Every wholesale business I walk into has the same person. You know the one. They've been there fifteen years, they know which supplier actually delivers on time, they remember why that customer gets special pricing, and they can tell you off the top of their head which product replaced the one you discontinued in 2019.

They're also the reason everything slows to a crawl the week they're on leave. And if they ever leave for good, a big chunk of how your business runs walks out the door with them.

That's not a people problem. It's an information problem. And it's the thing an "AI knowledge base" is supposed to fix — a phrase that's currently doing a lot of heavy lifting in a lot of sales decks, so let me tell you what it actually means, and where the money gets wasted.

The real problem isn't AI. It's where your knowledge lives.

Walk the floor of most wholesale or distribution businesses and the knowledge is everywhere except where you can find it. Some of it's in spec sheets. Some's in a supplier's email from eight months ago. Some's in a shared drive that three people organise three different ways. And a frightening amount of it is in one or two people's heads.

The day-to-day cost is interruptions. Someone needs to know a lead time, so they ask. Someone's not sure how to handle a return, so they ask. New staff take months to get useful because there's no way to learn the business except by asking. The questions almost always have answers — the answers just aren't findable.

The bigger cost is the one nobody puts on a spreadsheet: you're running a business on knowledge that exists in a couple of brains, and brains take annual leave and hand in resignations.

Why the obvious fixes don't fix it

Before anyone reaches for AI, they've usually tried two things.

A shared drive or a wiki. The trouble is that storing a document isn't the same as being able to answer a question. A folder full of PDFs doesn't tell you the lead time on the X range — you still have to know which file, open it, and read it. And the SME wiki that someone lovingly set up two years ago is, I'd bet, about six months out of date, because keeping it current was nobody's actual job.

Generic AI, usually ChatGPT. This is the one I get asked about most, so let me be blunt: ChatGPT knows the internet. It does not know your business. Ask it about your products, your suppliers, your pricing tiers, and it'll either tell you it has no idea, or — worse — it'll make something up that sounds completely plausible. Confidently wrong is more dangerous than "I don't know," because someone acts on it.

So what is an AI knowledge base?

Strip away the jargon and it's three things working together.

First, it's grounded in your material — your documents, spec sheets, processes, supplier emails, the stuff you actually run on. Not the internet.

Second, it answers questions in plain language. Instead of someone hunting through folders, they ask "what's the lead time on the X range for this supplier?" and get the answer.

Third — and this is the part that separates a useful tool from an expensive liability — it shows you where the answer came from. Every answer points back to the source document, so a person can check it in two seconds. And when the answer genuinely isn't in your material, it says so, instead of inventing one.

That's the whole thing. Your knowledge, made answerable, with its workings shown.

The five ways these projects waste money

This is where I earn my keep, because I've watched enough of these go sideways to know the failure modes. If you're going to build one — with me or anyone else — these are the things that decide whether it's worth the spend.

1. Garbage in, garbage out. A knowledge base is only ever as good as what you feed it. Point it at a drive full of contradictory, half-finished, out-of-date documents and you get confident, well-written, wrong answers. A real chunk of the work — the unglamorous chunk — is deciding what's actually authoritative, and getting the stuff that only lives in someone's head written down. There's no AI shortcut around that.

2. It has to cite its sources. If the thing can't show you where an answer came from, you can't trust it, and an answer you can't trust is worse than no answer because someone will act on it anyway. Grounding and citations aren't a nice-to-have. They're the difference between a tool and a hazard.

3. It has to be allowed to say "I don't know." Anyone selling you an AI that always has a confident answer is selling you a problem. The most valuable thing one of these can do is admit when something isn't in your material, so a human knows to step in. That feature is boring and unsexy and it's the one that matters most.

4. It will go stale, so plan for that. Products change, suppliers change, processes change. A knowledge base nobody maintains rots, and a rotten one — full of yesterday's answers delivered with today's confidence — is worse than not having one. Keeping it current has to be someone's job from day one, not a problem you discover in month four.

5. Some knowledge should stay human. The factual lookups — lead times, pricing, process steps — are exactly what this should take off people's plates. The judgment calls aren't. A good build is clear about that line, so the tool frees your expert up for the things that actually need them, instead of pretending to replace them.

Where to start

Not with a grand "let's put all our knowledge into AI" project. Those collapse under their own weight.

Start with one painful, well-defined area — the questions your team asks most, or the knowledge most at risk of walking out the door. Get that working, see whether people actually use it, then widen it. Small and proven beats big and abandoned, every time.

If you've got a person everyone leans on, or a folder nobody can navigate, or a sinking feeling about what happens when your most experienced staff member retires — that's the signal it's worth looking at. The hard part was never the AI. It's deciding what your business actually knows, and getting it somewhere everyone can reach.

That's the kind of thing the automation audit is for — working out what knowledge is worth capturing and what it's costing you to have it stuck. If you want to put a rough number on it first, the calculator takes about 30 seconds.