The Filing Cabinet Problem

Imagine you hire a new assistant. Brilliant, fast, works for coffee money. There’s just one catch: they can only remember things by color.

Every document you hand them, they file by color alone — not by topic, not by client, not by date. Red folder, blue folder, green folder. You ask them to find “that contract with the Bavarian supplier from March,” and they hand you every red folder in the building, because red is the closest match they have.

Would you trust this assistant with 100,000 documents? Probably not without asking one question first: how many colors do they actually have to work with?

This is, almost exactly, the question every business owner needs to ask before signing an AI vendor contract. The “colors” have a technical name — vector dimensions — and whether your vendor gives you 8, 800, or 8,000 of them determines whether your AI assistant finds the right contract or just the reddest folder in the room.

You don’t need to become an engineer to ask this question. You just need to understand what’s actually happening behind the sales pitch.

What “AI Understanding Your Documents” Actually Means

When a vendor says their AI “understands your business documents,” here’s what’s really happening, step by step:

  1. Every document gets converted into a list of numbers — think of it as a very long coordinate, like a pin dropped on a map, except the map has hundreds of dimensions instead of two.
  2. Documents about similar things end up with similar coordinates — close together on this map.
  3. When you ask a question, your question also gets a coordinate — and the system looks for the nearest documents on the map.

This process is called embedding, and the map is called vector space. The number of dimensions in that space — how many numbers make up each coordinate — is the “colors” in our filing cabinet story.

Here’s the part vendors rarely explain clearly: more dimensions doesn’t automatically mean better. It means more room. Whether that room is used well is a completely different question — and it’s the one your vendor’s price tag depends on you never asking.

The Neighborhood Analogy

Picture two neighborhoods.

Neighborhood A has 8 streets. Every house has to sit on one of those 8 streets, even if two houses have nothing in common except that all the “good” streets were already full. Neighbors end up next to each other by accident, not by similarity.

Neighborhood B has 800 streets. There’s room for a street for law firms, a street for logistics companies, a street for coffee roasters, a street for coffee roasters who also do logistics. Houses end up next to their actual neighbors.

If your business has 200 fairly similar contract templates, Neighborhood A might be perfectly fine — there simply isn’t much variety to get lost. If your business has contracts, support tickets, and technical manuals covering a dozen product lines in three languages, Neighborhood A will quietly force unrelated things to live next door to each other. Your AI assistant will occasionally hand a customer service answer that’s confidently wrong — not gibberish, just a plausible neighbor instead of the right house.

This is the single most common failure mode in business AI projects: not a crash, not an error message — just quietly wrong answers that sound right enough that nobody questions them until a customer does.

“But Bigger Must Be Better, Right?”

Not necessarily — and this is where most vendor conversations stop too early.

A neighborhood with 3,000 streets sounds more impressive than one with 800. But if the city planner (the company that built the AI model) did a sloppy job, half those streets might be empty, badly connected, or duplicates of each other in disguise. A well-planned 800-street neighborhood can outperform a poorly-planned 3,000-street one.

In other words: the size of the vector matters less than the quality of the model that built it, and whether it fits your specific documents. A vendor quoting you a big, expensive number is not automatically giving you a better system — they might just be giving you a bigger, more expensive filing cabinet with the same color-blindness problem.

This is exactly the kind of claim that sounds authoritative in a sales deck and completely falls apart when you test it against your own documents instead of the vendor’s demo data.

The Question That Actually Protects You

Before signing anything, ask your AI vendor one grounded question:

“Have you tested this against a sample of our actual documents — and can you show me it retrieves the right ones, not just plausible ones?”

If the answer is a shrug, a generic benchmark from someone else’s industry, or “our model is very advanced” — that’s your signal to pause. A vendor confident in their fit should be able to run a small test on your real contracts, tickets, or manuals and show you the results, not just the specification sheet.

This isn’t a technical nitpick. It’s the difference between an AI assistant that saves your team hours and one that quietly erodes trust every time it hands someone the reddest folder in the room instead of the right one.

The Real Lesson

The mistake most Mittelstand businesses make isn’t choosing the wrong AI vendor. It’s asking the wrong first question. “What can AI do for us?” is the easy, exciting question every vendor is happy to answer. “Does this specific systemactually work on our specific documents?” is the uncomfortable one that actually protects your investment.

The problem is rarely the technology. It’s skipping the step where you test the promise against your own reality before you pay for it.

Considering an AI implementation and want a second opinion before you commit budget to it? That’s a conversation worth having before the contract is signed, not after.