The why
remains.

There are now providers who sell exactly this one case as a finished product: interviews with those leaving, and from them a searchable knowledge store. That is decent work, and for some organizations it is precisely the right thing. The difference lies in what happens afterwards. There the project ends with an application that holds this knowledge. With us it is the first use case on a foundation that stays — the same knowledge is then available to every further process, without having to be gathered again.

If you are unsure what your case needs, we will also tell you when a point solution is enough for you.

What gets lost is not the what

The what is in the system: maintenance interval 800 hours, customer X has a three percent discount, supplier Y is not approved. That can be exported, backed up and handed over.

What is missing is the why:

The why decides whether a decision is repeatable. Without it, every exception is negotiated afresh, every mistake is potentially repeated, every handover becomes a blind flight. And it is written down nowhere, for three reasons:

It is extra work with no immediate benefit. Whoever knows why does not need the note. The benefit only arises once that person is gone — which is precisely the moment when nobody is left who knows it would have been missing.

It feels like a farewell. Writing down knowledge that only you hold implies: I could be replaced. That is rarely a conscious thought and almost always an effective brake.

And the why often contains a mistake — with the name of whoever made it right beside it. “We do it this way because something went wrong in 2019” is also a statement about a person. In an organization with a lived culture around mistakes, that gets written down. In every other one it stays spoken.

This is why thirty years of knowledge-management software did not solve the problem: it was never a tooling problem. A system that turns writing things down into a task fails for the same three reasons as every wiki before it.

Our approach is therefore a different one: the system does the groundwork, not the accountability. A question is asked, an agent researches and records the result — today it writes straight into the wiki. The intermediate step that matters is being built right now: the agent creates a proposal, a human adds, corrects and signs off. Writing things down does not disappear because of it. But it shrinks from a blank page to a review, and it happens where the work is being done anyway, instead of in a second, separate exercise.

That is not a technical detail but the core of it: whoever signs off takes responsibility — and that is exactly what makes an entry usable. A body of knowledge that nobody stands behind is just another archive. So the human belongs in the loop at the points where it counts. The evidence for it is already there: every change to the knowledge base is versioned — traceable who changed what, and when.

Integrity and sovereignty were meaningful, tangible values to us decades before these words became IT-business jargon. That is also why we are fascinated by the ability to build electronic “brains” for the benefit of all. And the problems that come along with that huge leap concern us equally.

Over the years we have seen tremendous efficiency gains in our IT landscapes. We have also seen growing dependencies — vendor lock-in, data lock-in — arise for individuals and organizations as soon as a once-frontier IT technology had been fully adopted.

Europe can learn from this — and send impulses into all markets instead of merely receiving them. We would like to contribute. Whenever a needed technological layer can efficiently be built around, or replaced by, an independent alternative rooted in Europe, we put it on our roadmap.

With the right strategy and the right tools, organizations can meet the challenges of using AI technologies and keep intellectual property in their own hands. We are developing Alveraign, driven by a deep desire to build a Europe-centric, yet decentral AI platform.

Alveraign provides agentic AI workflows — accessible via chat and API and backed by a multi-layered memory. Users build their own scenarios on the platform: work one-to-one with an agent, draw on predefined teams, or assemble their own sequence of existing agents at runtime for a particular task. A manager agent that distributes sub-tasks and consolidates the results has been demonstrated and is currently being merged into the main branch. Every step is traceable live as it happens.

Knowledge sources — manuals, procedural instructions, ongoing correspondence — can be connected: dropped into a watched folder, they are split up and made searchable by content. The body of knowledge is to grow through practical use — created by the system, reviewed and owned by people (see above); the sign-off loop for that is being built now. Relevant information carriers — important partner organizations, competitors, customer types — are modelled as representations an agent can step into on demand.

Product owners and administrators calibrate the system through a central system prompt as well as through roles, goals, backgrounds and tools per agent — declaratively in configuration files, applied while the system is running. Large language models can be used out of the box or set per agent as a default and fallback model.

You alone decide where a request goes. Whether all requests are processed by locally running models — or whether, in certain use cases, the frontier models of those cloud providers that are acceptable for your business environment come into play. You draw that line, not us, and you can draw it differently per agent.

Alveraign has been built entirely in-house — from an idea, with clear design principles and the care of someone who answers for every line himself. No construction kit, no borrowed architecture. What we offer you, we run in our own house first.

The prototype. Today Alveraign runs on a system in northern Germany. Two applications are live — a chat interface and an administration dashboard — both reachable through a European identity provider and protected by production-grade authentication. Login and security boundaries are verified end to end from the outside — hardening inside the local network is on our list and comes before productive operation. Agents solve tasks with tools, work together in teams and delegate among themselves. Documents are ingested through a watched folder and made searchable by content. Roles carry their own domain knowledge.

And what does not work today, we say just as plainly. The sign-off loop for knowledge entries is being built — until then an agent writes directly. Scheduled tasks run into the void and are under repair. Files attached in the chat rather than dropped into the watched folder land in the context unfiltered instead of being searched. For each of these there is an entry with cause and target date — that is the list we show you.

What this is and what it is not. A prototype with the fundamental feature set in place, not a finished product in the sense of an enterprise catalogue. We know the maturity level of every function — what runs, what runs with limitations, what does not yet. Anyone who gets specific gets the list. If that clarity means more to you than terse maturity claims, we are probably having the right conversation.

In autumn 2026 the productionization of the prototype begins — as the next stage of development: hardening for productive operation, role-based permissions, certification, first pilot partnerships. Until then we are having conversations — and looking for the use cases we can learn the most from.

In its guidance on the particularities of valuing small and medium-sized enterprises, the German Federal Chamber of Tax Advisers notes that intangible factors closely tied to the person of the owner “are used up over time” without that person’s continued involvement. As an example it explicitly names the owner as the carrier of specialised knowledge. The consequence, verbatim:

In plain terms: what exists only in one person’s head does not count towards transferable earnings power. That turns securing knowledge from a personnel measure into a valuation measure — and it concerns not only those handing over, but everyone who advises on or finances a handover.

That this is no fringe topic is shown by the succession report of the DIHK, based on more than 50,000 chamber-of-commerce advisory contacts: 70 percent of senior entrepreneurs have no emergency file, and three quarters first talk to anyone about it two years or less before the handover.

Since 2 August 2026 the transparency obligation of the AI Act applies: anyone talking to an AI system must be able to recognise that. The obligations for high-risk applications were postponed in July 2026 and now take effect on 2 December 2027 and 2 August 2028 respectively. Postponed is not cancelled — and the two requirements that matter technically, logging and human oversight, were not changed in substance. In Germany, the Federal Network Agency has been the competent authority since 29 July 2026.

Alveraign is designed for this: every working step of an agent can be followed and is logged, knowledge is versioned. As for human oversight, we are in the middle of building it — the sign-off loop is the step that turns a proposal into a decision someone owns. That does not replace legal advice — but it spares you the question of whether your own infrastructure can supply the evidence at all, once it is demanded in 2027.