Insights
Knowledge that's correct —
and gets cited
No content mill, just pieces that explain how reliable AI really works. Written for people who want to understand it properly — and structured so AI assistants pick it up correctly (GEO).
Why RAG without citations is worthless
RAG is not a miracle cure for hallucination. Without verifying the source, you get a model that lies convincingly. This is the difference between a demo and a production system.
Read the articleMore from Insights
How do you see what your agent did last night?
An agent that runs autonomously overnight should be traceable the next morning. What to log, how tracing with something like OpenTelemetry helps, and how to build a simple audit trail and dashboard.
ReadHow do you give an AI agent safe access to your files?
Keep the scope as small as possible: an explicit allowlist instead of a denylist, separate read and write permissions, backups before every mutation, and a tool layer that enforces the boundary outside the model.
ReadThe EU AI Act for SMEs: what do you need to have in place right now?
Since 2 August 2026, the AI Act's transparency rules apply and enforcement has begun. The high-risk system rules have been postponed to 2027/2028, but AI literacy, banned practices and chatbot transparency already apply. A practical overview for SMEs — not legal advice.
ReadWhy your AI operator should be allowed to do nothing
An agent that can intervene on its own feels powerful — and that is exactly why it is unreliable. On the separation between observing, reasoning, acting and remembering, and the guardrails that are the bare minimum around it.
ReadAI agents for SMEs: what actually works today (and what doesn't yet)
Not every "AI agent" delivers on its promise. What already works reliably for SMEs today — searching documents, fixed workflows, supervised monitoring — and where full autonomy is still a risk rather than a solution.
ReadDo You Need to Label AI-Generated Content?
The EU AI Act does not require a blanket visible label on all AI content. Article 50 distinguishes between technical marking by providers, deepfakes, and AI text on matters of public interest. A practical overview — not legal advice.
ReadWhich AI crawlers visit your site — and how do you spot them in your logs?
GPTBot, ClaudeBot, PerplexityBot and more do not all do the same thing: some gather training data, some build a search index, some fetch a page live at a user’s request. How to filter and verify them in your logs.
ReadAI and GDPR: what's allowed, what isn't, and how to get it right
Using AI with personal data is allowed under the GDPR — provided you have a legal basis, apply data minimisation, and know where the data goes. A practical guide: from legal basis to DPIA, and where the EU AI Act adds requirements on top.
ReadCan AI search case law without making things up?
AI can search case law without inventing judgments when answers are grounded in real sources. How RAG, source links and verification rules reduce hallucinations in legal research.
ReadAI literacy (Article 4): what does the law expect from your staff?
Article 4 of the AI Act requires providers and deployers to take measures supporting AI literacy among staff — proportionate to role and risk, with no mandatory certificate. In force since 2 February 2025, amended in 2026.
ReadAI in construction: making specifications, standards and drawings searchable
Specifications, permits, technical standards and drawings rarely live in one place. How RAG makes those documents searchable, why it is a difficult use case, and why on-premises processing often makes sense here.
ReadAI in legal practice: what can a law firm use it for today?
Contract review, case-law research and file search already work well — as long as every claim traces back to a real source. What a law firm can use AI for today, why hallucination is the core risk, and why client confidentiality means thinking about on-premises or EU hosting.
ReadAI in healthcare: why on-prem is almost always the answer
Health data is a special category of personal data under GDPR. That's why on-prem or a tightly controlled private environment is usually the most defensible architecture for AI working with patient records — not an absolute ban on cloud AI, but a much higher bar.
ReadWhat does an AI agent actually cost? Measure for a month instead of guessing
The monthly bill arrives and nobody knows which component caused it. How to log AI costs per call — and why input tokens are the real cost driver with agents.
ReadMonitoring AI services: which signals actually matter?
An AI service can be technically "up" while users get slow, expensive or low-quality answers. Which signals beyond uptime — latency, cost, drift, hallucinations — your monitoring actually needs to catch.
ReadAI for accountants: where's the upside, and where's the risk?
The upside of AI for accountants lies in faster document processing, journal-entry suggestions and periodic close support. The risk lies in hallucinated figures, tax interpretation, GDPR and liability.
ReadAI for Municipalities: Making Policy Documents Searchable
Policy papers, council records, permit files and local regulations become searchable by meaning instead of just keywords with RAG, with citations back to the original document. Why the Woo and data sovereignty matter, and when a municipal AI application can qualify as high-risk under the AI Act.
ReadAir-gapped AI: can a language model run without the internet?
Yes — a language model can run entirely disconnected from any network. What air-gapped AI is, how it differs from plain on-prem, what it takes technically, and why most businesses don't need it.
ReadArticle 50 explained: when do you have to disclose that it's AI?
Article 50 of the EU AI Act does not require businesses to disclose every use of AI. The transparency obligation applies specifically to chatbots, AI-generated content, deepfakes and emotion recognition. A practical overview for SMEs — not legal advice.
ReadShould you block AI crawlers or let them in?
Blocking every AI crawler is usually too blunt. GPTBot, ClaudeBot, PerplexityBot, Google-Extended and CCBot each serve a different purpose — how to configure robots.txt for training versus visibility.
ReadCaching for AI: which answers are safe to reuse?
AI responses are safest to reuse when the same request, under the same conditions, should produce the same valid result. Time-sensitive, personalised or creative answers usually aren't — and semantic caching has its own risks.
ReadChatbot disclosure: how do you say it without scaring visitors off?
A chatbot disclosure doesn't need to sound like a legal warning. Where the notice belongs, five example sentences from subtle to explicit, and when to be more explicit in healthcare, finance or with minors.
ReadChunking explained: how do you split a document without breaking its meaning?
Good chunking determines what your RAG system can retrieve in the first place. Why chunking goes wrong, how to choose chunk size and overlap, and how to chunk PDFs, tables and code differently.
ReadWhy your AI agents start over every session — and how to fix it
The most expensive problem in agent work is memory loss: every session starts blank. On the memory layer that turns separate conversations into one continuous, traceable brain.
ReadData residency: what does an 'EU region' really mean with a US cloud provider?
An EU cloud region does not automatically mean European control. What data storage location, access, the US CLOUD Act and cloud sovereignty really mean for your AI project.
ReadHow do you anonymise documents before they enter a model?
Documents should be anonymised before AI processing by detecting personal information and removing, masking or replacing it — ideally locally, before text reaches an external model or embedding API. Pseudonymisation is often more practical than full anonymisation, but the resulting data remains personal data under the GDPR.
ReadWhat is a DPIA and when do you need one for AI?
You need a DPIA for an AI system when its processing of personal data is likely to result in a high risk to individuals. AI itself is not the legal trigger; the combination of personal data, purpose, scale, technology and possible consequences is what determines the risk.
ReadTraining your own model? Start with RAG.
Many organizations dream of "our own LLM". In practice that is almost always the most expensive route to the wrong goal. Why RAG plus light fine-tuning usually wins — and when training genuinely makes sense.
ReadChoosing embeddings: which model understands Dutch best?
There is no embedding model you can call objectively "best for Dutch" without testing it yourself. E5-NL, Qwen3 Embedding, Jina Embeddings v5, BGE-M3 and multilingual-e5 are candidates — MTEB-NL and your own retrieval test should decide the winner.
ReadFallback chains: what does your system do when the provider goes down?
A fallback chain combines alternative AI providers or models with health checks, timeouts, retries and circuit breakers. Downtime is not the only failure mode — measurable quality degradation counts too.
ReadWhy Does Your FAQ Page Rank Worse Than One Good Question Per Page?
A long FAQ page makes several topics compete for attention, while search and retrieval systems often process content in smaller passages. When does a question deserve its own knowledge-base page, and when does it belong in the FAQ?
ReadGEO: how to get cited by ChatGPT and Gemini
SEO got you into Google. GEO gets you into the answer the user actually reads. What Generative Engine Optimization is, how AI assistants pick their sources, and what you can do today to get cited.
ReadGetting found by AI assistants: what GEO actually is
Classic SEO optimizes for a results list. GEO optimizes for being cited as a source in a generated answer. What that means technically and editorially — including checklists.
ReadWhat hardware do you need to run a language model locally?
RAM, VRAM, quantization and storage determine which LLMs you can run locally. Practical memory rules for 7B to 70B+ models on CPUs, GPUs and Apple Silicon.
ReadHow do ChatGPT, Perplexity and Gemini decide which source to cite?
Ranking well in Google is not the same as being cited in an AI answer. How ChatGPT Search, Perplexity and Gemini retrieve, assess and link sources to specific claims — and what that means for GEO.
ReadHow many documents do you need before RAG is worth it?
There is no minimum. RAG earns its complexity once documents no longer fit a single context, content changes often, many users need to search, or answers need citations — not at a fixed file count.
ReadWhat is a high-risk AI system — and are you running one?
Not every advanced AI system is high-risk under the EU AI Act, and not every simple one is exempt. Article 6 and Annex III decide it — along with the deadlines the Digital Omnibus pushed back in mid-2026. A practical overview for SMEs — not legal advice.
ReadAre you allowed to run customer data through ChatGPT?
Under GDPR it's allowed under conditions: a lawful basis, data minimisation, and — when OpenAI acts as a processor — a data processing agreement. But a personal ChatGPT account is not the same as ChatGPT Business, Enterprise or the API.
Readllms.txt: the new robots.txt — what it is and whether you need one
llms.txt is a proposed file in your site root that tells AI systems in plain text what your site is and where things live. Origin, structure, and an honest answer to whether your site needs one.
ReadLocal models in 2026: Llama, Qwen and Gemma compared fairly
On-prem AI needs an open model that fits your own hardware. An honest snapshot as of mid-2026: licensing, sizes and context of Llama 4, Qwen 3.5/3.6 and Gemma 4 — and why "which one is best" is the wrong question.
ReadMeasure your AI visibility: is it you in the answer, or your competitor?
A ranking in Google no longer says anything about whether ChatGPT, Perplexity or Gemini mention your business. How to measure, yourself, whether you show up in AI answers — and what to do when the answer is no.
ReadModel routing: why not every request should go to the most expensive model
Not every AI task needs the biggest model. How to route on task complexity, latency, cost and reliability — with classifiers, rules, fallbacks and cost ceilings.
ReadMulti-agent orchestration: how do you get several AI agents to work together without chaos?
Adding more agents rarely fixes the problem you deployed them for — it just moves the problem to the coordination between them. On one shared playbook, fixed roles and one log as the single source of truth, with lessons from our own lab.
Readn8n or code: when is a workflow tool enough?
When is n8n enough, and when should you switch to code? Where workflow tools are strong, when custom code fits better, and why a hybrid architecture is often the strongest approach for AI automation.
ReadOllama, vLLM or llama.cpp: which runtime fits your situation?
llama.cpp is the C/C++ inference engine, Ollama the user-friendly layer around it for local experimentation, and vLLM the production-serving engine for GPUs with many concurrent users. Which one fits your workload.
ReadOn-prem AI: when is it worth the investment?
Cloud AI is easy — until your bill and your DPO start paying attention. An honest trade-off: when does AI run better on your own hardware, and when doesn’t it.
ReadPrompt injection: how a single email can hijack your agent
An AI agent that reads email and takes action can be manipulated by text inside that email — no password, no hack required. What indirect prompt injection is, and how least privilege and human-in-the-loop limit the damage.
ReadHow do you measure whether your RAG system is improving?
Without a fixed evaluation set, improving a RAG system is largely guesswork. How to measure retrieval precision and recall, answer quality, latency and cost — and how to prevent regressions when you change chunking, embeddings or prompts.
ReadRAG explained: how to build an AI that doesn't lie
RAG (retrieval-augmented generation) pairs a language model with a live search over your own documents, so every answer can be traced back to a source. How it works, where it breaks, and what it still takes to actually stop hallucination.
ReadFrom RAG demo to production: the five things that break
You can build a RAG demo in an afternoon. In production it breaks in five predictable places: retrieval, chunking, fabricated citations, the missing benchmark, and scale. What we ran into on 3.6 million court rulings.
ReadRate limits and quotas: how do you stop your service from grinding to a halt?
Rate limits and quotas can slow down or interrupt AI services without warning. Retries, queues, model routing, budget guards and monitoring together turn temporary scarcity into controlled degradation instead of an outage.
ReadSchema.org for AI: which structured data do assistants actually understand?
Schema.org tells a machine explicitly what a page is, instead of leaving it to guess. Which types actually pay off for AI assistants, how to tie them together with @id, and where it goes wrong in practice.
ReadWriting for Humans and Models: A Structure That Serves Both
Content that works well for people often works well for language models too — if you write with the same discipline. Why the answer should come first, when a list or table helps, and how to write content that stays easy to retrieve without sounding robotic.
ReadFrom a score of 36 to 85+: what actually raises an agent-readability score
An online store scored 36 on AI readability and climbed to 85+ without changing a single line in the CMS. What that score measures, which fixes made the difference, and why they shipped through an edge worker instead of the CMS.
ReadWhen Should an AI Assistant Say 'I Don't Know'?
An AI assistant that never refuses is predictably unreliable. How relevance thresholds, missing-context detection and citation verification let a RAG system say "I don't know" exactly when it should.
ReadWhen does a human need to approve? A decision tree for AI agent autonomy
Not every action an AI agent can take deserves the same amount of freedom. A practical framework based on reversibility, impact, scale and input trustworthiness — with a concrete decision tree for when a human should approve first.
ReadWhat is MCP (Model Context Protocol) and why is it changing agents?
MCP is the open protocol that lets AI agents connect to tools, data and systems in a standardized way. What it is, how it relates to APIs, function calling and RAG, and what it means for security and on-premises AI.
ReadWhat is reranking and when do you need it?
Reranking is the second stage in a RAG pipeline that reorders candidate documents by relevance. When does it actually add value, and when is it unnecessary complexity?
ReadWhat does AI really cost? From API token to your own server
APIs bill by the token, a server you own bills in purchase price, electricity and upkeep. What a million tokens actually costs at Anthropic and OpenAI, when owned hardware pays for itself, and the hidden line items that inflate the bill.
Read