A while back I connected a Google Drive to an AI tool to see what it could do with our content. No curation, no preparation. I just pointed it at the Drive and let it loose.

The results were instructive, though not in the way I'd hoped.

There were 27 versions of a pitch deck in there. Copies of copies of copies, each one tweaked for a specific use case and never retired, because nobody had established which version was the one that should actually exist. The original template was in there somewhere. So were all the departures from it. Draft documents that had never been marked as drafts. Outdated process notes sitting alongside current ones with no way to tell which was which. The AI did exactly what you'd expect. It drew on all of it, confidently, and produced responses that were a blend of where we were now, where we'd been two years ago, and a few things that had never been true at any point.

It wasn't hallucinating. It was reading. The problem was what it had to read.

For as long as I've worked in learning, the dominant infrastructure conversation has been about platforms. Which LMS to buy. Whether to move to an LXP. How to consolidate the stack. Those aren't bad conversations. But they're conversations about containers. And a container is only as useful as what you put in it.

Most organisations have spent years optimising the container and paying almost no attention to the contents. The result is learning infrastructure that looks organised from the outside: a tidy platform, a structured catalogue, a sensible navigation. Underneath, it is quietly chaotic. Untagged content. Duplicate assets. Outdated material that nobody has got around to removing. Documents that were drafts and got treated as finished.

I've watched organisations spend six figures on a new platform, migrate everything across, and end up with exactly the same problem in a shinier container. The platform changed. The content didn't. Nothing travelled any better than it did before.

I've lived this more directly than I'd like. When we migrated from one platform to another, we hit the content before we could even start making decisions about what to move. No author tags. No last edit dates. No version numbers. Courses that could have been built by anyone, updated at any point, for reasons nobody had recorded. We spent some time untangling it before a single piece of content moved anywhere. The migration wasn't the hard part. The audit that had to happen first was.

As Courtney Sembler, Head of Zoom's Learning Centre, recently put it over on LinkedIn:

The platform is not the problem. It never was.

Some content is designed to be completed. It has a beginning, a middle, and an end. Someone works through it, demonstrates something, and a record is created. That's one category.

Everything else is knowledge. The process documentation, the product guides, the how-to articles, the best practice notes, the case studies, the answers to questions people ask every week. Knowledge exists to be found and used at the moment someone needs it, not worked through from start to finish.

When you build knowledge like a course, you make it unfindable at the moment it matters.

You give it a linear structure it doesn't need, lock it inside a platform that wasn't designed for retrieval, and format it in a way that makes it hard for AI to do anything useful with. The two categories need different structural logic. Mixing them up is how you end up with an AI that confidently gives someone a two-year-old answer because that answer was formatted like a course and sat at the top of the search results.

Anyone who has worked with data will recognise this problem immediately. A data analyst handed a clean, well-structured dataset can do remarkable things with it. The same analyst handed a folder of spreadsheets with inconsistent column headers, duplicate rows, figures in different currencies, and no documentation of what any of it means will produce something that looks like analysis but can't be trusted.

The output is confident. The conclusions are wrong. And the analyst isn't to blame. The infrastructure is.

Garbage in, garbage out. The difference with AI is that it gets it wrong at scale, with confidence, and the person on the receiving end has no way of knowing.

This is not a technical problem. The most sophisticated AI ingestion system in the world cannot compensate for knowledge that is duplicated, contradictory, or two years out of date. The pipeline is not the issue. What flows through it is.

This is not an argument for a six-month content audit before you're allowed to touch AI. It is an argument for being deliberate about how you build from here, and for knowing what you're dealing with in what you already have.

Clean knowledge infrastructure has four characteristics.

It is modular. Each piece of content answers one question or covers one topic. It can be used on its own, combined with other pieces, or surfaced in isolation depending on what someone needs. A large course built as a single unit looks efficient until something changes, and something always changes. Then you're rebuilding forty minutes of content to fix one outdated process. Modular content means you update the one piece that broke, not everything around it.

It is tagged. Not with a taxonomy that took three months to design and nobody uses consistently, but with enough metadata that the right content can find the right moment. Who it's for. What it covers. When it was last reviewed.

It has a source of truth. When there are 27 versions of the same thing, none of them is the source of truth. Deciding which one is, and retiring the rest, is unglamorous work. It is also the work that makes everything else possible. Without it, you are not managing knowledge. You are managing confusion.

It is maintained. This is the one most teams skip, and it's the one that quietly poisons everything else. An outdated how-to guide doesn't just sit there harmlessly. It gets surfaced, it gets used, and it sends someone in the wrong direction with complete confidence. A review cycle doesn't need to be complicated. It needs to exist, and someone needs to own it.

The piece I published last week was about context architecture. The personal, business, and environment layers that tell AI who it is talking to and what matters to them. This is the other side of the same idea. Context tells AI about the person. Knowledge infrastructure gives it something worth saying.

When both are in place, AI can surface the right asset, in the right format, through the right channel, for the specific person asking the question, whether that is a chatbot, an AI tutor, a CSM recommendation, an in-product tooltip, or something that doesn't exist yet.

Most organisations have no idea how tangled their content actually is. There's never been a reason to look closely. AI is the first thing that makes the mess visible, not by fixing it, but by trying to use it and failing.

The first time you point AI at your knowledge base and it gets something badly wrong, don't blame the AI. That's the moment your content infrastructure finally introduced itself.