Insights

Notes from the work. Written down so the next decision is faster.

Build notes from the products we run, how we put AI systems into production, and the reasoning behind design decisions. An outside claim carries its source; every note carries its date.

Topics: Build notesAI systemsGovernance and securityDesignField notes

Start here

Governance and security15 min read

The EU AI Act for companies outside the EU

The AI Act reaches a company with no office in the EU in two ways: placing an AI system or model on the EU market, or output used in the Union (Article 2). Your role, what applies when since the Digital Omnibus of July 2026, what a model vendor owes you, the fines, and a decision tree for a SaaS company or an agency.

Read the note

All notes

Ten notes, newest first

  1. Governance and security24 min read

    Under the AI Act, human oversight is a capability, not a duty

    EU AI Act Article 14 says what a high-risk system must let a reviewer do, and obliges nobody to intervene. The instrument that reaches a review which decides nothing is data protection law, and on the best evidence we found the sign of the result tracks which is the better judge, the person or the model.

  2. Governance and security22 min read

    What already binds your AI build, and what is only dated

    Most of what binds a company shipping AI today is not a statute but the terms of the model it builds on. Which rules reach a company abroad, which are only dated, and which are not law at all — read against the instruments on 3 and 4 November 2026.

  3. AI systems15 min read

    When should an AI agent act, abstain or escalate?

    Model capability is jagged: strong on one task, weak on its neighbour, in ways a benchmark score cannot locate. What is measured about confidence, self-checking and abstention — and the signals, thresholds and handovers to build instead.

  4. Governance and security15 min read

    The EU AI Act for companies outside the EU

    The AI Act reaches a company with no office in the EU in two ways: placing an AI system or model on the EU market, or output used in the Union (Article 2). Your role, what applies when since the Digital Omnibus of July 2026, what a model vendor owes you, the fines, and a decision tree for a SaaS company or an agency.

  5. Build notes10 min read

    A black rectangle is not a redaction

    Drawing a box over text leaves the text in the file. What a PDF still carries after the box, why Zero-Trust PDF Studio rebuilds a marked page as an image, and how it checks its own export before you send it.

  6. Field notes9 min read

    AI-generated software is a prototype until someone owns the failure modes

    Generating a working application has become cheap. What still costs money is knowing how it fails, who is told when it does, and what happens next: the work between a convincing demo and a system a business can run on.

  7. AI systems9 min read

    MCP gives an AI system tools. Permissions decide what those tools may do.

    The protocol hands a model a list of functions it can call. Whether a call should happen is a separate decision, and the specification says so itself. How we think that line should be drawn before an agent goes into production.

  8. Governance and security10 min read

    What an AI agent should log: the tool call, not the conversation

    A log that can say why an agent acted, who approved it and what it had read is an engineering artefact. The fields we think belong in each record, who may write it, what is redacted and how long it is kept, each with its source.

  9. Design10 min read

    When software does not know, it has to say which kind of not knowing it is

    A blank, a zero, a confident sentence and a spinner each hide a different gap. Seven kinds of not knowing an interface has to tell apart, what a reader needs from each, and the honest form, shown in products we run.

  10. Build notes10 min read

    When supplier quotes cannot honestly be compared

    Two quotations for the same material can differ in unit, tax basis, delivery point and version, and a ranking that ignores any one of them is a guess. The rules we gave Rateslip for refusing to rank, and what it does instead.

Reference

26 tables, lists and examples in the notes, each at its own address