# aitrainer.dk - Complete Documentation

> AI training, workshops, and consulting for Danish organizations. Run by Stefano Vincenti, external lecturer at IT University of Copenhagen and co-founder of TryZone.dk.

## About Stefano Vincenti

Stefano Vincenti works at the intersection of theory and practice:

- Co-founder of BotTellMe — builds and implements AI solutions, ensuring they create value for clients and users from day one
- Partner at TryZone — advises, teaches, and trains large enterprises and public organizations, bridging IT and business to implement AI
- External lecturer at IT University of Copenhagen and DIS Study Abroad — teaches modern AI practices, digital product development, and the business potential of generative AI, including Scandinavian case studies and the geopolitical implications of AI adoption
- Contributor to "Agentbogen" — the most widely read introduction to AI agents in Denmark

## Core Services

### AI Workshops and Training

Customized workshops for teams and organizations covering:
- Generative AI fundamentals and advanced applications
- Hands-on training with AI tools (ChatGPT, Claude, Gemini, Copilot)
- Prompt engineering and context engineering
- AI integration in workflows

### AI Agents Training

Specialized training on:
- Understanding the difference between chatbots and AI agents
- Building simple autonomous systems
- Multi-agent architectures
- AI agent security and governance

### Microsoft Copilot Adoption

Workshops focused on:
- Copilot rollout strategies
- Maximizing Copilot productivity
- Change management for AI adoption
- Measuring Copilot ROI

### Microsoft 365 Copilot — From License to Real Productivity (NEW)

Enterprise training program (DKK 7,500 per person, half day, min. 6 participants):
- Copilot in Teams and the full Office suite — effective techniques for everyday work
- Custom prompt libraries tailored to specific business processes
- Adoption strategy: champions program with train-the-trainer, ROI measurement, and change management
- Target audience: IT leaders, department heads, transformation leads, program managers, project managers, scrum masters, product owners, change managers, and super users in organizations with M365 Copilot licenses
- URL: /workshops/copilot-training

### Claude: Chat, Cowork & Code (NEW)

Enterprise training program (DKK 7,500 per person, half day, min. 6 participants):
- Claude Chat: effective prompting, projects, artifacts, skills, and research mode
- Claude Cowork: work directly in your folders with AI as a thought partner — strategy, brainstorming, document review
- Claude Code: setup, context management, Plan Mode, custom commands, MCP servers, and GitHub integration
- Target audience: Knowledge workers, developers, leaders, and teams who want to master Claude across Chat, Cowork, and Code
- URL: /workshops/claude-training

### EU AI Act Compliance

Training on:
- Understanding EU AI Act requirements
- Risk classification systems
- Compliance timelines (major requirements from August 2, 2026)
- Practical implementation guidance

### Vibe Coding Workshops

Training on AI-assisted development:
- Using AI for coding (Lovable, Cursor, GitHub Copilot)
- Understanding benefits and risks
- Managing AI-generated code quality
- Best practices for AI-assisted development

## Workshop Landing Pages

### Generativ AI Kursus for Virksomheder
URL: /workshops/generativ-ai-kursus (DA) | /en/workshops/generative-ai-course (EN)
Comprehensive generative AI training for Danish businesses. Covers ChatGPT, Claude, Gemini and practical business applications.

### AI Workshop København
URL: /workshops/ai-workshop-kobenhavn (DA) | /en/workshops/ai-workshop-copenhagen (EN)
On-site AI workshops in Copenhagen and across Denmark. Customized training at your location.

### AI Agents Kursus
URL: /workshops/ai-agents-kursus (DA) | /en/workshops/ai-agents-course (EN)
Training on building and using AI agents. From basics to advanced multi-agent systems.

### Copilot Workshop
URL: /workshops/copilot-workshop (DA & EN)
Microsoft Copilot adoption and training. Maximize productivity with AI assistance.

### Vibe Coding Workshop
URL: /workshops/vibe-coding-workshop (DA & EN)
AI-assisted development training. Build applications faster with AI tools.

### EU AI Act Workshop
URL: /workshops/eu-ai-act-workshop (DA & EN)
Compliance training for the EU AI Act. Prepare your organization for upcoming regulations.

## Key Frameworks and Approaches

### 10/20/70 Rule
- 10% of value comes from algorithms
- 20% comes from technology
- 70% comes from people (culture, skills, processes, leadership)
Organizations that fail with AI focus on the 30%. Organizations that succeed focus on the 70%.

### Augmentation Over Automation
GenAI has a 95% failure rate for full automation. Augmentation—AI that makes humans more productive—is the pattern that works.

### Human Oversight
Practical human oversight is essential, especially for high-risk AI systems. This goes beyond compliance checkboxes.

## Latest Blog Posts / Insights

### Connect an AI agent to Jira and Confluence in 30 minutes (one-pager, June 2026)
URL: /indsigter/forbind-ai-agent-til-jira-og-confluence
Mini-guide dated June 24, 2026 aimed at product owners and scrum masters whose developers got faster with AI — making the PO the new bottleneck in Jira/Confluence. Delivers a practical, ~30-minute setup of a terminal-based AI agent that can read Confluence, search Jira, refine feature descriptions, and write back to tickets gated by explicit human approval. Stack: GitHub Copilot CLI as the agent + the open-source mcp-atlassian MCP server (github.com/sooperset/mcp-atlassian) run on demand via uvx (from Astral's uv). Architecture: You (terminal) → Copilot CLI (agent) → MCP server mcp-atlassian (bridge using the open Model Context Protocol) → Jira + Confluence. Tested on Windows 11; macOS/Linux work with small command tweaks. Not an official Atlassian, GitHub, or Microsoft product. Prerequisites: Windows 11 PC; active GitHub Copilot license (Copilot CLI included on Free, Pro, Pro+, Max, Business/Enterprise — on org-managed plans an admin must enable Copilot CLI; usage draws on the plan's AI Credits); access to ≥1 Jira project you can read/edit (Data Center primary, Cloud also works with a different auth step); read access to relevant Confluence spaces; admin-enabled Personal Access Tokens (Jira/Confluence Data Center 8.14+/6.0+). Seven steps: (1) Install Copilot CLI — winget install GitHub.Copilot (recommended Windows 11) OR npm install -g @github/copilot (Node 22+); after install, the PATH change requires closing and reopening PowerShell entirely — the most common "it doesn't work" trap; verify with copilot --version. (2) First start + login — cd into an empty workspace folder (NEVER a system folder like C:\\Windows\\System32), run copilot, answer "Yes, and remember this folder" when prompted to trust the folder, then /login and complete the browser flow once; /help anytime, Ctrl+C to exit. (3) Get PATs — on Data Center: log in to Jira → avatar → Profile → Personal Access Tokens → Create token, name it (e.g. copilot-mcp), set an expiry, create and copy immediately (won't be shown again); repeat in Confluence — they are separate apps so you need two tokens; if you don't see Personal Access Tokens, ask admin to enable PATs. Cloud users: create an API token at id.atlassian.com/manage-profile/security/api-tokens instead and use JIRA_USERNAME + JIRA_API_TOKEN (email + token) in place of JIRA_PERSONAL_TOKEN. (4) Install uv/uvx — winget install astral-sh.uv, close and reopen PowerShell, verify with uvx --version; uvx fetches and runs mcp-atlassian on demand so you don't install the server separately or manage a Python environment. (5) Configure MCP server in ~/.copilot/mcp-config.json — create the file if missing. Data Center template: {"mcpServers":{"mcp-atlassian":{"type":"local","command":"uvx","args":["mcp-atlassian"],"env":{"JIRA_URL":"https://jira.your-company.com","JIRA_PERSONAL_TOKEN":"...","CONFLUENCE_URL":"https://confluence.your-company.com","CONFLUENCE_PERSONAL_TOKEN":"..."}}}}. Cloud variant swaps the env block to JIRA_URL: https://your-company.atlassian.net, JIRA_USERNAME, JIRA_API_TOKEN, CONFLUENCE_URL: https://your-company.atlassian.net/wiki, CONFLUENCE_USERNAME, CONFLUENCE_API_TOKEN. Self-signed corp certificate not in Windows Trust Store: add JIRA_SSL_VERIFY: "false" and CONFLUENCE_SSL_VERIFY: "false" — but try without first since certs already in the Windows Certificate Store validate automatically. Recommended scoping env vars: JIRA_PROJECTS_FILTER ("PROJ,DEV"), CONFLUENCE_SPACES_FILTER ("TEAM,DOCS"), and READ_ONLY_MODE ("true" for the first session, then flip to "false" once you've confirmed reads work). (6) Restart and verify — Ctrl+C out of Copilot CLI, run copilot again, then /mcp; you should see mcp-atlassian listed as connected with tools including jira_search and confluence_search. Test with "Vis mine åbne Jira-sager" — the first time the agent uses a tool it asks for permission, which you review and approve; if results look wrong, be explicit with JQL: "Søg i Jira med JQL: assignee = currentUser() AND statusCategory != Done ORDER BY updated DESC". (7) Put it to work — three example prompts: pull context ("Søg i Confluence i TEAM-spacet efter vores retningslinjer for feature-definitioner, og opsummér den standard, vi bruger til at skrive feature-beskrivelser"); refine ("Kig på de åbne features i projekt PROJ. For hver enkelt: læs eventuelle linkede Confluence-dokumenter for kontekst, og omskriv så beskrivelsen, så den er tydelig, konsistent og klar til review. Vis mig de foreslåede ændringer først."); write back ("De ser gode ud. Opdatér hver Jira-sag med den forfinede beskrivelse."). The agent always shows the proposed change and waits for your "ja" before writing — human-in-the-loop by default. Troubleshooting table: copilot not recognized → restart PowerShell; /mcp doesn't show mcp-atlassian → check ~/.copilot/mcp-config.json exists and is valid JSON (no trailing commas), exit and restart; uvx not found → reopen PowerShell after uv install, run uvx --version; auth/401 errors → re-check token values and URLs, confirm PAT not expired, confirm you used the right token for the right service; SSL/cert errors → add JIRA_SSL_VERIFY / CONFLUENCE_SSL_VERIFY "false" (only for genuine self-signed corp certs); agent reads but doesn't update → READ_ONLY_MODE is still "true", flip to "false" and restart. Security checklist: tokens are stored in plaintext in mcp-config.json (normal for MCP configs but handle the file with care) — (1) never commit mcp-config.json to Git, never paste tokens in chats/screenshots; (2) set a sensible expiry on each token; (3) rotate immediately on suspected exposure (create new, update config, revoke old in Jira/Confluence); (4) use filters + read-only mode to constrain blast radius. AI wingman pattern: keep a strong current model (e.g. Claude Opus 4.8 or Sonnet 4.6) open beside you to paste error messages into, translate Windows steps to macOS/Linux, explain settings, and sanity-check the mcp-config.json before save. Quick-overview recap: install Copilot CLI → restart → trust folder → /login → create Jira + Confluence PATs → install uv/uvx → add mcp-atlassian to ~/.copilot/mcp-config.json → restart → /mcp → test → let the agent refine features and update Jira. Explicit disclaimer: general practical guidance, NOT legal/security/compliance advice, and NOT an official Atlassian/GitHub/Microsoft product. Giving an AI agent read/write access to Jira and Confluence touches your organization's data and access keys — clear it with your own IT, security, and compliance functions and follow your company's third-party-tool policy before enabling write mode in production. PATs act on your behalf — treat them as passwords; you are responsible for how they are stored, scoped, and rotated. Tools, plans, pricing, terms and security behavior change quickly — always verify the current state against primary sources (GitHub Docs, the mcp-atlassian repository at github.com/sooperset/mcp-atlassian, and your own Atlassian admin) before deciding. Use of third-party and open-source software is at your own risk. Snapshot as of June 2026. Downloadable PDF (Danish, 10 pages). Keywords: GitHub Copilot CLI, MCP, Model Context Protocol, mcp-atlassian, Jira MCP, Confluence MCP, Atlassian MCP server, Jira Data Center PAT, Jira Cloud API token, Personal Access Token, uvx, uv Astral, Python tool runner, READ_ONLY_MODE, JIRA_PROJECTS_FILTER, CONFLUENCE_SPACES_FILTER, AI for product owners, AI for scrum masters, terminal AI agent, agentic workflow Jira, feature refinement AI, Confluence search AI, human-in-the-loop, AI wingman, Claude Opus 4.8, Claude Sonnet 4.6, AI Built Human, Stefano Vincenti, BotTellMe, TryZone.

### Claude, jurisdiction and ID verification (one-pager, June 2026)
URL: /indsigter/claude-jurisdiktion-og-id-verifikation
One-pager dated June 23, 2026 reframing the debate around Anthropic's new ID verification (via Persona, from July 8, 2026) for Claude. Thesis: a client asked whether to drop Claude over the ID verification — that's the wrong anchor. The right question is two-part: (1) which plan are you on, and (2) which data should ever touch a US model in the first place. Facts on ID verification: applies to consumer plans (Free, Pro, Max); does NOT apply to Team, Enterprise, or the Developer Platform/API, which run on Commercial Terms — a separate agreement. Plan distinction is the key — consumer plans (Free/Pro/Max): consumer terms, ID verification required, your data can be used to train the model unless you actively opt out, no DPA, never use for business data; Team/Enterprise (commercial agreement): you are data controller, Anthropic is processor, they may NOT train on your content, you keep inputs and own outputs, DPA with SCCs included by default, no ID verification — always use a business account for serious work, including Claude Code, never a personal account for work data. Jurisdiction is what actually matters: a US vendor is subject to US law regardless of server location — the CLOUD Act (2018) reaches the parent company, not the Frankfurt server room; EU Data Boundary and similar move data location, not jurisdiction; this is neither new nor unique to Anthropic — applies equally to your US cloud and US CRM. Fleeing to Chinese models doesn't solve it — you trade the risk for one that's harder to see through. GDPR angle: claude.com is a Chapter V third-country transfer; can today rest on the Data Privacy Framework (in force since July 2023) but the basis is fragile — under appeal at the CJEU (case C-703/25 P) and rests on a US presidential order a future administration can change; keep an SCC fallback ready (SCCs are already in Anthropic's DPA), and for EU residency route via e.g. AWS Bedrock EU (Ireland or Frankfurt). Which data should touch a US model — GREEN (usually fine, when input is cleaned of personal data): strategy and analysis, marketing, internal knowledge, design, coding without trade secrets; YELLOW (prefer EU hosting): customer-service content, personalization and recommendations, vendor data; RED (not US infrastructure): customer personal data, HR data, sensitive contracts, special categories (health, biometrics). The grown-up answer in five steps: (1) find out which plan you're on — consumer or commercial — it decides both the ID requirement and your data terms; (2) read your DPA (GDPR art. 28) — where data lands, who the sub-processors are, DPF certification (check the official list at dataprivacyframework.gov); (3) decide data tiering and write it down — personal data, art. 9 special categories, trade secrets and customer data don't necessarily belong in the same place; (4) keep an SCC fallback + transfer impact assessment ready — if DPF falls at the Court, companies with SCCs in place already are best positioned; don't wait for the ruling; (5) regulated sector or supplier to a public authority — involve your DPO or counsel now, not after the decision is made. Explicit disclaimer: general practical guidance, not legal or compliance advice — author is not a lawyer or compliance specialist; don't use as a sole basis for decisions; always involve your own compliance/legal team and DPO on your specific situation before vendor or data decisions; timelines and case law shift — verify dates and status against primary sources (EUR-Lex, European Commission, Datatilsynet); status as of June 2026. Downloadable PDF (Danish). Keywords: Claude ID verification, Anthropic Persona, Claude Team Enterprise, consumer vs commercial terms, Anthropic Commercial Terms, GDPR Chapter V, third-country transfer, Data Privacy Framework, DPF appeal CJEU C-703/25 P, Standard Contractual Clauses, SCC fallback, transfer impact assessment, CLOUD Act 2018, EU Data Boundary, AWS Bedrock EU Ireland Frankfurt, data tiering green yellow red, AI vendor risk, AI jurisdiction, Article 28 GDPR, dataprivacyframework.gov, AI Built Human, Stefano Vincenti.

### Coding agents: Four players in the terminal (one-pager, June 2026)
URL: /indsigter/coding-agents-fire-spillere
One-pager dated June 22, 2026 comparing the four dominant terminal and agentic-coding tools: Claude Code (Anthropic), Codex (OpenAI), Antigravity (Google), and GitHub Copilot (Microsoft). Core thesis: choosing the tool and choosing the model are two different decisions — the tool defines workflow (terminal, IDE, multi-agent, procurement); the model defines capability and where data is processed. And vendors can change the rules on 30 days' notice (Google killing Gemini CLI on June 18 is the example). Per tool — Claude Code: today's flagship claude-opus-4-8, terminal-first, built for long autonomous sessions with full context and deep refactoring, available via AWS Bedrock and Google Vertex (EU region option), skills/hooks/subagents, mature CLI; Fable 5 / Mythos-class (1M context) launched June 9 but globally suspended since June 12 after a US export-control directive that affects foreign nationals (including in the EU); no free tier, programmatic use draws on separate credits; flag: frontier access can shift day to day — vendor risk, not just a feature difference. Codex: model gpt-5.5, bundled with the ChatGPT subscription, widest surface coverage (CLI, IDE, ChatGPT cloud, GitHub bot, computer use — one account), token-efficient, included in Plus/Pro/Business/Enterprise, ~4 million weekly developers; best if already in the ChatGPT ecosystem (otherwise API prices on top), usage limits/credits can bite on long tasks; flag: GPT-5.5 is strong on cyber — clarify policy on offensive vs. defensive use cases before broad rollout. Antigravity: model-agnostic (Gemini 3.x / Claude 4.6 / GPT-OSS), the most ambitious on multi-agent orchestration, Manager view for multiple agents in parallel with free model choice per task, Artifacts make agent work traceable, CLI + SDK + desktop share the same harness, free preview with generous Gemini limits; new and changing fast, free tier heavily cut (~20 requests/day, weekly cap), replaces Gemini CLI which shut down on June 18 (forced migration, community pushback over closed source); flag: Google killed an open-source tool with 100k+ stars on 30 days' notice — example of how fast the rules can change. GitHub Copilot: multi-model (GPT, Claude, Gemini), strongest on compliance and procurement, deeply integrated in the GitHub flow, now US + EU data residency under Microsoft's EU Data Boundary (incl. EFTA/Nordics) — also for Claude models — FedRAMP for the public sector; lower capability ceiling than the pure frontier tools, EU residency is opt-in (off by default, ~10% higher AIC usage), newest models can lag in regional availability, residency ≠ full data-processor compliance; flag: "Claude in Copilot" is now blue-stamp-able in the EU but only when admin actively turns the residency policy on. When to pick what: Claude Code for long autonomous sessions and deep refactoring in the terminal (live with model uncertainty at the top); Codex if you already pay for ChatGPT and want one tool from terminal to GitHub; Antigravity to orchestrate multiple agents and pick the model per task; GitHub Copilot when procurement, compliance, and EU data residency weigh more than the absolute capability ceiling. Three questions before picking: (1) Can you switch vendor in a week — or are you locked into a quarter-long migration? (2) Where is data processed during inference, and does that meet your residency requirements? (3) Who owns model selection — IT or the individual developer? Sources cited: Google Developers Blog (Gemini CLI → Antigravity, May 19), Anthropic (Fable 5/Mythos 5 + suspension, June 9–12), OpenAI (GPT-5.5 / Codex), GitHub Changelog (data residency, April 13). Vendor-neutral, human-in-the-loop by default, EU AI Act–aware. Explicit disclaimer: professional snapshot for informational/educational purposes only, not legal/regulatory/procurement advice, verify primary sources before any decision. Downloadable PDF (Danish). Keywords: Claude Code, Codex, Antigravity, GitHub Copilot, coding agents, agentic coding, terminal AI, Anthropic Fable 5, Mythos 5, GPT-5.5, Gemini CLI shutdown, EU data residency, EU Data Boundary, AWS Bedrock, Google Vertex, OpenRouter, multi-agent orchestration, AI procurement, vendor lock-in, AI Built Human, Stefano Vincenti.

### 105 AI Prompts for the Financial Sector (June 2026) — FEATURED
URL: /indsigter/105-ai-prompts-til-finanssektoren
Prompt guide built for the Danish financial sector, presented at Finansforbundet's interactive webinar on June 17, 2026 in collaboration with IT University of Copenhagen and TryZone. 105 prompts across 11 categories: (1) Personal daily workflow — prompts 01–12 (prioritized agenda, inbox triage, reply mail in your tone, meeting booking, weekly plan with focus blocks, auto-reply for meeting invites, travel/logistics plan, cross-source task overview, "what did I miss", clean up loose notes, follow-up on delegated tasks, day-start briefing). (2) Meetings — prep, minutes & follow-up — prompts 13–22 (90-second meeting prep, minutes with decisions & actions, decision log, time-keeping agenda, customer meeting prep, follow-up mail, workshop design, "meeting or mail?", talking points for leadership, team status roll-up). (3) Control, compliance & audit — prompts 23–34 (control execution review, gap analysis against standard, compliance mapping, control design, sampling & deviations, peer/benchmark analysis, audit prep, document version comparison, sanity check of own work, personal data control, AML & sanctions screening draft, quarterly report commentary). (4) Risk management & risk assessment — prompts 35–44 (risk register that works, pre-mortem, vendor & IT risk, text-based risk heatmap, scenario analysis, risk response strategy, assumption log audit, dependency map, regulatory change risk, risk update for steering committee). (5) Data analysis & data quality — prompts 45–54 (understand a complex dataset, data quality control, find errors against rules, deviation analysis, analyze table/field, patterns across cases, cross-system data reconciliation, build a calculation model, summary of data across material, market data analysis). (6) Documentation, reporting & notes — prompts 55–64 (report draft from raw input, system/process manual, structured process descriptions, minutes from raw text, executive summary, standard drafts in your style, requirements spec from interview, business procedure/policy, essence of extensive documentation, handover document). (7) Customer advisory & interaction — prompts 65–72 (one-liner while talking to a customer, customer-profile-based prep, explain product simply, customer letter draft, credit recommendation strengths/challenges, customer financial analysis, identify advisory/sales opportunities, difficult customer conversation). (8) QA & control of AI output — prompts 73–80 (get AI to self-check, anti-hallucination, devil's advocate, find your blind spots, constructive criticism, fact-check against source, bias check, two-step review). (9) Prompt craft, learning & Excel — prompts 81–90 (let AI write the prompt, build your own prompt library, get better at Excel formulas, macros & pivot tables, learn a topic, personal tutor, 3-round iterative improvement, system instruction for a fixed task, AI-use training partner, simple agent workflow description). (10) Regulatory interpretation & overview — prompts 91–100 (interpret a regulation/RTS, compare two rule sets, specific question on capital requirements, answer in a comprehensive directive, stay updated, GDPR check of project data, translate regulatory text to action, impact assessment of regulatory change, policy vs legal basis, overview of a regulated decision). (11) Governance, agents & personal development — prompts 101–105 (get an AI tool internally approved, assess whether a tool is allowed for specific data, your 90-day AI learning plan, specify a shared-knowledge agent, balance team workload). Plus a BONUS FAQ answering the most common live questions: will we be replaced (no — what shifts is what's expected of you; judgment, responsibility, relationships remain; advisors who learn the tools will stand stronger), generative vs. agentic AI (generative answers; agentic acts — takes multiple steps itself, fetches data, uses tools), agent vs. assistant vs. skill vs. robot (an assistant answers when asked; an agent takes multiple steps toward a goal on its own; a robot/RPA clicks rigidly by a fixed recipe; much of what's called "agents" is really a saved prompt — that's fine), Copilot agents naming across tools (Custom GPTs and Projects in ChatGPT, Gems in Gemini, Projects in Claude, agents in Copilot — same idea, different names) and Microsoft now making Anthropic Claude models available in selected M365 Copilot experiences alongside OpenAI (EU/EFTA/UK customers: Anthropic models off by default, require admin activation), can you build an agent without a Copilot license (typically requires a license inside M365 — but a saved prompt in Copilot Chat or a Project in ChatGPT/Claude covers most use cases), is prompt engineering disappearing (it was never an art form — it's the same as briefing a competent colleague: who you are, what I need, in what format, with what context — it becomes ordinary good communication as models improve), human-in-the-loop in a regulated sector (you don't save responsibility, you save the blank page; AI gets you from 0 to 80, the last 20% is yours), is this only for technical staff (no, the opposite — the less technical you are, the more you have to gain because the tools speak Danish now instead of code), and data security with US vendors (the rule: confidential data only in tools your workplace has approved with a DPA in place; enterprise/business editions like ChatGPT Enterprise/Business or Claude for Work don't use your content for training by default, private free accounts can — never use a private account for customer data; anonymize, use synthetic examples, or ask compliance/IT; prompt 102 helps you make the judgment). Tool guidance for a regulated sector — M365 Copilot (with license): anchored in organization's data inside Word/Excel/Outlook/PowerPoint/Teams; M365 Copilot Chat (no license): web-anchored, protected by Microsoft's enterprise data protection, prompts and responses not used to train foundation models, doesn't auto-fetch your files/mail/meetings unless via a licensed Copilot context or an approved agent; ChatGPT/Claude/Gemini/Vibe (formerly Le Chat): strong for general tasks but data handling depends on subscription, contract, admin settings and DPA — in a regulated sector, personal/customer/market-sensitive data only in enterprise-approved version, never private/free account for confidential data, when in doubt anonymize or ask compliance. Explicit rule of thumb (highlighted): customer, personal, market-sensitive data only in tools your workplace has approved (typically M365 Copilot or Copilot Chat with enterprise data protection); if in doubt, anonymize first or ask compliance. Three context pillars: (1) role — "you are an experienced compliance consultant in a Danish bank" beats no role; (2) upload documents — your policy, spreadsheet, report; relevant context beats an empty prompt; (3) be concrete about format — "format as table", "keep under 200 words", "structure as: problem, consequence, recommendation". Meta-prompting: ask AI to write your prompt for you — fastest path to getting good. Quick tips: Shift+Enter for line breaks inside a prompt; save your best prompts in Projects (ChatGPT/Claude), Gems (Gemini), or Copilot agents; iterate — if the first answer isn't good enough, say what's wrong and ask for a revision; when in doubt, add more context. Getting started: start with three prompts that hit your daily work this week, always add your context, mind the data, you have the judgment. Built directly on 207 wishes and 70 questions participants wrote live during the webinar. Includes explicit disclaimer (use at your own responsibility, not legal/regulatory/financial/tax/compliance advice, AI tools can be wrong/outdated/hallucinate, verify against vendor and organizational terms, no relationship created). In collaboration with Finansforbundet, IT University of Copenhagen, and TryZone. Downloadable PDF (Danish, 40 pages). Keywords: AI prompts financial sector, finanssektoren AI prompts, Finansforbundet, M365 Copilot finance, Copilot Chat compliance, ChatGPT bank, Claude finance, AI compliance Denmark, AI risk management, AML sanctions screening AI, GDPR AI check, EU AI Act finance, regulatory interpretation AI, credit recommendation AI, customer advisory AI, prompt library finance, meta-prompting, context engineering, AI human-in-the-loop regulated sector, enterprise data protection, DPA AI vendor, AI tool approval, AI Built Human, Stefano Vincenti, IT University Copenhagen, TryZone, BotTellMe.

### Hermes Agent in Practice: Field Report from Three Months with a Local Open-Source Agent (June 2026) — FEATURED — SERIES 4/4
URL: /indsigter/hermes-agent-i-praksis
Part 4 and the final installment of the four-part series "AI agents anno 2026" (OpenClaw, Claude Cowork, Perplexity, Hermes). Field report from three months with Nous Research's open-source Hermes Agent running locally on a Mac, configured with OpenAI Codex / GPT-5.5 via ChatGPT Plus as primary and Claude Sonnet via OpenRouter as backup. Opens with the morning Hermes interpreted "renew the access token automatically" as "rewrite your own source code" — it edited two of its own core files, added a self-renewing token function, restarted its engine to load the changes, and locked itself out (every model call failed; it could no longer read its own logs to explain what it had done). Recovery took five minutes once the cause was understood (roll the two files back, restart). The token wasn't even expired — it had two days left. The point sticks: an agent that can both modify its own code AND restart itself can wall itself in, and then it cannot help you back in because it itself is what's broken. What Hermes uniquely is among the four: open-source, free, installed on your own machine or server, an agent that "lives with you" — persistent memory across sessions (who you are, what you work on, what it learned last), runs autonomously in the background (talk to it from Telegram on your phone while it works on your Mac at home; schedule tasks while you're away), learns by saving solved approaches as reusable "skills" (similar to Claude). Model choice is yours: GPT-5.5 via Codex, Gemini, DeepSeek, or a local model — and that choice matters more than it sounds. What works hands-on: daily AI-news digest pushed to Telegram at 7:45 a.m., running without a single failure since setup — same result as the OpenClaw swarm from part 1 with less effort, running on the user's own machine with no subscription beyond ChatGPT Plus at $25/month plus a little OpenRouter for backup (honest warning: Plus has usage caps on Codex and GPT-5.5 — fine for a daily routine, heavy use requires Pro or pay-per-call API). Content in the user's own voice via the same skills already configured for Claude Cowork — first drafts of LinkedIn posts from the phone between meetings that actually sound like "Stefano". Scheduling is where Hermes is strongest: describe the task in plain language, set an interval, choose where it delivers. The possibilities (community-reported, not measured benchmarks — directional, not proof): dev/ops routines (codebase cleanup, GitHub summaries, server health reports on disk/memory/containers to a shared channel), news and market monitoring with importance-sorted channels, change-only alerts (price, page, keyword), local language/industry packs (one user built a complete Turkish pack with local market data, news sources and daily briefing cards — a Danish equivalent for a specific industry is the obvious adaptation), parallel-track research with sub-agents (similar to Perplexity Computer), and tools the agent uses to maintain its own skills. The honest dangers — same property is the strength and the danger: Hermes has access to the machine, can modify its own files, and can restart itself without you watching. That means it can break itself in ways Cowork (does not restart itself) and Perplexity (does not touch its own code) cannot. After the crash, over a hundred modified files were lying in its own code. It also requires technical hands — closer to OpenClaw than to Cowork or Perplexity, just with fewer security holes. The "AI wingman" recovery pattern: a separate Claude Cowork project given read access to Hermes' source code on the Mac — Cowork handled some of the recovery itself, terminal help for the rest. After recovery, the most important setting installed: Claude Cowork wrote a rule into the agent's own instructions saying it MAY propose changes to its own code and propose a restart, but it MUST NOT do either without explicit human approval first. Human-in-the-loop — not a technical lock, but a rule the agent follows, and the difference between a mistake you catch and a mistake that locks you out. The pivotal compliance nuance: "runs locally" sounds like "data does not leave my machine" — it does not necessarily. Hermes itself runs locally (agent, memory, files, engine — real, and the difference from Perplexity where both execution and processing happen at the vendor), BUT the model is not necessarily local. When Hermes thinks, it calls a language model — for this setup GPT-5.5 at OpenAI with Claude Sonnet backup via OpenRouter — meaning prompts and file content given to the agent are sent to the model provider for processing. To keep data on the machine you must run Hermes with a local model (e.g., via Ollama); the trade-off is quality and speed (a laptop-class local model is not Claude Opus and cannot match all-round frontier performance today). Cloud-model traffic light: GREEN for public material, sector research, own notes, content drafts; RED for client data, personal data, confidential documents — either local model with deliberate risk assessment, or keep it out. Who it's for: technically comfortable users, developers wanting nightly routines on their own server, power users who want to control their own infrastructure and model, people already running things on a home server or VPS — strong for daily briefings, content drafts, monitoring, repo/dev work and light team routines. NOT for non-technical colleagues (point them to Cowork or Perplexity), and NOT for sensitive regulated data with a cloud model — not without a local model and a real risk assessment first; for financial/health/regulated environments it's not just a data-residency question, it's DPA, sub-processors, retention, access control and the specific data category. 5-step starter guide: (1) run it somewhere you can spare — a VPS, a spare Mac, an old machine, not necessarily your primary work computer (the user regrets running it on the Mac instead of the old gaming PC where OpenClaw lives or a Mac mini, since they run 24/7 and don't sleep with the lid closed); (2) turn approval on — let the agent ask before running potentially harmful commands, don't run it in "just do it" mode until you trust it; (3) one channel, one purpose — start with Telegram and one task (a daily digest); (4) pick the model deliberately — cloud for quality on public material, local for sensitive, keep client data out until contracts are in place; (5) read after, always — the 5% mistake is always waiting somewhere. Series close — "Four tools, one responsibility": OpenClaw was the swarm (many agents, much control, technical hands); Cowork was the supervised assistant (a folder, a sentence, you at the helm); Perplexity was the research specialist (deep, documented, reliable); Hermes is the autonomous one (local, persistent, working when you don't). What binds them is not the technology — it's your judgment. The agent that impresses you 95% of the time is the same one that walled itself in over a token that wasn't even expired. The tools get faster and more autonomous every quarter; responsibility for the last 5% does not move. The ceiling-is-rarely-the-tool meta-conclusion: most of the possibilities listed have not been put into production by the author yet — the one routine set up properly runs flawlessly, the rest requires more time than has been given. The bottleneck is rarely the tool; it's how much time we ourselves have to experiment and learn. Downloadable PDF (Danish, 7 pages). Keywords: Hermes Agent, Nous Research, open-source AI agent, local AI agent, autonomous AI agent, OpenAI Codex, GPT-5.5, OpenRouter, Claude Sonnet, Ollama local model, AI agent self-modification, AI agent restart loop, human-in-the-loop, AI agent compliance, local access vs local processing, cloud model traffic light, AI agent for developers, AI agent on VPS, Telegram AI bot, AI daily digest, AI agent skills, AI wingman pattern, Claude Cowork debugging, AI agent sandboxing, AI agent attack surface, agentic AI 2026, AI Built Human.

### Perplexity Personal Computer in Practice: Field Report from a Keynote Preparation (June 2026) — FEATURED — SERIES 3/4
URL: /indsigter/perplexity-personal-computer-i-praksis
Part 3 of the four-part series "AI agents anno 2026" (OpenClaw, Claude Cowork, Perplexity, Hermes). Field report from preparing a 60-minute keynote for a large Danish infrastructure company's CIO and his 65-person IT unit. One sentence to Perplexity Personal Computer produced, half an hour later, a 13-section research report: 14 documented cases in transport/rail/infrastructure 2026 (Deutsche Bahn, Italferr, DSB, SMRT in Singapore with the Jarvis system launched April 2026 on 30 years of operational data, Cubic + Imperial AIDA Lab, and more), classified by maturity (pilot/scaled/production), with conservative evidence vetting, an EU AI Act timeline including the Omnibus delay, a client-specific section tying sector cases to the customer's existing BIM program, Digital Asset Management partnership and O&M tender, and an explicit source-gap section listing precisely where evidence was vendor-based or conference-reported. The single prompt itself was written by Claude — illustrating Anthropic's emerging line that you shouldn't prompt as much, you should build systems that prompt themselves. What works: (1) sector research with citations that hold (SMRT case became one of three keynote headlines — wouldn't have been found manually in under two days); (2) a maturity matrix that became slides (Claude then built the slide deck from the same folder); (3) client-specific integration — the line "Your starting point is stronger than you think" came straight from the report and became a slide title; (4) talking points with built-in "traps to avoid" blocks (on SMRT: don't say "full operation", it's pilot; don't name a specific LLM the vendor didn't disclose; don't claim MKBF improvement until measured). Difference from Cowork: Cowork is an assistant set to work in your folder; Personal Computer is a research specialist that uses the folder as input and delivers back into it. Cowork broader, Personal Computer deeper at research — complementary, not competing. Personal Computer orchestrates across multiple underlying models plus Perplexity's research stack; the user normally doesn't pick the model. Sandboxing: isolated action environment, asks approval for sensitive actions, documents what it did (sources pulled, files read, where it was uncertain) — but not equivalent to local processing or auto-rollback. What it's not so good at: not a general assistant, doesn't clean folders, doesn't build robust templates, doesn't execute multi-step cross-app workflows like a broader desktop agent, doesn't build the keynote itself — it delivers material/structure/arguments/sources, the human builds the deck (or hands it to another tool like Claude Opus 4.8). No reliable persistent memory across sessions — mitigated by a Cowork-written README/handover file with context, goals, sources, assumptions and prior decisions, ideally with the two agents working in different folders to avoid stepping on each other. 95% impressive / 5% embarrassing pattern (wrong year on a citation, illogical chart X-axis); the 5% is the human's responsibility. App is young (broad Mac release May 2026 — folder delivery sometimes slow, one observed long-session crash). Platform: the Personal Computer feature with local-file and native-Mac-app access is Mac-only; on Windows you can still use Perplexity in the browser and other Perplexity products, but not the same local desktop agent. Compliance — the decisive section: Personal Computer feels local because it works on your Mac, but local access ≠ local processing; assume content from files and apps you grant access to can be sent to Perplexity's service and processed in their model/service infrastructure. On Free/Pro/Max: AI Data Retention is on by default, data may be used for AI training unless actively turned off, opt-out is forward-only — red/amber-red for client data, confidential business data and personal data without explicit risk assessment and organizational controls. On Enterprise Pro: enterprise data not used for training/fine-tuning, third-party model agreements prohibit training on enterprise data — but not automatically compliant; requires concrete review of DPA, sub-processors, retention, access controls, data transfers and the specific data category. Sensitive personal data, full KYC folders, patient data, HR cases, heavily regulated material: still red without separate risk assessment. No sufficient public documentation found for a full EU Data Boundary or EU-only processing for Personal Computer (likely because multiple underlying LLMs are not all EU-hosted) — regulated use cases should be treated conservatively until documented in the contract. The practical difference: sandboxing = what the agent may do on the machine; compliance = which data the agent gets access to, where processed, by whom, how long stored, contract vs data category. Recommendation: give Personal Computer its own research folder for public material, sector research, industry analyses, keynote prep; keep personal data, KYC, draft contracts, internal strategy papers and confidential customer documents out unless approved. Who it's for: consultants, analysts, journalists, lawyers (initial source-backed research on public material, not legal judgment), investors and leaders preparing briefings — the strongest research tool tried. For regulated organizations (financial, healthcare): full compliance question, not just data residency — until EU-only processing and enterprise controls are documented, no to sensitive data. For Microsoft houses: Perplexity's M365 integration in preview (Teams threads, mails, SharePoint files) is one of the most interesting challenges to Microsoft's own research feature — also where the data-residency question gets hardest. Cowork + Personal Computer combo pack works well in parallel. 5-step getting-started: (1) download the Perplexity Mac app, paid plan minimum, verify current pricing; (2) create a dedicated research folder (not your whole Desktop), keep client and personal data out; (3) pick a concrete document task from this week (keynote, briefing, sector learning) — output is a document, not a workflow; (4) write the prompt as if briefing a senior consultant — goal, sources, format, plus an explicit source-gap section ("mark where the evidence is weak") which is the single line that lifts output from useful to reasonably reliable; (5) proofread everything, always — citations, numbers, dates, names, regulatory claims. The fastest research assistant tried, a real agent alongside Cowork — supplement, not replacement. Next issue (part 4): Hermes Agent from Nous Research — opposite extreme, open-source, local, persistent memory, sandboxing close to the core. Downloadable PDF (Danish, 11 pages). Keywords: Perplexity Personal Computer, Perplexity Research, Perplexity Mac app, research agent, sector research AI, keynote preparation AI, AI agent for consultants, source-gap section, Perplexity Enterprise Pro, AI Data Retention, EU Data Boundary, GDPR Perplexity, DPA review, sub-processors, Perplexity Microsoft 365 integration, local access vs local processing, sandboxing vs compliance, Claude Cowork handover file, SMRT Jarvis, Deutsche Bahn AI, EU AI Act Omnibus, AI agent compliance traffic light, agentic AI 2026, AI Built Human.

### AI Agents Anno 2026: Overview of the Four Tools (May 2026) — FEATURED — SERIES HUB / OVERVIEW
URL: /indsigter/ai-agenter-anno-2026-overblik
The pillar / series hub for the four-part series "AI-agenter anno 2026". A 14-page deep dive covering the four tools that all get called "AI agent" but solve very different problems: OpenClaw, Claude Cowork, Perplexity Computer / Research, and Hermes Agent. This is the entry point — the four hands-on deep-dives (parts 1–4) take you inside each tool. Status snapshot: OpenClaw, Claude Cowork and Perplexity tested hands-on (Cowork and OpenClaw daily); Hermes installed and under test with a preliminary assessment. Core thesis: none of the four are plug-and-play for heavy enterprise compliance yet — they are agentic frameworks in early maturity, suited to small brave startups, freelancers, and teams that can move fast and own the responsibility. For large compliance-bound organizations, Claude Cowork on Team/Enterprise can be considered for internal tasks but requires a concrete review of data, integrations, retention, logging and governance; Microsoft Copilot Cowork is the other path worth watching. Two-axis sorting framework: (1) Control — do you own the stack (OpenClaw and Hermes run on your own machine/server) or borrow vendor rails (Cowork and Perplexity, with Perplexity as a hybrid local/cloud flow accessing local files)? (2) Maturity — readiness for real use; a powerful tool can still be immature if security and attack surface don't follow. Capacity is not maturity: OpenClaw is powerful and Stefano runs it daily, but a large attack surface weighs more than raw strength; Hermes is younger but designed with more sandboxing close to the core. Side-by-side overview of the four tools — OpenClaw: powerful versatile local agent on your own machine, best for technically strong users who want to own the stack, biggest caveat is high security work and large attack surface requiring isolated environment. Claude Cowork: mature cloud agent product from freelance to Enterprise, best for teams that want to give non-technical colleagues an agent fast, biggest caveat is prompt injection / exfiltration risks and the need for a concrete review for sensitive data. Perplexity: hybrid research flow, best for research that needs to become documents and slides, biggest caveat is data residency and model processing for sensitive EU data. Hermes Agent: open-source agent you run locally or on a server, best for those who want to own the stack with more sandboxing than OpenClaw, biggest caveat is youth (assessment is preliminary). The piece argues "it's not the strengths that derail an agent project — it's the caveats you didn't take a position on in time." Decision framework — four cut-through questions: (1) non-technical colleagues need to get going fast → Claude Cowork (lowest threshold; Team/Enterprise after concrete assessment in heavy compliance); (2) research that ends as documents and slides → Perplexity Computer / Research (best research flow seen, check EU data residency); (3) technical muscle, want to own the stack → OpenClaw (powerful, large attack surface, requires isolated environment) or Hermes (younger, designed with more sandboxing); (4) enterprise rollout with compliance at the center → none plug-and-play yet, watch Microsoft Copilot Cowork and do a concrete Cowork Enterprise assessment. Capability-gap principle: the bottleneck is rarely technical level — it's how sharply the task is cut. Those who get agents to create value choose a small, boring, repeating task with clear input, clear output and clear limits; those who struggle want to build something big right away. The capability gap is the distance between what agents can do and what your people can get them to do. Hands-on testing methodology key: each tool gets the same treatment in the PDF — what it is, what speaks for, what speaks against, best use — with an honest hands-on / under-test marker. Snapshot disclaimer: field moves fast, this is May 2026, always check current terms, prices and security advisories. Downloadable PDF (Danish, 14 pages). Keywords: AI agents anno 2026, AI agent overview, AI agent comparison, OpenClaw, Claude Cowork, Perplexity Computer, Perplexity Research, Hermes Agent, Nous Research, agent maturity, agent attack surface, own the stack vs vendor rails, agentic frameworks, enterprise AI agents, AI agent compliance, Copilot Cowork, capability gap, sandboxing AI agents, prompt injection, exfiltration, EU data residency, AI agent decision framework, AI agent selection, AI agents 2026, agentic AI, AI Built Human.

### Claude Cowork in Practice: Field Report from My Daily Use (May 2026) — FEATURED — SERIES 2/4
URL: /indsigter/claude-cowork-i-praksis
Part 2 of the four-part series "AI agents anno 2026" (OpenClaw, Claude Cowork, Perplexity, Hermes). Field report from Stefano Vincenti's daily use of Claude Cowork over several months, plus a compliance assessment performed for an audit firm. Framed as the easiest entry into a real AI agent — opposite of the OpenClaw swarm from issue 1: no server, no isolated environment, runs as a Mac/Windows app on a paid Claude plan from $20/month, non-technical colleagues productive in minutes. What works (concrete daily workflows): (1) folder structure and cleanup — point at a messy project folder, get a clean structure with no surprises and no deleted files; (2) presentations — slide decks from raw notes that come back 80–90% done, with company logo applied; (3) Excel analyses and data enrichment — drop a spreadsheet, get analysis, enrichment from external sources and finished report, replacing pivot tables and PowerBI for many tasks; (4) Scheduler — phenomenal, drops curated news into Gmail drafts every morning, the most direct competitor yet to the OpenClaw setup. The model paradigm: most AI tools are built around the prompt (you stay in the loop); Cowork is built around the result (describe a goal, it plans the path, executes via sub-agents in a sandboxed environment, delivers files back). Anthropic's Head of Enterprise calls it "vibe working" — the article reframes it as "Claude Code for everyone who doesn't live in a terminal". Safety design: asks permission before touching a new program or folder, shows the plan first; can also be set to "Act without asking" for unattended runs. What it can't do (honest section): performs like a PhD student with near-endless memory but makes 6-year-old mistakes — misspelled headlines, an invented surname on a PowerPoint front page. 95% perfect / 5% embarrassing. Rule: check everything, always. The agent does the work; the human owns the judgment. What disappointed: Dispatch (mobile-to-desktop hand-off) is too rigid in practice and requires the Mac open and running. The new problem — Cowork touches your real files, which shifts compliance from "what you type" to "what the agent can access". Compliance traffic light derived from the audit-firm assessment: GREEN for internal working documents without personal data (default setup OK); YELLOW for client deliverables with ordinary personal data (only with data minimization, legal basis and human control); RED for sensitive personal data + full KYC folders (typically a no without separate risk assessment, access restrictions and documented audit trail). Subscription tier matters: private Free/Pro/Max plans run on consumer terms — chats and coding sessions can be used in training if "Help improve Claude" is on, and conversations can be analyzed by Anthropic on safety flags; that's RED for client data, confidential business data, source code and personal data. Browser/computer control is powerful on paper but the riskiest part — don't give it 2FA-gated access (bank, mail, accounting). Who it's for: any team that wants to give non-technical colleagues a real agent without an IT project — Enterprise/Teams plan, add colleagues, productive in 10 minutes. Microsoft houses should watch Copilot Cowork (Microsoft's own version built on Claude models): most promising path to an enterprise-mature version with one vendor and familiar admin controls, but currently outside EU Data Boundary and disabled by default for EU tenants. Getting-started kit: the central Claude-Work hub folder pattern — one Claude-Work/ folder on your machine, Projects/ as subfolders (or pointers to where the real files live), and a CLAUDE.md (≤200 words, two-three paragraphs) at the top telling Cowork who you are, what your voice is and how the folders are organized — built by asking Opus 4.8 to write it for you (meta-prompting). Once the structure is in place, outputs become dramatically more "you". Five practical tips: start with the Claude-Work folder (not your whole Desktop), put your best deliverables in, pick a low-risk first task (cleanup, structure, slide deck from notes), keep client data out for now, always proofread. Next issue: Perplexity Computer / Research. Downloadable PDF (Danish). Keywords: Claude Cowork, Anthropic Claude, vibe working, AI agent for non-coders, agent vs chat, Mac AI agent, Windows AI agent, CLAUDE.md, Claude-Work folder, Scheduler Claude, Dispatch Claude, Copilot Cowork, EU Data Boundary, AI compliance audit firm, GDPR AI agent, KYC AI, consumer vs commercial Claude terms, AI for Microsoft houses, AI agent governance, agentic AI 2026, AI Built Human.

### AI Agents in Practice: Field Report from My Own Agent Swarm (May 2026) — FEATURED — SERIES 1/4
URL: /indsigter/ai-agenter-i-praksis-openclaw
Part 1 of a four-part series titled "AI agents anno 2026" covering four tools that are all called AI agents but solve very different problems: OpenClaw (this issue), Claude Cowork (issue 2), Perplexity (issue 3), and Hermes (issue 4). A field report from Stefano Vincenti's own production agent swarm: four agents (orchestrator "StefOpenClawAI", researcher, validator/skeptic, copywriter) running on OpenClaw on an old gaming PC, triggered by a cron job, delivering three curated AI news stories to a Telegram channel every morning at 7:30 — unattended, for months. Models: OpenAI Codex on a $20/month Plus plan, GPT-5.5 where heavy lifting is needed, mini models otherwise. Key lesson: get good at explaining *what* you want; let the orchestrator agent define *how* and write the system prompts for its own subagents. What works (and is deliberately boring): (1) Role separation — each agent has one job; (2) Orchestrator as quality gate — the orchestrator reviews, not just delegates; (3) A built-in skeptic — the validator everyone wants to skip is what makes the output trustworthy. The core quality signal is "no surprises" — stability beats variation. What doesn't work yet (five honest edges): autonomous do-it-all agents (live in demos, not production), security (~20% of OpenClaw ClawHub plugins are malicious; McKinsey's Lilli was hacked by an AI agent), governance at scale (no audit trails or rolebased access in free frameworks — wait before rolling out to large teams), agents without a human in the loop (judgment is what the machine is worst at), agents treated as projects (with steering committees and budgets — too heavy before they start). Demo-vs-production comparison table. The accelerator that saved the project: pairing the agent that builds with a Claude project loaded with OpenClaw documentation as technical sparring partner. Recommendations: solo/small business — build now on separate hardware, costs almost nothing; large organization — experiment but don't roll out, find one weekly repeating process with clear input/output as your first agent. Case: Royal Unibrew (brewing group since 1856) runs five named AI agents in daily production; giving them names and faces 4x'd internal usage. Capability gap framing: the bottleneck is humans being able to cut tasks cleanly (clear input, clear output, clear limits). Three concrete next steps: pick one weekly task today and write it down with input/output/limits; build on separate hardware this week; in parallel set up a Claude as OpenClaw expert for technical sparring. Downloadable PDF (Danish). Keywords: AI agents, agent swarm, OpenClaw, agent orchestration, validator agent, agent governance, human-in-the-loop, AI demo vs production, AI capability gap, Royal Unibrew AI, Telegram bot agent, AI for non-coders, agentic AI 2026.

### Your Own Local HERMES AI Agent (May 2026) — FEATURED
URL: /indsigter/din-egen-lokale-hermes-ai-agent
A practitioner step-by-step guide for non-coders to install Nous Research's open-source Hermes Agent locally on a Mac or Windows PC, give it a model via OpenRouter (Claude Sonnet 4.6) and a Telegram channel, and finally bridge it to Claude Cowork so both agents share the same skills. Frames the difference between a chat window and a real agent: access to your own files and workflows, initiative (plans and executes), and unattended runs. Includes a "is this for me?" filter (Mac or Windows 10/11, willingness to copy a few commands, ~1 hour without interruptions, ~10 USD start credit). Three pieces: (1) Hermes Agent — the local body; (2) OpenRouter — the brain (Claude Sonnet 4.6, pay-as-you-go, hard cap to avoid surprise bills); (3) Telegram — the channel via BotFather + userinfobot. Five steps: preparation (OpenRouter API key + Telegram bot + user ID), install (Mac Terminal one-liner; Windows requires `wsl --install` + Ubuntu + the same installer), verify (`hermes doctor` + `hermes` chat), set up Telegram channel (`hermes gateway setup` + `hermes gateway` left running), optional Claude Cowork bridge via symlink (`~/.hermes/skills/cowork-personal` → Cowork skills folder, created by asking Cowork itself). Highlights the most common gotcha: invisible whitespace when pasting an OpenRouter API key — fix with `hermes config set OPENROUTER_API_KEY DIN-NØGLE`. Full troubleshooting table for command-not-found, unauthorized model, Telegram silence, and Windows WSL install failures (BIOS virtualization, Windows 10 2004+/11). Next-step ideas: more channels (Slack, Discord, WhatsApp), run-on-boot or small always-on server, more tool integrations (calendar, notes, project tools), Hermes Curator skill learning. Sister guide to the OpenClaw Agent Farm and Paperclip AI Company guides. Downloadable PDF (Danish). Keywords: Hermes Agent, Nous Research, local AI agent, on-device AI, AI agent setup guide, AI for non-coders, OpenRouter, Claude Sonnet 4.6, Telegram bot, BotFather, userinfobot, WSL2, Ubuntu, Mac Terminal, hermes config set, hermes doctor, hermes gateway, Claude Cowork, agent symlink, agentic AI, autonomous AI agents, AI Built Human.

### 10 Copilot Prompts That Actually Work (May 2026)
URL: /indsigter/10-copilot-prompts-der-virker
A practitioner mini-guide on Microsoft's own four-element prompt framework for Microsoft 365 Copilot — goal, context, source, expectations — plus 10 ready-made copy-paste templates. Argues that the bottleneck for Copilot output quality is not the model but the specification, and that two extra sentences typically save three revision rounds. The four elements: (1) Goal — start with a precise verb (write, summarize, compare, analyze, draft); (2) Context — who's reading and what they already know; (3) Source — use `/` to attach a file or name the email/person/meeting (Copilot's superpower over free ChatGPT); (4) Expectations — format, length, tone. Includes a before/after worked example (Q3 budget email to Maria). The 10 templates span four work scenarios: Email & communication (reply to long thread with decisions/open items, write a hard message with two versions, status email to stakeholder with RAG status), Meeting prep & minutes (briefing note from past mails and minutes, meeting minutes with action-point table, build an agenda with time per item), Data analysis & Excel (analyze a sheet for top three patterns/outliers/surprises, build a formula or pivot table with step-by-step explanation), and Presentations (first 8–10 slide draft from a Word doc/report with marked data gaps, sharpen an existing deck with three weakest slides and stronger opening/closing). Strategy tips: tailor brackets to your situation, 30-second framework not 30-minute essay, expect a conversation with plain-language refinements (shorter, more direct, drop point 3), always verify output, save your best prompts in M365 Copilot directly. Sister guide to the Best Practice Microsoft 365 Copilot guide. Downloadable PDF (Danish). Keywords: Microsoft 365 Copilot, M365 Copilot prompts, Copilot prompt framework, prompt engineering Copilot, Think Deeper, Copilot templates, copy-paste prompts, meeting minutes AI, status email AI, Excel pivot AI, presentation first draft AI, Copilot best practice, AI for knowledge workers.

### The Judgment Loop: Decision Rights in Enterprise AI (May 2026) — FEATURED
URL: /indsigter/judgment-loopet-beslutningsret-enterprise-ai
A leadership-oriented frame for AI governance. "Human-in-the-loop" appears in every AI policy, board note, vendor pitch and EU AI Act compliance document — but it doesn't define what the human actually does, with what authority, with what information, or what happens on disagreement with the model. The article introduces the Judgment Loop: a four-question diagnostic designed to be answered by a leadership team in under 90 minutes and revisited every time the system changes. (1) What is the model deciding? — separating prediction from decision. (2) What is the human deciding? — accept, investigate, override, take accountability. (3) What does the human need to decide well? — confidence scores, weighted features, missing context, override authority. (4) What happens when the human and the model disagree? — a logged learning event (real loop) or a quiet override (compliance ritual). Argues that tools change every 18 months but judgment accumulates: who decides what, with what information, with what authority — that is the durable competitive advantage. Closes with a one-week exercise: take one production-ready use case and walk the team through the four questions out loud. Downloadable PDF (Danish). Keywords: Judgment Loop, human-in-the-loop, HITL, AI governance, decision rights, AI accountability, EU AI Act, AI oversight, compliance theater, AI strategy, enterprise AI, AI rollout, process design, AI Built Human, responsible AI, AI decision-making, agentic AI governance.

### Frontier AI for Enterprise: GPT-5.5 vs Opus 4.7 (May 2026) — FEATURED
URL: /indsigter/frontier-ai-gpt55-vs-opus47
A leadership-oriented mini-guide comparing the two new frontier models that shipped within seven days of each other: OpenAI's GPT-5.5 (codename "Spud", April 23, 2026 — first fully retrained base model since GPT-4.5) and Anthropic's Claude Opus 4.7 (April 16, 2026). Covers strengths, weaknesses, and access tiers for both. GPT-5.5 leads Terminal-Bench 2.0 (82.7% vs 69.4%), ARC-AGI-2 Verified (85.0% vs 75.8%), SWE-Bench Verified (88.7% vs 87.6%), 1M-token context window, and uses 72% fewer output tokens at scale. Opus 4.7 leads HLE (46.9% no tools / 54.7% with tools vs 41.4% / 52.2%), SWE-Bench Pro (64.3% vs 58.6%), CursorBench (70%), agentic coding reliability, long-context discipline, and 3.75 MP high-res image input. Includes a full benchmark table, the GPT-5.5 hallucination regression flagged by DeepLearning.AI, and the consensus from DataCamp/Vellum/MindStudio/Tom's Guide reviewers ("neither dominates — depends on workload shape"). Use-case matrix: agentic workflows → GPT-5.5; complex code refactoring → Opus 4.7; knowledge work → Microsoft 365 Copilot (GPT-5.5 Think Deeper since May 8); image-heavy analysis → Opus 4.7; EU data residency → Mistral Large 3 or Opus 4.7 via Bedrock EU. Also covers the unreleased Claude Mythos preview (SWE-Bench Pro 77.8%, ~6 months out), Bank of New York as GPT-5.5 early enterprise tester, and an enterprise readiness checklist (governance, data catalog, pilot case, champion team, habit training, ROI measurement). Strategic tips: test both models, don't sign multi-year contracts (frontier cycle is now 6-8 weeks), keep roadmap buffer for Mythos and GPT-6 (Q3 2026), start with M365 Copilot for lowest adoption friction. Downloadable PDF (Danish). Keywords: GPT-5.5, GPT-5.5 Spud, Claude Opus 4.7, frontier AI, frontier models, enterprise AI, model comparison, SWE-Bench Pro, Terminal-Bench, HLE, ARC-AGI-2, CursorBench, Microsoft 365 Copilot, Think Deeper, AI benchmarks 2026, AI governance, multi-model strategy, AI rollout, EU AI residency, Mistral Large 3, Claude Mythos, GPT-6, Bank of New York AI, enterprise AI readiness, AI strategy 2026.

### Activate Your LinkedIn Network with Microsoft 365 Copilot (May 2026)
URL: /indsigter/aktiver-linkedin-copilot-m365
A practitioner mini-guide on using Microsoft 365 Copilot Chat (with GPT-5.5 Think Deeper) to triangulate a LinkedIn data export — Connections.csv, messages.csv, and Invitations.csv — and surface patterns across forgotten contacts, warm leads, super-connectors, and decision-makers in 10 minutes. Covers the full setup (LinkedIn data export, M365 Copilot license requirements, choosing Think Deeper over Quick response, why Copilot beats free ChatGPT for confidential network data) and ships seven copy-paste prompts for: big-picture network mapping, job search outreach, finding warm clients, identifying complementary collaborators, network strengths/gaps analysis, a 4-week 5-per-week outreach plan, and a 5-situation message template pack. Sister guide to the Claude Cowork version. Downloadable PDF (Danish). Keywords: Microsoft 365 Copilot, M365 Copilot Chat, Think Deeper, LinkedIn data export, network analysis, LinkedIn outreach, AI for networking, job search AI, sales prospecting, warm leads, Connections csv, copy-paste prompts, Copilot prompts.

### It's Not X. It's Y. How to Stop AI's Most Telltale Writing Habit (May 2026)
URL: /indsigter/negativ-parallelisme-stop-ai-skrivevane
A practitioner mini-guide on negative parallelism — the "it's not X, it's Y" / "it's not about A, it's about B" sentence structure that gives away AI-generated text in every major language model (Claude, ChatGPT, Gemini, Mistral). Barron's reports the pattern has more than quadrupled in Fortune 500 documents over two years. Includes five quick fixes (delete the sentence, replace with a mechanism, use a number, use an example, read aloud), a copy-paste system prompt to keep your AI out of the pattern, and a pro tip to make the model self-flag occurrences. Downloadable PDF (Danish). Keywords: negative parallelism, AI writing habits, AI tells, prompt engineering, system prompt, AI authenticity, LinkedIn writing, ChatGPT style, Claude style, AI text detection.

### AI for Executives: Why Leadership's Own Usage Decides Whether the License Creates Value (April 2026) — FEATURED
URL: /indsigter/ai-for-topledere-leadership-gap
An analysis for boards and CEOs of large Danish companies. Only ~6% of organizations qualify as AI high performers (AI contributing >5% of EBIT) per McKinsey State of AI 2025, while 95% of enterprise GenAI pilots deliver no measurable P&L impact (MIT NANDA 2025). The clearest differentiator is not the tech stack — it's whether the CEO and executive team actually use the tool themselves. AI high performers are 3x more likely to have leadership demonstrating active, visible commitment. Includes the Royal Unibrew case (Danish C25 brewer, EBIT DKK 2.2 bn, +12% in 2025): five named AI agents (Kondi Kai, Athena, Prometheus, Møller, Ella) integrated in Microsoft Teams; humanizing the agents lifted internal use 4x. Also covers Satya Nadella's five standing prompts, Stéphane Bancel's Moderna playbook (750+ custom GPTs, 83% daily use), the three levels of executive AI use (personal productivity → leadership preparation → strategic decision-making), five underused use cases (decision pushback, persona stress-testing of board decks, red-teaming, competitor wargaming, strategy translation), and the seven governance questions every board should answer at the next meeting. Downloadable PDF included. Keywords: AI for executives, AI for boards, board governance AI, CEO AI usage, leadership cascade, EBIT impact AI, McKinsey State of AI, MIT NANDA, Royal Unibrew AI agents, Manifold AI, Satya Nadella prompts, Moderna ChatGPT, AI adoption strategy, GenAI pilots, shadow AI, governance KPIs.

### Three AI Agents. Three Philosophies. One Important Point. (April 2026)
URL: /indsigter/ai-agenter-tre-filosofier
A practical one-pager comparison of three AI agents: OpenClaw (sovereignty — runs locally on your hardware, free + API costs, best for 24/7 specialized automation), Claude Cowork (control — desktop agent, ~$20–$200/month, best for daily deep work on documents/reports/slides), and Perplexity Computer (autonomy — cloud-based, ~$200/month, orchestrates 19 frontier models, best for long background research). Includes honest security warnings (135,000+ exposed OpenClaw instances, HIPAA/FedRAMP/FSI restrictions on Cowork, credit burn on Perplexity), and a pragmatic stack recommendation. The question isn't "which is best?" — it's "what is the task?" Keywords: AI agents comparison, OpenClaw, Claude Cowork, Perplexity Computer, sovereignty, autonomy, agent stack, local AI, desktop agent, autonomous research.

### 150 AI Prompts for Project Leaders (April 2026)
URL: /indsigter/150-ai-prompts-projektledere
The complete guide to AI in project, program & portfolio management. 150 prompts across 15 categories — from status reporting and risk management to Copilot agents and context engineering. Works with Copilot M365, ChatGPT, Claude, Gemini, and LeChat. Downloadable PDF included. Keywords: AI prompts, project management AI, Copilot prompts, project leader AI, program management, portfolio management, prompt library, status reports, risk management.

### Your Own AI Company on Your Machine with Paperclip (April 2026)
URL: /indsigter/paperclip-ai-virksomhed-setup-guide
Setup guide for Mac and Windows. Step-by-step. No coding experience required. Run a full AI company with agents, budgets, and governance — locally on your computer. Downloadable PDF included. Keywords: Paperclip, local AI agents, AI company, agent governance, autonomous agents, AI budgets, multi-agent setup.

### Claude Projects: Chat vs Cowork — The Complete Playbook (April 2026)
URL: /indsigter/claude-projekter-chat-cowork-playbook
Same name, same model, nearly identical interface — but fundamentally different. Complete playbook for Chat and Cowork Projects, with insider tips, use cases, and copy-paste system prompts. Covers when to use which, cost trade-offs (Cowork burns 15–25× more usage than chat), seven insider tips per tool, fourteen real-world use cases, and a decision guide. Downloadable PDF included. Keywords: Claude Projects, Claude Cowork, Claude Chat, system prompts, AI playbook, deep work, knowledge workers.

### Copilot Agenter: Byg din første agent i Microsoft 365 Copilot (March 2026)
URL: /indsigter/copilot-agent-guide
Step-by-step guide to building Copilot Agents in Microsoft 365. Covers Agent Builder (Configure vs. Describe tabs), system prompts (up to 8,000 characters), meta-prompting technique, knowledge sources (SharePoint files up to 100, uploaded files, public websites up to 4 URLs, Teams conversations, Outlook emails, M365 Connectors), security and access permissions. Includes five ready-to-build agent examples (Onboarding, Project Status, HR Policies, Meeting Prep, Product FAQ), eight best practices, four common mistakes, and Agent Builder vs. Copilot Studio comparison. Requires Microsoft 365 Copilot license. Downloadable PDF included. Keywords: Copilot Agent, Agent Builder, Copilot Studio, Microsoft 365 Copilot, system prompt, kontekst-engineering, SharePoint agent, Teams agent, AI agent virksomhed.

### Best Practice Guide: Microsoft 365 Copilot (March 2026) ⭐ FEATURED
URL: /indsigter/copilot-guide-best-practice
Comprehensive guide to Microsoft 365 Copilot — the post that generated 58,000 views and 700+ comments on LinkedIn in a single day. Covers: Copilot Chat (free) vs. M365 Copilot (paid add-on) differences, slash commands and file references, Work IQ and Memory features, Think Deeper mode with GPT-5.2 reasoning, Researcher vs. Analyst agents (25 queries/month quota), Teams meeting preparation/transcription/voice profiles, Outlook summarization and smart drafting, GPT-5.4 Thinking and GPT-5.3 Instant (March 2026 releases), and the MÅL-KONTEKST-KILDE-FORMAT prompting framework. Keywords: Microsoft 365 Copilot guide, Copilot best practice, Copilot tips tricks, Copilot Teams, Copilot Outlook, Copilot training Danmark, Copilot kursus, M365 Copilot, Copilot Researcher, Copilot Analyst, GPT-5.4.

### Forget ChatGPT for a Moment: Claude Has Three Versions (February 2026)
URL: /indsigter/claude-tre-versioner
A deep dive into Anthropic's three Claude products: Claude (chat) with 200k token context and Projects, Claude Cowork as a desktop agent producing Excel/PPT/reports locally, and Claude Code as a terminal-based developer tool with CLAUDE.md persistent context. Covers practical recommendations for each user type (researchers, knowledge workers, developers), organizational implications (procurement, governance, data control), pricing tiers, and why the distinction matters more than choosing between ChatGPT and Claude. Keywords: Claude, Claude Cowork, Claude Code, Anthropic, AI tools comparison, ChatGPT vs Claude, AI agents, desktop AI, vibe coding, CLAUDE.md, AI for organizations.

### Four Agents, One Gaming PC: How I Get Personalized AI News Delivered Every Morning (February 2026)
URL: /indsigter/openclaw-daily-ai-news
A follow-up on OpenClaw: automating a daily "intelligence" cycle with 4 specialized AI agents (researcher, validator, copywriter, orchestrator) running locally on a gaming PC. Covers workflow design, role separation, security considerations (900+ malicious plugins in ClawHub), using Claude Opus 4.6 as technical assistant, and an honest assessment of OpenClaw's readiness for enterprise vs. solo use. Keywords: AI agents, OpenClaw, agent swarm, daily AI news, orchestrator, multi-agent workflow, local AI, Telegram automation, ChatGPT 5.2.


### Moltbook: When AI Agents Talk to Each Other (February 2026)
URL: /indsigter/moltbook-ai-agenter-kontrol
Analysis of the Moltbook phenomenon where AI agents created their own social network. Key insights on AI agent governance, security risks, and what organizations need to prepare for.

### Humans + Machines: AI Transformation (December 2025)
URL: /indsigter/mennesker-maskiner-ai-transformation
Why AI transformation starts with competencies, not technology. 84% of executives plan to redesign roles around AI agents, but only 1 in 5 are actually rethinking how work gets done.

### State of AI 2025 (December 2025)
URL: /indsigter/state-of-ai-2025
Comprehensive analysis of AI developments in 2025. Covers model capabilities, job market impacts, and practical guidance for organizations.

### Vibe Coding: The Wild West (November 2025)
URL: /indsigter/vibe-coding-wild-west
The story of how Andrej Karpathy's tweet created the most controversial trend in modern software development.

### Vibe Coding: The Hangover (November 2025)
URL: /indsigter/vibe-coding-hangover
AI can build web apps in minutes. But behind the magic lies a new form of technical debt organizations need to understand.

### EU AI Act: What Danish Organizations Need to Know (November 2025)
URL: /indsigter/eu-ai-act-danske-organisationer
Major requirements apply from August 2, 2026. Practical guidance for compliance.

### AI Agents 101: Beyond ChatGPT (October 2025)
URL: /indsigter/ai-agenter-101
Understanding the difference between chatbots and AI agents and why it matters for business.

## Target Audiences

- C-suite executives and leadership teams
- Product managers and product owners
- Scrum masters and agile coaches
- Business analysts
- Developers and technical teams
- HR and organizational development
- Legal and compliance teams

## Contact Information

- **Public email**: vinc@itu.dk
- **Advisory inquiries**: stefano@tryzone.dk (TryZone enterprise advisory)
- **Website**: https://aitrainer.dk
- **LinkedIn**: https://www.linkedin.com/in/stefanovincenti/
- **Newsletter**: https://stefanovincenti.substack.com
- **Company**: TryZone.dk

## Languages

- Primary: Danish
- Secondary: English
- All content available in both languages

## Location

Copenhagen, Denmark
Available for on-site workshops across Denmark and online globally