Sweden took its AI in-house

A decision about AI at work is a decision about where an organisation’s data goes. The Swedish government’s block on ChatGPT and Claude shows what that means in practice.

Its staff have lost access to both chatbots on their work computers.¹ Whatever they typed there went to OpenAI and Anthropic, the companies behind them. The government had no say over what happened to it next.² The AI itself stayed. They now reach models from the same two companies through Intric, the government’s own platform, hosted in Sweden.^{1 3} So what was closed was one way of using AI, the one where the state’s data left its hands.

Data is what gives AI its value at work. A chatbot brings general knowledge, while everyday tasks need the organisation’s own facts in the chat. The more of them staff paste in, the better the answers fit.

So the data that makes AI useful is the same data the organisation has to protect.

Every use of AI at work puts some of it somewhere.

From a personal account to a model of one’s own

Where the data goes depends on how staff reach the AI. There are four levels, each giving the organisation more say:

1. The first is a personal consumer account. OpenAI trains on what is typed there unless the user opts out.⁴ Anthropic does so only with consent, and can then keep the chats for up to five years.^{5 6} At this level the data leaves without the organisation knowing.
2. The second is a business plan. It keeps inputs out of training, but the data still sits with the vendor.^{4 7} The organisation knows which tool is in use, while the provider holds what goes into it.
3. The third is Sweden’s choice. The organisation runs the platform itself on servers in its own country, so it decides what goes on to the models behind it.^{1 3}
4. The fourth is an open model installed on in-house hardware, so the data stays on the premises. Each step up moves the decision about the data away from the user and the vendor and towards the organisation.

Seen this way, Sweden moved its government staff up to the third level. The public chatbots it blocked sit at the bottom of that scale, where it had no say at all.

A block works when there is a replacement

Much AI use at work sits at the lowest level. Two thirds of office workers use AI they believe their employer forbids, and a third have put customer data into public tools.⁸ In firms of more than 1,500 people, almost three quarters do so.⁸ They use it because it helps them do their jobs. A block on its own leaves that reason untouched.

The only trend data available shows what works instead.

Use of personal AI accounts fell from 78% to 47% of users at companies that offered approved apps and blocked unwanted ones.⁹

Many switched to the sanctioned tool once it existed. Sweden followed the same sequence. Intric was running before the block came into force.^{1 3}

Three steps, in this order

For a company, Sweden’s lesson is the sequence. The first step would be finding out which AI tools and accounts staff already use. Each of them keeps the company’s data at one of the four levels.

When Sweden’s transport agency checked, it found about ten different AI products in use.^{10 11} That made ten places where its data could end up.

The second step would be choosing the level each kind of data needs. Personal data typed into a public chatbot falls under the GDPR.¹² A trade secret can lose its legal protection if the firm takes no steps to keep it confidential.¹² At the second level such data still leaves the organisation, so its safety rests on a vendor’s terms. It belongs higher up, at the third or fourth. What is already public can sit lower, even on a personal account.

The third step would be giving people a tool at the chosen level, and only then closing the levels below it.

Blocking first leaves staff with the same work and fewer tools to do it. Done this way round, employees keep their AI and the organisation keeps its data.