Articles
These essays explain AI from a first definition of a token through cost, workplace use, EU rules, model architecture, and security. They point at public pages you can open and check.
Articles
AI agents: what they are and why they change everything
An AI agent plans, chooses tools, and performs tasks on its own, without you sitting beside it. A chat window only answers when you type something.
AI and accounting — what can be automated (and what cannot) 2026
Receipt scanning, bank import, automatic coding, VAT returns at the touch of a button. AI promises to remove all the tedious parts of accounting. But what actually works today for a small or medium-sized Swedish company — and what is still just marketing?
AI for customer service — chatbots, automated responses, and what actually works in 2026
A customer asks a question at 10:30 PM on a Friday. Without AI: no one responds until Monday. With AI done right, the customer can get a correct answer right away. A wrong answer can do more harm than no answer. Here is what actually works for Swedish SMEs.
AI for your marketing: concrete tools and workflows for Swedish companies
Most Swedish companies know they "should use AI in marketing" — but stop at generating a LinkedIn post every now and then. That's not enough. There is a systematic way to use AI that cuts the time you spend on marketing, without losing your voice.
How to build an AI strategy for your company — a 90-day plan
Most Swedish SMEs know they should have an AI strategy. The problem: they don't know where to start, and are worried about getting stuck in projects that cost money but don't deliver. The plan is written for companies with 5–100 employees — without consultant fees and without tearing up existing systems.
Automate with AI Agents: A Framework for Swedish Companies
Most companies have three to five processes that are repetitive enough to be automated — but vague enough that no one has actually done it. That is where an AI agent can help the most. This article provides you with a concrete framework to identify, build, and measure your first production-ready agent.
Claude vs ChatGPT vs Gemini vs Grok — which one should you pay for?
Four models, four pricing plans, four different philosophies. The difference between the right and wrong choice is often a noticeable sum every month — and hours lost when the model simply doesn't understand what you are trying to do. Here is the framework that actually works for Swedish entrepreneurs who want to choose once and avoid switching every quarter.
EU AI Act: What Swedish companies actually need to know
The EU AI Act takes effect in stages from 2025 to 2028. What does this mean for an average Swedish company using ChatGPT and Claude? More than most think — but often less dramatic than the headlines suggest.
Local AI models: When you no longer want to send your data to the USA
Ollama, open models, and local hardware make it possible to run powerful AI directly on your server — without a single word leaving the office.
The five most common mistakes with AI — and how to avoid them
Most people who "tried AI and got nothing out of it" make the same five mistakes. The article shows what works instead.
Prompt engineering for business use: 10 techniques that actually work
Most people use AI as a glorified search engine. Those who understand prompt engineering — and use it systematically — get results that are far better from the exact same model. Ten techniques. Swedish business examples. Everything ready to copy.
Stop using AI as Google
You ask questions and expect answers. Give the AI a task instead of a question and you will more often get what you want.
What does AI cost for Swedish companies? — An honest cost guide 2026
"AI is expensive" some say. "AI is free" others say. The answer depends on what you choose, how many people use it, and if you are calculating correctly. This guide gives you the numbers.
Suggest a question for the next test edition
Your suggestion goes to an editorial queue before anything is published.
We save your question and optional name for editorial review. If you provide email, we use it only to notify you when the test is published. To stop spam we keep a scrambled code made from your IP address for two days. The address itself is never saved. Privacy.
Vote on candidate questions
- Can a local AI handle customer support in natural Swedish without sharing data?0
- Which open model is best at reviewing Swedish contracts and finding hidden risks?0
- How well do local models transcribe and summarize Swedish meetings and calls?0
- Can a local model interpret Swedish receipts and invoices as reliably as the cloud?0
Beginner
Tokens, prompts, and five early mistakes
A large language model continues text one small piece at a time. Those pieces are tokens. A token is a chunk of text, often a word or part of a word, stored as a number. The model reads those numbers, predicts the next one, and turns it back into text. A short English word is often one token. A long word or a name may be several. The context window, how much text the model can consider at once, is counted in tokens, so a pasted contract fills it faster than a short question.
The message you type is a prompt. A useful prompt names the job, the reader, the facts to stick to, and the shape of the reply. "Write about AI" leaves every choice open. "Explain tokens to a shop owner in about 150 words, with one example, and stop" can be finished and checked.
Five mistakes show up in weak first tries. The real task is buried in a long story. A fact is requested and no source is named. Private customer data goes into a public chat. Fluent writing is treated as proof. One reply is asked to research, decide, and send the message. Split the work. Ask for a draft, ask what could be wrong, then check each claim on a page you open yourself. Essays here start with a definition, a picture, and a checklist. Swedish editions are on the Swedish articles page.
Rewrite a vague request
You are teaching an adult who has never used an AI chat. First, explain what a token is, using one ordinary English sentence as the example. Then rewrite this request so a model can finish it: "Write something about AI for my business." The rewrite must name the reader, the job, the length, and one thing to leave out. After the rewrite, list three ways the answer could still be wrong, and name what I should open to check it.
Intermediate
Cost, a ninety-day trial, and the EU rules
These essays cover what a model costs and what a workplace may do with it. A hosted chat can look free and still bill you per token. A local model has no per-token bill. You still pay for a computer, electricity, and setup time. Name the unit, the task size, and the hours spent checking drafts.
A small laptop model can sort email and draft replies. A larger hosted model suits a long source. A reasoning model spends extra tokens on steps before the answer. That can help with plans and sums, and it costs more. An agent calls tools, such as search or mail, in a loop. Set a short tool list and a stop rule.
Spend ninety days on one task. Month one: write the prompt for routine customer replies and compare its output with a reply you wrote yourself. Month two: time twenty real cases and count errors, while a person still sends every message. Month three: keep or stop, and record where personal data went. Marketing, support, and bookkeeping drafts fit the same trial. A person still approves any invoice, promise, or filing.
Regulation (EU) 2024/1689 on EUR-Lex is the binding EU text. The Commission service desk lists start dates, including Digital Omnibus changes, so confirm them there. IMY says the GDPR applies when personal data is handled as AI is built or used. The 2026 AI Index reports that industry produced over 90 percent of notable frontier models in 2025, and that documented incidents rose from 233 to 362.
Plan a ninety-day trial
Help me plan 90 days of AI use for one task only: drafting replies to routine customer emails at a small company. A person still sends every email. Ask me at most five questions, then give a month-by-month plan. Include what to measure, which personal data must stay out of a public chat, and which pages to open before I use a tool on customer text: the EU AI Act on EUR-Lex (Regulation (EU) 2024/1689), the European Commission AI Act implementation timeline, and the Swedish IMY guidance on GDPR and AI. Do not invent prices. If a price is required, name the vendor page I should read.
Expert
Measurements, architectures, and ways the system fails
Treat every comparison as a dated measurement. The LMArena dataset stores human votes between two answers. A vote is one preference on one prompt. Also test your own files. The 2026 Stanford AI Index gathers tests, spending, and reported incidents. An arXiv cs.AI preprint is a draft: read the method, and look for code or a later paper, before you change a system.
A Transformer, from Attention Is All You Need (2017), scores each token against the others in the context. Cost rises fast as the context grows. A state-space model keeps a running memory. Mamba (Gu and Dao, revised 2024) makes that memory selective. The authors report linear scaling, about five times Transformer throughput, and a 3-billion-parameter model matching Transformers twice its size on their tests. The paper is Creative Commons BY. Cite those figures as their results.
Quantization stores weights in fewer bits so the model fits in memory. llama.cpp (MIT license) supports 1.5-bit to 8-bit integer copies and downloads a GGUF file in its quick start. AWQ, activation-aware weight quantization, protects channels that activations mark as important. The authors say protecting about 1 percent of salient weights greatly reduces error, with no training. EXL2, in archived ExLlamaV2, mixes 2-bit to 8-bit layers for an average of 2 to 8 bits per weight. Further work points to ExLlamaV3.
Scratch work in a reasoning model can be wrong. An agent adds a tool loop. Prompt injection hides new instructions in a page or an email. A jailbreak pushes past provider limits. Read NIST AI 600-1 and AI 100-1.
Check a model comparison
I will paste a claim that one AI model is better than another. Treat the claim as unproven. Ask for the task, the date, and whether the number is a human preference vote (as in the LMArena leaderboard dataset), a fixed test, or a vendor post. Then say how to check it against the Stanford HAI 2026 AI Index and the original paper or model card. In a separate paragraph, explain a Transformer versus a selective state-space model such as Mamba, and limit any performance figures to results I can open in the paper. Name one long-context test that would strain a home computer. Mark every point you cannot verify.
Check a claim before you act
Before you act on a paragraph, do this:
- Write the claim in one sentence.
- Open the linked page. Record its date next to the essay's date.
- Keep a measured figure apart from a forecast.
- For an EU duty, read the article on EUR-Lex and the matching row on the Commission timeline.
- For a model ranking, write down the task, the date, and who ran the test.
- If the next step would paste customer or staff data into a chat, stop and read the IMY guidance.
EU AI Act dates to confirm
The European Commission's AI Act Service Desk timeline says these dates include amendments from the Digital Omnibus on AI. The binding text is Regulation (EU) 2024/1689 on EUR-Lex in English and in Swedish. Confirm a date on both before you depend on it.
| Date | What the service desk says starts |
|---|---|
| 1 August 2024 | The regulation enters into force. |
| 2 February 2025 | General provisions (definitions and AI literacy) and the prohibitions. |
| 2 August 2025 | Rules for general-purpose AI models, meaning models built to serve many tasks. Member states must have designated authorities, and EU-level governance must be in place. |
| 2 August 2026 | Most of the remaining rules, including transparency duties in Article 50. Enforcement begins for the rules already applicable, including general-purpose models, prohibitions, transparency, and AI literacy. |
| 2 December 2026 | Further prohibitions, including systems that generate non-consensual sexual deepfakes or child sexual abuse material. Some synthetic-content systems already on the market before 2 August 2026 get until this date to meet Article 50(2). |
| 2 August 2027 | Each member state should have at least one AI regulatory sandbox running. |
| 2 December 2027 | Rules for high-risk AI systems listed in Annex III. |
| 2 August 2028 | Rules for high-risk AI built into products already covered by the EU product laws in Annex I. |
Courses, papers, and licenses
Free courses, if you want a path beside the essays:
- Elements of AI (University of Helsinki and MinnaLearn) starts from ordinary language. The license page says the lessons are all rights reserved: cite them, and get written permission before you copy them.
- Google's Machine Learning Crash Course is the short technical introduction.
- The Hugging Face LLM Course is for people who already write Python. It is Apache 2.0: keep the license credit if you reuse a lesson.
Papers and code named above:
- Attention Is All You Need (2017), the Transformer paper.
- Mamba: Linear-Time Sequence Modeling with Selective State Spaces, revised 2024, Creative Commons BY 4.0.
- AWQ: Activation-aware Weight Quantization. See the arXiv license on that page.
- llama.cpp, MIT license, local inference, GGUF in the quick start.
- ExLlamaV2, archived; the README points to ExLlamaV3. EXL2 is documented there.
- New preprints: arXiv cs.AI recent.
Our World in Data on artificial intelligence publishes the long-run charts. OWID's own articles and visualizations are Creative Commons BY 4.0, so name the authors if you reuse one. Data from other organizations keeps its original license.
A shared community DGX Station would give a place to rerun the quantization, long-context, and agent tests in the expert essays when a home computer runs out of memory. About the shared station
Sources
- Stanford HAI 2026 AI Index ReportYearly figures on technical results, investment, incidents, education, and policy. The PDF is linked from the page. Open the chapter before you repeat a number.
- Our World in Data: Artificial IntelligenceCharts on compute, data, investment, and use. OWID's own articles and visualizations are Creative Commons BY 4.0; credit the authors. Third-party data keeps its own license.
- Regulation (EU) 2024/1689, EUR-Lex (English)The binding text of the EU AI Act. The Swedish text is on EUR-Lex in Swedish.
- EU AI Act implementation timelineEuropean Commission service desk. Dates from entry into force through August 2028, including Digital Omnibus changes. Recheck before you rely on a date.
- IMY: guidance on GDPR and AISwedish Authority for Privacy Protection. A technical part and a legal part. GDPR applies when personal data is processed while AI is developed or used.
- NIST AI 600-1, Generative Artificial Intelligence ProfilePublic risk profile for generative AI (PDF). The broader AI Risk Management Framework is NIST AI 100-1.
- Elements of AIFree introductory course from the University of Helsinki and MinnaLearn. All rights reserved. Cite it; do not copy lessons without written permission. See the license page.
- LMArena leaderboard datasetPublic dataset of human preference votes between model answers. A vote records a preference on that prompt. It is a separate question from accuracy on your own files.