What a public assistant can do, and what it cannot
Generative AI took hold in German companies unusually fast, because it needed no project. Opening a website was enough. Within a few months a novelty had become an everyday tool: for drafts, translations, summaries, first ideas.
These tools are good at what they were built for. They have a command of language and of general knowledge about the world. What they do not have is access to what decisions in a company actually rest on: orders, costings, contracts, inventory, customer histories, complaints.
The difference becomes clear in the kind of question. “Draft a rejection letter to a supplier” is something a public assistant answers well. “Which of our projects in the past six months ran more than ten percent over the estimated time, and what did they have in common?” it cannot answer. Not because the model is too weak, but because it does not have the data.
That second kind of question is the one that matters commercially. And that is exactly where the difference begins between a general tool and one that moves something in your own business.
How an assistant works on your own data
One misunderstanding persists here that makes many discussions unnecessarily complicated. The common idea is that you have to “teach” an AI your company documents, that is, train the model on your own data.
That is not how it works in practice, and for good reasons. Knowledge baked into the model goes stale with every price change. It comes without a citation, so the answer cannot be checked. And it cannot be restricted per person: once something is in the model, it is in there for everyone.
The usual route is a different one, and it fits into a single image. The assistant does not learn things by heart before the exam. It sits the exam with the manual open in front of it.
In practice this means: for every question, the system first pulls the relevant passages out of the company’s systems, from tables, documents, metrics. It puts those passages in front of the language model together with the question. The model writes the answer from those alone and names what it rests on.
Two consequences follow from this, and both weigh more than any technical detail.
First, every answer can be checked. If you do not trust the number, you click through to the source. That is the difference between an interesting tool and one that someone negotiates a price on.
Second, the quality of the answer is a question of the data, not of the model. If the same metric is defined differently in three systems, no assistant in the world can give a binding answer. It will pick one of the three variants and phrase it plausibly. Ordered, harmonized data is therefore not a preliminary stage of the AI project. It is the project’s substance.
Everyone sees exactly what they are allowed to see
One point decides whether an assistant on company data can be rolled out widely: it has to follow the same rules that already apply in the company.
The principle is simple. The assistant always works in the name of the person who is asking. It sees exactly what that person is allowed to see anyway, and nothing beyond that. A sales director asks for the numbers and gets their own area. Management asks the same question and gets the full picture. Both use the same application, the same data and the same metric definitions.
For that to work reliably, permissions sit in one place and not separately in every system. They are taken from the existing user directory, apply down to the individual row of data, and take effect regardless of whether someone opens a report, calls up a dashboard or asks a question in plain language. Anyone who changes role sees the matching slice from the next day on. Nobody has to think about it, and nobody has to maintain a second analysis for it.
This is why governed access and AI belong together. They are not a brake on adoption, they are the condition for making an assistant available to all business teams in the first place. Where the rights are cleanly recorded, the circle of users does not have to be kept small, and data can be available across the whole company without anyone losing control over it.
What it delivers
The benefit shows up in two places, and both are well known in the German Mittelstand.
The first is knowledge transfer. In many companies, decisive knowledge is tied to individual people: the controller who knows the pitfalls of a certain type of order. The scheduler who knows which supplier turns unreliable over the summer. The service technician who spots a fault from three sentences in a log. Much of this knowledge already sits in the company’s own systems, in old orders, quotes and service reports. It is just not findable. An assistant on that data makes it retrievable, and above all for the people who need it most urgently: new employees and stand-ins.
The second is access. Today, the route to an analysis usually runs through a request to IT or controlling, through an export and through a few days of waiting. Often enough the question is no longer current by then. If it is asked in plain language instead and answered in seconds, that changes not only the speed but the kind of questions. People follow up. People try things out. Anyone with a hunch checks it instead of keeping it to themselves.
That is the real effect: working with data leaves the circle of a few specialists and becomes a tool for large parts of the company.
Conclusion
The question is no longer whether a company works with AI. It is already in use, in many companies without any connection to their own data.
The real question is whether that use rests on a body of data that belongs to the company, is in order, and whose access is governed. Only then does an assistant answer the questions that really count in the business, and only then are its answers solid enough to base decisions on.
The way there does not run through the model, it runs through the data. That is the uncomfortable and at the same time reassuring news: the work this takes pays into reporting, data quality and governed access anyway. It would have been the right thing to do without AI as well.