If that sounds familiar: you are not the problem. And neither is the AI.
The number everyone quotes — and the half nobody adds
In 2025, MIT examined more than 300 AI deployments in its study “The GenAI Divide” and produced the number that has opened every panel discussion since: 95 percent of GenAI pilots deliver no measurable return. What gets quoted far less often is the finding right next to it: more than 80 percent of companies experiment with tools like ChatGPT — and report genuinely positive experiences. The disappointment doesn’t come from trying things out. It comes from the attempt to move from trying out into everyday operations.
The study names the cause clearly, and it’s less comfortable than “the tool is no good”: the tools are not integrated into processes and data. They learn nothing about your company. They don’t know your orders, your contracts, your machine history.
Run the test yourself
Ask your assistant three things. How many units of product X do we have in stock? What does our maintenance contract with customer Y say about response times? Which complaints are piling up this quarter?
Three times, no usable answer — of course not. The assistant has no access to your ERP, your contract archive, your ticket system. It can write brilliantly and supply solid general knowledge. But the questions that make up your working day happen to be about your company.
That is the entire explanation of your disappointment, and it has a consoling side: the problem is solvable. It just sits somewhere other than where the licence purchase suggested.
What the failed trial reveals about your data
Treat the failed pilot as a free diagnostic. It has shown you two things.
First: your numbers aren’t connectable. For an assistant to answer questions about stock or revenue, it needs a central, clean, documented data foundation underneath — otherwise it guesses, and guesses eloquently, which makes things worse. If sales and accounting already sit in meetings with different numbers today, an AI won’t know which ones are right either.
Second: your knowledge isn’t accessible. Contracts, service reports and manuals are scattered across folders and inboxes — and across heads. According to an Atlassian survey, office workers in Germany spend around ten hours a week searching for information; more than half only get at knowledge by asking colleagues. And those colleagues, by the way, are leaving: current projections by the German Economic Institute (IW) expect around 1.3 million people per year to retire out of the German labour market through 2036.
The path, in the right order
Whoever simply buys the next tool after a failed pilot repeats the experiment with the same outcome. The sequence that works is unspectacular: foundation first, then the AI on top.
For the numbers side, that means a central data foundation with automated preparation — which first makes your figures reliable, and then askable. For the knowledge side, it means a curated knowledge base on which an assistant operates that answers exclusively from vetted company knowledge, with sources. And your existing licences? Keep them. For generic tasks — drafting, summarising, translating — they are already worth every euro today. They were never the wrong tool. Just the wrong promise.
Which side comes first in your case is what our Data Readiness Check shows you in five minutes. And if you like, tell us about your Copilot trial in the first conversation. We hear this story often — and by now we know fairly precisely how it can continue.