Field notes
Long-form writing on the parts of this work that are actually difficult. Written when we have something to say rather than to a content calendar.
The Article 50 exemption is the specification
The Commission's Article 50 guidelines landed twelve days before enforcement. The exemption in paragraph 4 is the useful part: no disclosure needed where a human reviewed the content and a named party holds editorial responsibility. That is a specification, and it has to be evidenced.
Read the essayYour supplier is now your advisor
Anthropic has co-founded an enterprise AI services firm. The company advising you on which model to use is now part owned by one of the companies selling models. Not a scandal, a change in the terrain.
6 min read
The grace period is over.
EU AI Act enforcement starts on 2 August 2026, and the general-purpose obligations it penalises have applied since August 2025. Readiness is assessed against a record that already exists, or already does not.
7 min read
The market grew up. Did your AI?
Capital is concentrating, the frontier labs are heading for public markets, and enforcement has a date. A maturing market stops rewarding the loudest model and starts rewarding the one you can prove.
8 min read
Agent adoption outran agent security.
65% of organisations had a security incident caused by an AI agent this year. That is not a missing-tool problem. Autonomy, credentials and data reach were granted before anyone governed them.
7 min read
Your clients are not your fact-checkers
KPMG withdrew a report after 40 of its 45 citations turned out to be fabricated. The errors were caught by the organisations it named, in public, after publication. A control failure, not a model failure.
8 min read
Vibes don't ship. Systems do.
A vibe-driven engagement looks confident and produces deliverables that sound right. Six months later nobody can explain how the recommendation was reached.
7 min read
Vibes don't comply.
Article 50 is the smallest part of the EU AI Act. The real work is provenance, and it cannot be retrofitted the week before enforcement.
9 min read
Regulation is the head start.
Your governance posture is the most defensible thing on your cap table. Earlier, cheaper, and harder to copy than any feature.
7 min read
Nobody is winning enterprise AI yet.
Copilot ROI is disappointing, pilots stall, and the incumbents are trading credibility for velocity. The prize is still unclaimed.
8 min read
Everyone's an agent now.
Every marketing platform now sells agents. What a category-wide rebrand, quoted from the vendors' own pages, says about where the real work has moved.
9 min read
When the model layer goes public.
The labs are crossing multi-billion revenue lines and heading for the public markets. What listed model suppliers mean for everyone building on top.
8 min read
Governance is code, not policy.
Most governance lives in handbooks nobody reads. The only model that scales is enforcement in the architecture itself.
7 min read
Why most AI knowledge systems fail GDPR.
Most AI knowledge systems store everything they are given and can delete none of it. That is an architecture gap, not a policy gap.
9 min read
Your AI governance architecture is your Art.50 strategy.
The transparency deadline lands in August 2026. Organisations treating it as a labelling problem are solving the wrong problem.
8 min read
The side that loops fastest wins.
Boyd proved tempo beats capability in air combat. The same loop decides which organisations compound and which merely consume.
8 min read
We found the cheat code by getting it wrong.
A deployment ran for weeks with its scheduled quality checks absent. No error, no alert, nothing to debug. Recovery took hours because the orientation had been written down before it was needed.
6 min read
We migrated to ADO. Here is what actually changed.
Five days after making the work-tracking system the source of record: what a real cutover to disciplined delivery actually changes.
6 min read
Day zero.
Twenty-seven case-law entries from our first customer install. What shipping into a real tenant teaches that staging never will.
8 min read
If you're not the model, you're the harness.
The frontier labs ship models. Everyone else ships the harness: the scaffolding that decides what an agent remembers, touches, and must never do.
9 min read
The AI Typewriter Trap.
Using AI to do the same things faster is the typewriter trap: a revolutionary tool spent on volume instead of thought.
5 min read
Knowledge Equity, the economic footprint.
Organisations measure outputs but not the downstream value of what their people know. Knowledge equity makes that footprint measurable.
6 min read
Week 1 is the worst it will ever be.
Most software depreciates. A governed knowledge system compounds: week one is the worst it will ever be.
6 min read
We forgot our own engineers.
We built a platform to reduce cognitive load, then shipped an install that pushed the load onto our own engineers. Catching the contradiction.
6 min read
ORCA Glossary.
Every key term in the ORCA platform, defined: brains, confidence classification, knowledge equity, constitutional governance.
3 min read
- OpenAI
- Anthropic
- Google Gemini
- Meta
- Mistral AI
- xAI
- DeepSeek
- Cohere
- Qwen
- Ollama
- Hugging Face
- NVIDIA
the model layer we work across · routing, evaluation and fallback Marks are the property of their owners, shown to describe our own stack. No endorsement, sponsorship or partnership implied.
Tell us what you are accountable for.
The useful first conversation is usually about the thing that worries you, not about our capabilities. If it turns out we are the wrong firm for it, we will say so.
One reply from a person who has read what you sent.