Companies

This page shows how to read a workplace that builds or uses AI: who it is, where the system runs, and which duties come with the data. Named company profiles, logos and an employer form on Polaris open later.

Beginner

What workplaces use AI for

Workplaces meet AI in ordinary work. A factory may use it to flag faults on a line. A law firm may use it to search documents it already holds. A shop may use it to draft customer replies. Some workplaces build the tool and sell it. Many buy a service and point it at their own files.

Eurostat's 2025 survey covers enterprises with at least 10 employees. In the EU, 20.0 percent used at least one listed AI technology, up from 13.5 percent in 2024. Sweden stood at 35.0 percent, behind Denmark (42.0 percent) and Finland (37.8 percent). The survey counts analyzing written language, speech-to-text, generating text, code, pictures or sound, recognizing objects in images, machine learning, workflow automation, and machines that move from sensor data. Method notes on the article name the industries included. If you quote the figures, name Eurostat and the date you read them.

Bolagsverket, the Swedish Companies Registration Office, holds the legal name, organization number and registered address. A job ad from the Swedish Public Employment Service often adds the town and whether the work is remote. AI Sweden, a non-profit national center, says it works with more than 180 partners in companies, public bodies and universities. Those names are on AI Sweden's website. The partner page and the company register answer different questions.

Named profiles on Polaris, logos included, open later. Until then these sources show where a workplace is registered and what the official survey counts as use of AI.

Read one workplace

I will paste a public page about one workplace (an about-page, a report excerpt, or a job ad). Using only that text: (1) list the tasks they say AI is used for, in plain words; (2) mark each point as written on the page or not stated; (3) list what is still unknown: where the model runs, whether customer or staff personal data is included, and which town the work sits in. Do not add numbers, customers or quotations that are not in the paste. If the page is only a general claim, say which facts are missing.

Intermediate

Cloud, own machines, and the rulebook

Keep two facts apart: where the model runs, and whose software it is.

Question	Cloud service	Own machines
Files	Provider computers, sometimes abroad	Machines the workplace runs
Bill	Subscription or price per request	Cards, power, and staff time
Software	Vendor product and its terms	Weight file under the license in its readme, or an in-house model
EU AI Act	Deployer: the body that uses the system	Also the provider, if it places the system on the EU market

IMY, the Swedish Authority for Privacy Protection, says the Artificial Intelligence Act and the GDPR apply together. The GDPR covers personal data used to build or run a system, training included. Personal data can identify someone, such as a customer name or a support recording.

IMY names four risk levels: unacceptable, high, limited, and minimal. The provider assesses which risk level applies to its system. High risk covers some safety components and Annex III uses, including recruitment and credit scoring, and those uses need documentation before sale or use. The Act took effect on 1 August 2024. Banned uses apply from 2 February 2025. On 18 June 2026 several later dates were postponed, so check IMY for the live calendar.

For bookkeeping, customer service and marketing, write the task, which fields leave the building, and who checks the text before sending it. In 2025 the most common EU use in the Eurostat survey was analyzing written language, at 11.8 percent of enterprises.

Compare hosting for one task

We are choosing how to run one AI task. I will describe the task and the kinds of records involved (for example invoices, support emails, or product text), and whether any record can identify a person. Compare a cloud service with running the model on machines we administer. For each option cover where the records would sit, what kinds of cost we would pay (do not invent prices), whose license applies, and whether we would be a deployer, a provider, or both under the EU AI Act. Then list the documents to read before a pilot: vendor terms or the model readme, our record of processing, and the Swedish Authority for Privacy Protection (IMY) pages on the AI Act and the GDPR. Mark anything you are unsure about.

Expert

Stack, people, and duties

Record only what a public page states. Leave a cell blank when the page is silent.

VRAM is memory on the graphics processor. A local language model runs on a machine you choose. Describe it by its parameter count, in billions, and by the gigabytes of its compressed weight file. Note the card type, the VRAM, and how many cards serve one request. A small file may fit one workstation. A large file needs more cards or a rented server.

Name a serving program, orchestrator or queue only when a job ad, a paper or a repository names it. Kubernetes is one orchestrator: software that keeps services running across a group of machines. Hugging Face organization pages list companies and labs that publish models. The license is in the model card, its readme. AI Sweden links to github.com/aidotse. LeakPro, one public project there, targets privacy risks when AI is used.

Job ads from the Swedish Public Employment Service are CC0 open data: you may reuse them commercially with no permission step. JobStream carries live Platsbanken ads. Historical ads on the same page go back to 2006 and keep the employer's own job titles. Wikidata's structured facts are CC0 as well. query.wikidata.org returns a website and a location when a contributor has filled them in.

The provider places the system on the EU market and classes its risk. The deployer uses it. IMY says the GDPR still applies to personal data, including during training. The Polaris declaration form opens later, and a logo needs written permission.

Build a stack sheet from public pages

I will paste public sources: job ads, a GitHub organization, a Hugging Face model card, or a company page. Turn them into a sheet and leave a cell empty when the source is silent. Columns: model name; license as written on the card or in the contract; parameter count if stated; where it runs; graphics card model; VRAM in GB; number of cards; serving program; orchestrator; data categories; whether personal data is in scope; provider or deployer under the EU AI Act; and the URL for each fact. Add the job titles exactly as the ads write them. Do not infer cluster sizes, towns, or customers. End with questions an engineer and a privacy officer should answer together before the system is used on customer or staff records.

Read one workplace from public pages

Use this order before you trust a paragraph about a firm.

Legal identity. On Bolagsverket, note the exact name, organization number and registered address.
Place of work. In a current job ad, note the municipality and any line about remote work.
The task. Copy the sentence that says what the AI system does. If the page has no such sentence, write "task not stated".
The data. List the kinds of records that would have to enter the system for that task to work, such as invoices, emails, images or case notes.
People. Mark whether those records can identify a person. If they can, the GDPR applies, on IMY's reading, including when the records were used to train a model.
The supplier. Name the product if the page names it. Leave the cell blank when the page only says "AI".
Source line. Keep the URL and the date you read it next to every fact.

Treat partner badges and award logos as optional. Note where the data is stored, as stated on the product page or in the contract, on its own line.

Questions for a Polaris AI declaration

You can ask a workplace these questions now. The Polaris form that stores the answers opens later. Keep each answer short enough that an outsider can check it.

Which tasks use an AI system this quarter, in one sentence each?
For each task, is the system bought as a cloud service, run on machines you administer, or both?
Which model is in use, and which license is named on its model card or in the contract?
Which categories of data are sent to it, and which of those are personal data?
In which country are those data stored while the system handles them?
Are you the provider, the deployer, or both, under the EU AI Act?
Which risk level has been recorded (unacceptable, high, limited, or minimal), and where is that written down?
Who reviews the output before it is sent to a customer, an employee, or a public decision?
Where is the public code or the model card, if there is one?
May Polaris show your logo beside this declaration? Only a written yes counts.

Leave unknown lines blank so a later reader can see which facts are still unpublished.

What the workplace profile will do

When the workplace area opens, a firm will be able to submit its own profile: legal name, organization number, towns where staff work, public links, and the declaration above. That profile will also be where a firm attaches a job ad. The ad will have to name a real employer that matches a Bolagsverket record, describe the work as concrete tasks, and be free for the applicant. Polaris will show a logo only after the workplace sends permission. Until that area opens, use the public sources on this page and keep your own notes.

A shared community DGX Station could read the open job-ad stream overnight and group workplaces by the tools and skills the ads name, and later run a model check on that machine so files a workplace submits for the check stay in the building. About the shared station

Sources

Eurostat: 20% of EU enterprises use AI technologies2025 survey release: EU, Sweden, Denmark and Finland shares, and the most common techniques. Figures are for enterprises with at least 10 employees. Reuse needs source acknowledgement. Copyright terms: https://ec.europa.eu/eurostat/help/copyright-notice (editorial text under CC BY 4.0; statistical data reusable with attribution).
Regulation (EU) 2024/1689, the Artificial Intelligence ActOfficial text on EUR-Lex. The page also points to the consolidated version. Use it for provider and deployer duties, banned practices, and high-risk rules including Annex III.
IMY: questions and answers on the AI ActSwedish Authority for Privacy Protection. Risk levels, high-risk examples such as recruitment and credit scoring, the application calendar (including the 18 June 2026 postponement), and the statement that the GDPR applies in parallel when personal data is used, including in training.
Swedish Public Employment Service: job-ad open dataPlatsbanken ads as open data under Creative Commons CC0, including the JobStream API for live ads and the historical collection described as running from 2006. Useful for employer name, place, and job titles.
AI SwedenNational non-profit center for applied AI. The homepage states that it works with more than 180 partners in the private sector, the public sector and academia. The footer links to public code at https://github.com/aidotse, including the LeakPro project.
Bolagsverket: open dataSwedish Companies Registration Office. Start here for the legal name, organization number and registered address, and read the terms on the product page you open.
Wikidata licensingStructured data on Wikidata is released under CC0. Other namespaces on the site use CC BY-SA 4.0. Query what volunteers have entered at https://query.wikidata.org/ (official site, headquarters, coordinates when present).
Hugging Face organizationsPublic directory of companies, universities and non-profits that publish models. The license for each model is in its model card, documented at https://huggingface.co/docs/hub/en/model-cards. Read that card before using the weights.