic will inevitably shift to humans and machines producing knowledge in tandem, machines managing the knowledge with oversight from humans, and both humans and machines consuming the resulting knowledge.

It is my belief that a non-trivial proportion of web traffic in the future will be semi-autonomous agents pursuing long-horizon tasks. These tasks could include anything from market research, competitor analysis, appointment management, trip booking and much more. However, utilisation of agentic AI in the real world will require orchestration of general agents by humans, initially in to a large extent, and in the future much less. Whilst we have control over these systems and they act in our interest, will always want to have the final say in decisions made by our agents.

In order to be able to manipulate and orchestrate these systems, we will initially need to integrate them as reasoning engines inside larger, traditional and deterministic software systems. In the same way that Kubernetes orchestrates containers, so too will we need to be able to manage our agentic intelligence as it processes tasks.

As we progress into the future, much of this scaffolding will be effectively integrated as learned circuits embedded in the neural structure of the model itself, but for now, it will need to be programmed.

But, for the forseeable future, I think agentic, general intelligence will work most effectively as reasoning engines in the confines of traditional software scaffolding.

These agents need an AI-first knowledge base to report back to, where human operators can have final say over results and updates.