Garbage In, Garbage Out: Why Your AI Is Only as Good as the Data You Feed It
Last month I was pulling together a response to a tender - the kind of document where the organisation’s reputation rides on getting the detail right. I asked my new bestie, Claude, to draft the technical capability section using our SharePoint sites and our iManage workspaces as sources. Within minutes, it had produced something plausible, confident, and completely out of date. It had lifted a case study from a project we hadn't worked on in seven years, quoted a fee structure we updated years ago, and cited a team member who left the company in 2022.
Nothing it wrote was invented. Every word traced back to a real document sitting in our own systems. The problem wasn't Claude. The problem was what we'd been feeding it. This holds true for any AI agent connecting to your enterprise data.
What Your AI Is Really Reading
Most organisations I know are sitting on a staggering volume of content: SharePoint sites from every matter and project going back decades, iManage workspaces full of drafts, superseded versions, and emails nobody ever cleaned up. A lot of it has no business being there anymore. In records management circles, this has a name: ROT -Redundant, Obsolete, and Trivial information.
-
Redundant means the same thing saved in six places.
-
Obsolete means the version that got superseded eighteen months ago but never got archived or deleted.
-
Trivial means the "quick draft, ignore this" email chain that somehow became the file everyone finds first.
None of that was a serious problem when a human was doing the searching, because a person could usually tell immediately that a document looked stale. An AI agent connected to those same repositories doesn't have that instinct. It retrieves what matches the prompt, and if the best-matching document happens to be an obsolete precedent or a redundant duplicate with the wrong figures, it will use it — fluently, and more worryingly, confidently. The output reads beautifully. It's just wrong, or worse, subtly wrong in a way that's hard to catch on a quick read.
Why this matters more once AI is in the loop
When I'm drafting a document by hand, I instinctively filter as I go. I recognise a clause I haven't seen in years and pause. I see a project methodology that I know has changed - and I disregard it. I see an author I don’t trust, and I move on. An AI agent searching across an organisation’s Windows, SharePoint and iManage content doesn't have that hesitation built in. It treats a 15-year-old draft and yesterday's finalised version as equally valid inputs unless something in the metadata or the content tells it otherwise. Give it a clean, current, well-classified library and it produces genuinely excellent first drafts. Give it decades of unmanaged sprawl and it will happily assemble a document out of whichever fragments happen to rank highest, with no idea that half of them belong in the bin.
That's what happened with my tender response. The retrieval wasn't broken, but the library was.
Finding the ROT Before Your AI Does
This is where information governance stops being a compliance tick box and becomes a precondition for decent AI output. Morae's managed InfoGov solution, which I've now seen work across Windows file shares, SharePoint and iManage environments, is built around exactly this problem: finding the ROT before it finds its way into an AI-generated document. It scans across repositories to classify content at scale, flagging duplicates and superseded versions. It also identifies what's genuinely redundant, obsolete, or trivial - all with full audit trails, approvals, and legal hold protections built into the workflow, so nothing gets disposed of carelessly.
The output isn't just a tidier file plan, it's also a trusted source library that reliably informs AI agents or initiatives.Current precedents are correctly tagged as current, superseded drafts either archived or clearly marked, duplicates resolved down to a single authoritative copy. When I re-ran that tender draft after creating a better context, the difference was immediate. My bestie Claude pulled the right case study, the current rate structure, and the right people. Same prompt, same AI, dramatically better output, because the data it was drawing from was clean and current.
The Lesson for Organisations Leaning into AI
Every organisation I talk to is racing to plug AI into their document systems, and rightly so; the productivity gains are real. But there's a less visible piece of groundwork that determines whether those gains show up or backfire: is the underlying content actually fit to be read by a machine that takes everything literally? An AI agent doesn't know your SharePoint site has three versions of the same NDA template, or that most of your iManage content consists of matters that closed years ago, or trivial emails. It just knows what it can retrieve.
Before you measure the quality of your AI drafting, it's worth asking a less exciting but more important question: When did anyone last clean up the library it's reading from? Fix that, and the AI you already have gets noticeably better, without a single change to the model.
Morae can help you eliminate the ROT, clean up your library and keep it clean, for higher-quality outcomes. And this makes me and my bestie, Claude, happy.
