Architecture decisions, governance gaps, and what really goes wrong when enterprises deploy AI - from real projects.
PII masking is supposed to make enterprise data safe for LLMs - and fails on a detector with unprovable recall, on utility errors that pull against the safety axis, and on a stateful translation layer in the hot path. Why establishing the legal basis is the shorter route, and what remains of the building block.
Business cases for AI automation weigh savings against running costs - and leave one real position unbooked: the loss of the fallback option. Why substitution isn't symmetric, which five paths lead to the same outage, why the ability to train people is the more uncomfortable case, and four measures that make the position quantifiable.
An LLM sounds assured, phrases things politely, and still carries no responsibility whatsoever. Why trust signals run into the void with language models - and the four architectural consequences that follow: deterministic validation, a sandbox with defined rights, task-scoped access, and Dissemination Control.
The debate around context windows is all about tokens, cost, and summarization - and it misses the question that actually comes first: which information gets to reach which agent? Why the context window isn't storage but an egress point, and how memories get policy-filtered with four attributes and a deterministic layer instead of dumped.
Stack Overflow for Agents ships one of the most careful skill.md files out there - and that's exactly why it exposes a structural problem in the whole skill economy: mutable remote instruction sources without integrity. A close read of 692 lines and why it takes technical controls, not prose.
The usual AI architecture puts the frontier model at the center and a local one as a stopgap behind it. This piece flips the dependency around: the local model becomes the foundation, the frontier model the teacher that makes it better through distillation and works itself out of the critical path by design - for a risk profile a regulated company can actually take responsibility for.
Why an AI that knows nothing about your business is the safer choice: instead of training knowledge into the model, the agent receives only the data of the one case it's handling per request. The real work isn't in data access - it's in controlling what the agent is allowed to pass on: Dissemination Control.
Chaotic free-text fields are nearly impossible to clean with conventional means. How a local LLM (Ollama/Qwen) recognises context, a regex reverse-check catches hallucinations, and the data stays GDPR-compliant without ever leaving the company network.
Authentication and authorisation aren't enough for AI agents with tool access. This post explains the missing third layer - Dissemination Control - and a four-tier architecture to implement it incrementally.
An AI agent wiped an entire production infrastructure. The industry responded by stripping agents of access - and misses the point entirely. The real fix is architectural.
A practical evaluation of automated PII anonymization for LLM pipelines: spaCy vs. Flair (via Microsoft Presidio) on German business texts, including the limits of contextual personal references and a recommended approach (regex + NER + documented residual risk).
AI agents are not autonomous pipelines - they are team members with defined roles, permissions, and boundaries. A practical report on dissemination control, persona agents from real customer data, and the architecture that makes collaboration between humans and AI secure.
AI assistants can bypass permission boundaries by mixing information from different project contexts. This article shows through a real scenario why context separation matters more than the choice between MCP and API - and which three architecture approaches solve the problem.
RAG on enterprise data rarely fails because of the model - it fails because of access control. An architecture guide on chunk-level permissions with OPA, GDPR masking, and the structural conflict between data protection and AI usage.
Why most risks of using AI are scaled versions of familiar enterprise problems - and why that's good news.