Hey, Luca here! — welcome to a ✨ monthly free edition ✨ of Hybrid Hacker! Every week we write original articles about engineering management, career growth, and productivity, to more than 160K friends! To read all the editions and join the private community, subscribe to the paid version: Today’s article is written by my friend Matteo Cellini, who writes a great newsletter about the future of work and is obsessed with managing knowledge, which is today’s topic! In Italy we say: “non mettere il carro davanti ai buoi” — don’t put the cart before the horse. The “horse” is AI, pulling that knowledge forward. Without the graph, the horse is powerful but directionless.
This is why AI assistants and agents, despite their promise, stumble on the basics of daily work:
These aren’t edge cases — they’re the fabric of everyday work, and the place where most AI fails. Not because the models are weak, but because organizations haven’t mapped their knowledge. Until knowledge is treated as capital — invested, structured, and maintained — AI in the workplace won’t reach its full potential.
🪨 What’s hard about itMost of the talk right now is about model sizes, benchmarks, and prompts. But inside companies, the bottleneck is fractured knowledge:
An agent asked to help is like someone given a box of puzzle pieces without the picture on the lid: the pieces exist, but the context is missing — and without that, agents end up shallow, no matter how advanced the technology. Work itself is also shifting: as argued in The Decentralised Workforce, teams are more fluid, roles shift constantly, and information gets even more scattered. AI doesn’t fix that. It rather amplifies what Luca described in his article on cognitive debt: the hidden cost of fragmented knowledge and half-finished systems. 🏦 Knowledge as capitalSo, knowledge capital is what your company knows, and how easily that knowledge can be applied. Traditionally it spans:
Especially in an age where information has become abundant and cheap — it’s knowledge, the ability to turn information into decisions, that generates advantage. (for more on this topic, check out the book Knowledge Capital: Organizational Discretional Intellectual Capital). So treating information as a portfolio (not just a pile of documents) is what can turn it into a productivity and competitive advantage. Like financial capital, knowledge capital has returns and risks:
This is why it calls for a concept like knowledge debt to be introduced and measured. Just as engineers talk about technical debt, organizations accumulate knowledge debt — shortcuts and gaps that feel efficient in the moment but create downstream costs. AI can make things worse and amplify that debt: shifting teams into “review mode” (quick edits on AI drafts) risks anchoring thinking unless the underlying context is sound. So to act like real capital, knowledge must be:
🗃️ The Enterprise GraphTo do this, we need to introduce the enterprise graph. This acts as the organizational nervous system that makes knowledge capital usable:
Together they form a living map of how work actually happens — updating as projects shift, collaborators change, and tools evolve. Example — feature request triage in Jira
The result: context surfaces where and where work is happening and can be used to feed assistants, decisions, and build agents that have the right information at hand. Once we build the map (the graph), we need to give agents a way to act. That’s where Model Context Protocol (MCP) comes in. It’s quickly emerging as the standard that lets agents interact with enterprise tools through a common client–server pattern. Instead of one-off integrations, MCP defines simple calls: agents act as clients, systems expose capabilities (search, retrieve, update). This is the way that the various Of course a lot of questions remain — what tools to expose, how to govern permissions, but this could become one of the new types of tasks/roles that can be created to help facilitate and coordinate implementation and evolution of the graph.
❓ But who gets to use it?Another big problem with AI (just like any technology) is adaptation and accessibility: the real opportunity is when both engineers and non-engineers can build and use them. For example: when accessible, HR can build an onboarding assistant, Ops can track workflows, Sales can fork a forecast bot. Engineers stop being the bottleneck, and teams closest to the work design the agents they need. This mirrors what has already happened with AI writing tools: they don’t just generate text, they guide clarity and tone. Agent builders need to do the same — helping non-engineers shape workflows without learning query languages. As we argued in What’s Good for Humans is Good for AI, this is the only way agents can scale and enable further personalization of any solution. 🔨 How to implement an Enterprise GraphTo build the enterprise graph we need to think in stages, where value comes from proving usefulness early and layering complexity over time. Here is what a 90-day plan may look like: 1) 🌱 30 days → address knowledge debt that hurts most
2) 🪴 60 days → expand across teams and signals
3) 🌳 90 days — scale with accessibility and monitoring
I talked about this with the guys at Glean, who helped shape my thinking about this topic. They are building a platform that is trying to solve this, both at enterprise and personal level. They build graphs, on top of which they add functional layers like Search, Agents and Assistants. 📌 Final thoughtsWe already discussed at length how AI shouldn’t be seen as another tool, but a good excuse and technology that helps us rethink how we work — what patterns we reward, how teams learn, how decisions flow. In medium and large organizations, fragmentation and dispersion are the norm everyday bottlenecks. No model will fix that until the infrastructure of work changes. We need to design systems that mirror or capture real human workflows: messy, adaptive, intertwined. The enterprise graph is one foundational move toward that as it starts making context durable, shareable, and actionable across domains. That’s when we gain visibility not only into what teams do, but how they do it, and this matters not only for individuals, but also for team performance. In “There’s No AI in Team” we discussed that AI can’t replace trust and alignment but that it can enforce consistency, detect drift, highlight gaps, and even surface invisible dependencies before they trip someone up. In that environment, performance becomes not just a reflection of good tools, but of shared structure. If we treat knowledge capital as something we invest in — design, measure, maintain — and not just as “data to feed models” we can start evolving new forms of work and coordination. And that’s it for today! I wish you a great week Sincerely
|
The Rise of the Enterprise Graph 🗃️
Thursday, 2 October 2025
Subscribe to:
Post Comments (Atom)
No comments:
Post a Comment