Help your AI help you - clean the corner you're standing in

Organisations diving into the internal use of AI to provide answers often realise that their file system is barely organised chaos, and their document management practices have been missing in action for years. There's a tidy, responsible-sounding answer to the "your documents are a mess, so AI will just confidently lie to you" problem. Fix the documents first. Sort out the knowledge base, retire the old document versions, assign ownership, get it all true, then build an AI bot to rule them all. It sounds like grown-up advice. It's also a brilliant way to do nothing for a year.

I read a piece on our optimisation obsession this week that also applies here. The 80:20 rule, we all know it: somewhere around 80% of anything you get diminishing returns, and the last stretch costs far more than it brings in value. We accept this about exercise and diets. We forget it the second someone says "knowledge management". The dream of a pristine SharePoint or Google Drive isn't a goal, it's the same shiny thinking that says AI fixes everything, pointed at your filing system. The real world is always messy. It was messy before AI, it'll be messy after, and no model is going to tidy years of organisational cruft into a single source of truth.

So the question isn't "how do we clean everything". It's "where's the smallest patch with decent content and a visible payoff". A few places I'd actually start:

One team, not the whole org. Pick a team whose domain is reasonably well documented and understood, give that corner an owner, point AI at that and nothing else. You learn what good looks like at a scale you can hold in your head. You might even be able to write a set of evals that measures accuracy and improvements. You can get a feeling for what kinds of questions people ask, where they struggle.

Find the stuff that's already roughly current. Some teams do daily and weekly updates, they're written content which makes excellent AI fodder, they're recent, and almost nobody reads the ones from other teams. Synthesise them, share the summary, and people suddenly have visibility into work happening three desks away that they'd otherwise duplicate next quarter. Low effort, because the raw material already writes itself.

Onboarding. A new joiner doesn't need a perfect knowledge base; they need a guide through an imperfect one. An assistant that can differentiate between confident answers and less certain ones, where "the real answer is probably in here somewhere, and you'll want to ask Magnus to confirm" beats leaving them to discover by accident which parts of the wiki or document share are fiction.

This is incidentally how I already work on my own projects with Claude Code. I don't maintain a perfect project brain, I instruct Claude to maintain decisions and reference documentation as the project evolves. I tell Claude to keep a lean index in key folders and pull in detail only when it's relevant to what I'm asking Claude to build next. (I wrote a post about trimming your CLAUDE.md files for exactly this reason.) Same instinct, albeit smaller footprint.

Your messy documents will still be there next year, and the year after. Getting people to maintain good document management practices is a losing battle, always. Waiting for it to clear is going nowhere. Pick a corner to clean up, ship something people can see the value of, and let that gain you the trust to widen it. It's like organising a festival, if you ever had the pleasure. Everyone had a great time, then whisked away off home and left you alone in a field full of rubbish. You won't clear it all in one go, so clean the corner you're standing in. And then the next corner. And the next. And...