Our position

AI rarely fails on the model — almost always on the data before it

When quotes, parts lists and maintenance records sit scattered across folders that grew over years, a language model adds nothing at first. The first useful step is usually an unspectacular one.

Our position

Before a language model is worth anything in a company, the documents have to be findable — so the first useful step toward AI is usually order, not intelligence.

Why does AI deliver so little in many companies at first?

Because the model is not the problem. In the companies we work with, quotes, parts lists, inspection and maintenance records live in folder structures that grew over years, in mailboxes and on individual machines, under file names only the person who chose them can decode. A language model cannot make anything reliable out of that, because it faces the same question as any new employee: which version actually applies? So when NDVDL is called into an AI project, the first question is rarely which model fits, but which documents exist in a state where an answer built on them would hold up. That is the uncomfortable part, and it is also the part that carries everything else.

01

A model can only read what can be found

The idea that a language model simply searches "everything we have" underestimates how much of a company never becomes a file at all. A great deal lives in email threads, in handwriting on a printed record, in a spreadsheet one person maintains, or in the head of the colleague who has looked after the machine for years.

No software can open up what was never captured. A tool built on that basis does not give wrong answers because it is a poor tool — it gives incomplete answers because the ground under it is incomplete. For the business the effect is much the same, but the distinction decides where the work has to start.

02

An answer is only as unambiguous as the filing

More common than missing documents is having too many. The same quote exists in several versions, the maintenance record sits once in the project folder and once in the customer folder, and which one counts is something people know from who touched it last — not from its name or its location.

A model has none of that context. It will pick one of the versions and phrase it convincingly. That is the awkward property: an answer drawn from unclear filing does not look uncertain. Before you ask people to trust answers like that, it has to be settled which document applies.

03

The workflow decides whether the result lands

Even a correct answer is worthless if nobody knows who checks it, who signs it off, and at which point in the working day it is actually needed. Tools that sit beside a workflow instead of inside it stop being opened after a few weeks — that is as true for AI as for any other software.

This is why we do not see the preparatory work as a delay but as the same job in a different order. Deciding where documents belong, what they are called and who owns them improves the business on its own terms. If a model arrives afterwards, it has something solid to stand on. If it never arrives, the work was still not wasted.

The case against our position

The strongest argument against us is that tidying data without a concrete purpose has no natural end: there is no yardstick for when it is good enough, and that is exactly where such efforts quietly die — the cleanup continues, it is never finished, and nobody ever saw what it was for. An early, deliberately imperfect pilot inverts that: within days it shows which documents people genuinely reach for and which ones nobody has opened in years, and it produces the justification for cleanup work that nobody wanted to fund beforehand. On top of that, people can judge a tool once they have used it; they cannot judge a folder structure. Waiting for clean data very often means never starting at all, and that risk is real enough that we have to weigh it in every conversation.

What this means in practice

  • We say plainly when we think an AI project is premature, even though it means not winning the larger piece of work for now and recommending unglamorous preparation instead.
  • We say before the start that the first phase produces nothing you can demonstrate in a meeting — anyone who needs a quickly visible result should know that before committing.
  • We cut the first step as small as we can: one area, a manageable set of documents, a clearly named purpose, so the preparation has an end and does not turn into a permanent project.

Does that sound like your situation?

Then let us talk about what it concretely means for your business.

IT infrastructure
Follow-ups

What we get asked about this

No. It means the order matters. We do not suggest spending a year tidying up before starting; we suggest defining the first use case narrowly enough that the required order stays manageable — and beginning in the area where the documents are already in the best shape.

Settling where a document belongs and where it does not, how files and folders are named, who owns a given body of records, and which version counts as current. It also includes the uncomfortable step of noting what only exists verbally today and has never been written down anywhere.

They open up whatever is already digital, and that is a genuine step forward. What they cannot do is decide which of three versions applies, or surface what was never recorded. Such tools make unclear filing more accessible — they do not make it unambiguous.

Not necessarily. If it is running, the most useful question is: where were the answers unusable, and was that the tool or the underlying material? A running pilot is a good diagnosis, as long as you are willing to read its results as a statement about your data too.

Wondering where AI could actually start in your company?

We look at the state your documents are really in first — and tell you honestly whether an AI step makes sense now or later.

Request a first consultation+43 800 007075