--- title: "Insights" description: "Articles on building with AI tooling while keeping quality high. Concrete methods, with sources." url: https://torn.studio/en/insights locale: en --- # Insights Every article opens with a short, quotable answer and cites a source wherever there are numbers. We hold content we build for others to the same standard. ## Insights - [Why agency prices have not moved since 2022](https://torn.studio/en/insights/why-agency-prices-have-not-moved) — Agency rates have held near 2022 levels because time-based billing turns every efficiency gain into supplier margin. Forrester and the 4A’s found 75% of US agencies absorbing generative AI as overhead. A survey of 73 UK agencies found 40% holding prices and delivering more work, while 11% passed the saving on. (2026-10-13) - [Does the EU accessibility law apply to our website?](https://torn.studio/en/insights/does-the-accessibility-law-apply-to-us) — The law names the consumer services it covers: e-commerce, banking, electronic communications, passenger transport and e-books. A website outside that list falls outside the law. Companies under ten employees and €2 million in turnover are exempt from the service requirements. Where it applies, the standard is EN 301 549, which today means WCAG 2.1 level A and AA. (2026-09-29) - [The European Accessibility Act: does it apply to your website?](https://torn.studio/en/insights/the-european-accessibility-act-and-your-website) — In Sweden the European Accessibility Act is lag (2023:254), in force since 28 June 2025. It reaches consumer-facing services: e-commerce, banking, electronic communications, passenger transport, e-books and television. Companies with fewer than ten employees and at most €2 million in turnover are exempt. PTS assesses that the requirements cover the parts of a website belonging to the service. (2026-09-29) - [When an AI agent fills in your form](https://torn.studio/en/insights/when-an-ai-agent-fills-in-your-form) — An agent reads a form through the accessibility tree, the same surface a screen reader uses. Give every field a real label, every action a real control, and hold the layout still while the page loads. Chrome's Lighthouse scores all three now, and asks for an llms.txt as well. (2026-09-29) - [The EU AI Act: what you must label from August 2, 2026](https://torn.studio/en/insights/eu-ai-act-what-you-must-label) — Since August 2, 2026, Article 50 of the EU AI Act has required a chatbot or voice assistant to say it’s AI from the first interaction. Deepfakes have to be clearly labeled, and generative AI systems have to mark their output in a machine-readable form. The duties differ for whoever builds the system and whoever runs it. (2026-09-28) - [Grok Bot: why people love it and what is actually new](https://torn.studio/en/insights/grok-bot-why-people-love-it) — Grok Bot is SpaceXAI’s always-on AI agent, released in beta on 11 August 2026. Each bot runs on its own cloud computer, signs into the tools you already use, finishes multi-step jobs and asks for approval before anything consequential. People love it because the work lands finished inside their own tools. How reliable it is on messy workflows is unmeasured. (2026-09-16) - [How to structure a landing page that converts](https://torn.studio/en/insights/how-to-structure-a-landing-page-that-converts) — A landing page converts when each section answers one question a stranger has, in the order they ask it. The eight questions run from what this is and which problems you solve, through proof, alternatives and price, to what happens if it goes wrong. A question is inaudible until the ones above it are answered. (2026-09-10) - [AI in product discovery: do the synthesis yourself first](https://torn.studio/en/insights/ai-in-product-discovery) — Use AI in product discovery as a second reader. A person synthesizes every interview first, a model synthesizes it second, and the two readings are compared. Teresa Torres’s December 2025 rule is that dumping transcripts into a model is no shortcut. Models are good at desk research, clustering and drafting, and they miss what a customer left unsaid. (2026-09-07) - [Evals are the spec for an AI feature](https://torn.studio/en/insights/evals-are-the-spec) — An AI feature is specified as an eval: a written set of cases the feature must handle, the cases where it must refuse, and a tolerance for how often it may miss, agreed before the first prompt. Hamel Husain’s March 2024 finding: unsuccessful AI products almost always share one root cause, no robust evaluation. The product role writes the eval. (2026-09-07) - [When building gets cheap, deciding gets expensive](https://torn.studio/en/insights/when-building-gets-cheap) — AI tooling has made producing software cheap, and the gain is smaller than it feels: METR’s 2025 trial found experienced developers 19% slower with AI while believing they were 20% faster. The constraint moves to deciding what to build, how to know it worked, and whether users can trust it. That is the product role, and it got harder. (2026-09-07) - [AI in your existing systems: one bounded job at a time](https://torn.studio/en/insights/ai-in-your-existing-systems) — Yes. A language model connects to your existing systems over their APIs and does one bounded job there: reading invoices, classifying tickets, answering from your own documents. Swedish supplier Swivrr publishes the bands — a simple API integration at 50,000–80,000 kr, ERP systems like SAP or Dynamics at 150,000–250,000 kr, plus 15,000–40,000 kr per extra system. (2026-09-04) - [Automating in-house or hiring: where the line sits](https://torn.studio/en/insights/automating-in-house-or-hiring) — A flow that stays inside one system, with rules that look the same every run, is yours to build: n8n Cloud costs €20 a month, Zapier from $19.99 and Make from $12. The line sits at flows that cross systems and need exception handling, monitoring and someone who answers when they stop — that is where agency work begins. (2026-09-04) - [Which processes to automate first?](https://torn.studio/en/insights/which-processes-to-automate-first) — Start with the process that repeats every week, follows the same rules every run and eats the most hours. Measure three things per process: hours per week, frequency, and what an error costs. Rank by hours times hourly cost, automate the top of the list — and leave processes that change every month untouched. (2026-09-04) - [The most common automations to hire an agency for in 2026](https://torn.studio/en/insights/most-common-agency-automations) — Accounting tops the search data: nine of ten markets probed complete their automation phrase with bookkeeping or invoicing — “ai to automate accounting” in the UK, “ai automation for accounting” in the US. Adoption statistics lead elsewhere: 52% of AI-using US firms apply AI in sales and marketing, per the Census Bureau. Agencies get hired where ready-made tools stop. (2026-09-03) - [AI consultant hourly rates in 2026](https://torn.studio/en/insights/ai-consultant-hourly-rates) — An AI consultant costs $150–200 an hour at a boutique firm, $200–350 at principal level and $300–600 at the big consultancies, according to Jahanzaib’s 2026 guide. UK advisory runs £150–400 an hour (Axivon), and platform specialists charge $40–200 (GolmTech, Ciphernutz). Torn Studio sells fixed prices: €4,200–8,800 per automation. (2026-09-01) - [What does AI automation cost in 2026?](https://torn.studio/en/insights/what-ai-automation-costs) — A single automation costs £1,000–5,000 in the UK and $300–3,000 in the US at the simple end, with first projects at £4,000–12,000 and pilot builds at $10,000–50,000, according to published supplier guides. Torn Studio prices each build fixed at €4,200–8,800, after a €2,200 automation review. (2026-09-01) - [When does automation pay for itself?](https://torn.studio/en/insights/when-automation-pays-for-itself) — Automation pays for itself when the yearly cost of the manual process clearly exceeds the build cost. The arithmetic is short: hours per week times loaded hourly cost times working weeks. Snabba Pucks publishes 120 000–300 000 kr as the typical yearly cost of such processes in Sweden; a Torn Studio build costs €4,200–8,800, once. (2026-09-01) - [Case: a gynecology clinic, its care flows and its website](https://torn.studio/en/insights/case-clinic-care-flows-and-website) — A Swedish women’s health clinic opened its physical and digital practices at once. The care flows in the platform came before the website: visit types, questionnaires, booking and fees were ready for opening day on 15 December 2023, and the website was rebuilt over the three months that followed. The order was a decision that cost visibility during the launch. (2026-08-31) - [What does an AI-built website cost?](https://torn.studio/en/insights/what-an-ai-built-website-costs) — An AI-built website costs roughly what a hand-built one costs. DigitalApplied’s 2026 benchmarks show AI-assisted build times fell 22–34 percent in a year while pricing held flat — the freed hours went into design, performance and measurement. Production time was never where the money went. (2026-08-31) - [AI crawlers and agents: how they read your site](https://torn.studio/en/insights/ai-crawlers-and-agents-on-your-site) — AI company bots do three different jobs: training models, building search indexes, and fetching pages for a user in the moment. Every major lab now documents separate bots per purpose, so you can block training while keeping visibility in AI answers. User-triggered fetchers generally ignore robots.txt, though, and Cloudflare introduces new default blocks on September 15, 2026. (2026-08-28) - [AI search in late 2026: the full picture for your site](https://torn.studio/en/insights/ai-search-late-2026) — In 2026, AI answers became the default in search: Google’s AI Mode passed a billion monthly users in May, Bing wrote GEO into its webmaster guidelines in February, and ChatGPT now runs its own search index. The foundation of visibility holds: indexable content, direct answers, verifiable numbers. Meanwhile the bots, the measurement tools, and the value of a click changed. (2026-08-28) - [AI website agency vs traditional agency](https://torn.studio/en/insights/ai-website-agency-vs-traditional-agency) — A traditional agency staffs a team of specialists and plans three to five months for a Swedish B2B project, with a 15–30% surcharge to compress it into six weeks. An AI-assisted studio staffs one person and plans in three to six weeks. The difference is coordination time. (2026-08-28) - [AI website builder or an agency?](https://torn.studio/en/insights/ai-website-builder-vs-agency) — AI builders now ship server-rendered HTML: Lovable moved new projects to server-side rendering on May 13, 2026, and v0 emits Next.js. The old objection that a crawler sees an empty page has largely fallen away. What a builder still leaves to you is accessibility, multilingual routing, and the checks that hold a codebase together after the first month. (2026-08-28) - [The best AI web agencies in Sweden in 2026](https://torn.studio/en/insights/best-ai-agencies-in-sweden) — Six Swedish agencies build websites with AI tooling in 2026: Bueno, Kumpan, Sabest, SEOS Design, The Generation and Torn Studio. Kumpan, Sabest and Torn Studio publish their prices; The Generation states a team of thirty. The list runs alphabetically and rests only on what each agency publishes itself. (2026-08-28) - [Fixed price or retainer for a website?](https://torn.studio/en/insights/fixed-price-or-retainer) — Pay a fixed price when the scope can be written down, and a retainer when the work is continuous. Swedish agencies publish hourly rates of 850–1,800 kr and ongoing maintenance at 3,000–15,000 kr a month. A fixed price moves the overrun risk to the supplier, and you pay a margin for that. (2026-08-28) - [Google AI search in 2026: AI Mode, reports, rules](https://torn.studio/en/insights/googles-ai-search-2026) — Google AI search in 2026 is AI Overviews and AI Mode merged into one experience, past a billion monthly AI Mode users. Sources are retrieved through the ordinary ranking systems plus query fan-out, so AI visibility rests on the same index as before. New this year: generative AI reporting in Search Console, an explicit AI control, and llms.txt officially ignored. (2026-08-28) - [How to choose a web agency in 2026](https://torn.studio/en/insights/how-to-choose-a-web-agency) — Choose a web agency by deciding the job first, then checking what you can verify on your own: whether they publish a price, how their own site performs on real-user data, and whether the text stays readable with JavaScript turned off. Brief three of them identically, and compare on scope, languages, integrations and deadline. (2026-08-28) - [Do you own the code to your website?](https://torn.studio/en/insights/owning-your-website) — You own the code to your website when the contract says so. Swedish copyright law transfers a computer program to the employer only where an employee wrote it; a commissioned agency keeps the rights until a written assignment transfers them. Ask for the transfer clause, the repository, the domain and the hosting account before you sign. (2026-08-28) - [What does website maintenance cost?](https://torn.studio/en/insights/what-website-maintenance-costs) — Maintaining a website costs 10–20% of the build price per year in Sweden. A simple site lands at 1,000–7,000 kr, a business website at 8,500–17,000 kr, and a platform with integrations at 25,000–50,000 kr. Hosting and a domain are rarely in the quote. (2026-08-28) - [Where AI assistants get their answers in 2026](https://torn.studio/en/insights/where-ai-assistants-get-answers) — AI assistants pull answers from their own indexes and their own fetches: ChatGPT queries an OpenAI-built index where only 1.5% of results overlap Bing’s top list, Copilot builds on Bing, Claude on Anthropic’s own search crawler, and Perplexity on its own. Profound’s analysis of 11.84 billion citations shows 57% point at brand-owned websites. (2026-08-28) - [Winning customers when AI answers first](https://torn.studio/en/insights/winning-customers-when-ai-answers-first) — High-intent customers in 2026 are won by pages that answer buying questions directly: prices, comparisons, and process. Clicks are fewer, with 68% of US Google searches ending in zero clicks, but visitors arriving via AI convert 60% better, per Adobe’s July 2026 measurement. Build to be cited, track the platforms’ new AI reports, and keep pages visibly dated. (2026-08-28) - [Deciding what to build next](https://torn.studio/en/insights/deciding-what-to-build-next) — A prioritization decision holds when it is written down: what gets built, which problem it solves, what is set aside, and what has to be true in two months for the decision to have been right. One page per decision is enough, read aloud at the next review. (2026-08-27) - [Turning your expertise into content that gets cited](https://torn.studio/en/insights/expertise-into-content-that-gets-cited) — Content that gets cited comes from people who already know the subject. The method is a 40-minute interview, a draft the same day, a fact-check by the expert, and publication within the week. Freshness carries weight: about half of what AI systems cite is under 13 weeks old, per Profound’s citation-decay analysis. (2026-08-27) - [How long does an AI-built website take?](https://torn.studio/en/insights/how-long-an-ai-built-website-takes) — A website built with AI tooling takes 3–6 weeks, the range our own website service quotes. Week one sets positioning and the sitemap, weeks two and three the design system and the first pages, weeks three to five the remaining pages and content, and the final week measurement and handover. (2026-08-27) - [When a design system pays for itself](https://torn.studio/en/insights/when-a-design-system-pays-for-itself) — A design system consists of tokens for color, typography and spacing, components built from them, and a build check that fails code carrying loose values. In our experience it pays for itself around the tenth page, or the second time a brand has to change across a whole site. (2026-08-27) - [How to appear in AI-generated answers in 2026](https://torn.studio/en/insights/appearing-in-ai-generated-answers) — Content gets cited in AI-generated answers when it is technically easy to read, answers the question directly within the first hundred words, contains specific numbers with a source, and is mentioned on other credible sites. Neither llms.txt nor extra schema markup is required — Google has said so explicitly. (2026-08-26) - [Holding quality when AI writes the code](https://torn.studio/en/insights/quality-when-ai-writes-the-code) — When AI writes most of the code, responsibility for quality moves from review to automated gates. Strict type checking, tests, lint rules that encode architectural decisions, and a single verify command that must be green before every commit together catch what a human stops noticing after the third file. (2026-08-26) - [What a website actually costs in 2026](https://torn.studio/en/insights/what-a-website-actually-costs) — Swedish price guides publish 8,000–19,000 kr for what most small businesses pay and 80,000–200,000 kr for a mid-complexity B2B site. Both are true, because they describe different products. Name the product first — pages, languages, integrations, deadline — and the number follows. Torn Studio publishes five fixed tiers from 10,000 to 250,000 kr. (2026-08-26)