Stop Calling It Automation: AI Is a Knowledge Augmentation Layer, not a Replacement
Every few years, someone in the enterprise reintroduces the same bad idea wearing new clothes—that a piece of technology is finally sophisticated enough to make human knowledge work obsolete. Search engines were going to do it. Then enterprise portals. Then chatbots. Now, it’s large language models (LLMs), and this time, the claims come with a demo convincing enough that boards are writing checks before anyone has asked what the technology is actually for.
I think that’s a mistake, and I think it’s a mistake we’re about to make at scale. Not because the underlying models are weak—they aren’t—but because “replace the knowledge worker” is the wrong target. The right target is narrower, harder to put on a slide, and far more valuable: Use AI to make an organization’s existing knowledge legible, retrievable, and usable at the moment someone needs it. That’s augmentation, not automation, and the distinction is not semantic. It determines whether these systems earn trust or destroy it. Organizations keep buying AI to eliminate knowledge work. They should be buying it to save the knowledge work they already have.
What AI Actually Changes
Traditional KM ran on keyword search and folder structure, and it was honest about its limits: If you knew what to search for, you could probably find it, and if you didn’t, you were out of luck. LLMs break that constraint. They can retrieve across ambiguity, synthesize across documents, and produce something that reads like an answer instead of a list of links. That is a genuine capability leap, and I don’t want to undersell it.
But it comes with a catch that most rollout plans conveniently skip: The same fluency that makes an AI system useful also makes it persuasive when it’s wrong. A returned document is either relevant, or it isn’t. A generated synthesis can be confidently, articulately incorrect, and most people have no reliable way to tell the difference from the outside. That asymmetry— between how authoritative an output sounds and how authoritative it actually is—is the central problem of AI-enabled KM, and it’s the reason I keep pushing back on the automation framing.
Retrieval-Augmented Generation Is the Right Instinct
If there’s one architectural pattern worth defending in this debate, it’s retrieval-augmented generation (RAG). RAG systems don’t let a model answer purely from what it absorbed during training; they force it to ground its output in retrieved, sourced material that the organization actually controls. That single design choice does more for trust than any policy memo I’ve written, because it makes provenance checkable instead of theoretical. You can log what was retrieved. You can audit what was withheld. You can trace a generated answer back to the document it came from, which is exactly what a subject-matter expert would do if you asked them to defend a conclusion.
Where I part ways with the more enthusiastic adopters is their assumption that RAG solves the trust problem on its own. It doesn’t. It narrows the space in which hallucination can occur; it doesn’t eliminate the space in which a model can still misread, overgeneralize, or blend two retrieved passages into a claim neither one actually supports. RAG is necessary. It is not sufficient.