The hard part is the data. At Telekom, a platform decides what leaves the building.

What the article says

The minister's point was that putting a chatbot in front of a customer-service desk is not enough. In his words, the data that AI systems can be built on has to be merged first, and the state's systems, which today run in many places on different technologies, have to be made to work with each other. Data protection and data security, he said, have to be given a leading role. He described a state that does not know everything about everyone, where people have real control over their own data and it is transparent who accesses it and for what purpose.

Rékasi opened with a lesson. One of Telekom's first big projects was to replace its push-button customer-service system with a voice assistant, and the company budgeted eighteen to twenty-four months for the system to learn Hungarian. Before that work was finished, the global large language models had essentially solved the problem on their own. His conclusion: the speed of AI development is itself one of the biggest corporate risks, because it is easy to put a great deal of money into something that becomes technologically obsolete within a short time. Telekom's answer is a hybrid model. Keystone AI decides per task between internal processing and an external model; keeping costs in check is one stated goal, while the company's own internal data can still be used. On jobs, he said there is no mass elimination of roles for now, but that some tasks and roles can be taken over entirely, and the resources freed up can be redirected.

The rest of the panel covered education and AI literacy, with OECD and reading-comprehension figures. Those describe the population, not a company's AI rollout, so we leave them aside.

Two very different organisations reached the same diagnosis

The state says: merge the data and connect the systems before the agent. Telekom's lesson: the eighteen to twenty-four months it had set aside for its own build were overtaken, because the capability arrived from elsewhere first. What neither of them can outsource is the part underneath: where the data is, what state it is in, who may access it, and which data may go to which model.

We see it the same way. A model running on your server is not yet a platform; it is a running program. The work is in the integration, the permissions, the audit log and the release path. Our customers, 20-to-200-person companies in regulated or data-sensitive sectors, reach the same point at a smaller scale. In place of the state's patchwork of systems, they have an ERP, a document store and a few spreadsheets. Merging comes before the agent there too.

Keystone names the real question: where is the line

The most valuable sentence in the article is the one about Telekom's platform: it decides, per task, what stays inside and what goes out. That is the decision layer. Anyone who deploys AI on customer data builds this layer, whether or not they give it a name.

This is where the paths diverge. The article names one stated goal behind Keystone: keeping costs in check, while the company's own internal data can still be put to use. Our decision rule is the classification of the data. Confidential data in a production workflow, the prompts, documents, logs and embeddings, stays on the platform, on EU infrastructure, with no US data processor in the path. In development and test environments, on synthetic or anonymised data, public AI is allowed. So the line exists in our world too. We draw it by environment and data classification, write it into the contract, and enforce it at the network level: production has no route out to a public model. The GDPR and CLOUD Act side of this is in our GDPR-compliant AI hosting guide.

The difference looks small; from the buyer's side it is the whole point. A cost-based decision is re-made on every task. A classification-based decision is made once, in the contract, and the customer can verify it. Which one belongs behind a customer-service bot depends on the data it touches: if contract numbers, health data or account numbers show up in the questions, the second.

What we think, including the parts that cut against us

This article is a validation and a competitor announcement at the same time, and we do not want to dress the second part up as agreement.

First: Keystone AI forwards to OpenAI and to Gemini. The country's incumbent telecoms operator answered with a hybrid. That runs straight into our own blunt sentence, "the data does not leave the company." Almost every real architecture is hybrid, ours included. The sentence is true of customer data in a production workflow; for development environments it holds only with conditions. So the argument is about where the line runs, what may cross it and who decides, not about whether one exists.

Second: Telekom shipped first. That is a fact. The article describes a named platform that routes work task by task. On our side there is one public, anonymised customer story, a certification body that spent two years on failed attempts and then got a modern application with four environments in about two months, plus our own products on our own platform. For a company of 20 to 200 people we are the right size. Next to Telekom we are small.

Third: the ministerial part is a statement of intent at a conference. There is no tender, no budget and no deadline behind it. "One to two years" is a stated goal. We do not read a strategy into it, and at this size the state is not our customer anyway.

Fourth: state-scale data consolidation says nothing directly about a 200-person company. What carries over is the order of operations: data and access first, agent second. The scale does not.

Fifth: our argument is self-interested. We sell precisely the layer we are claiming matters here. That is why we spelled out a checkable rule alongside the slogan.

Where we stand

We agree with both speakers' diagnosis. You buy or download the model. You clean up the data, connect it, and decide what may go out. The article describes Telekom deciding task by task, with cost control among the stated goals.

One testable sentence: if a workflow runs on customer data, you should be able to say in one sentence which part of it may leave the building, and who decided.

Guides: Private AI for business · Private ChatGPT · Private AI hosting · Private AI platform · Your own private AI · GDPR-compliant AI hosting