The series · 10 parts · ~90 min end to end
AI is arriving inside the software you already pay for
Not as a purchase you evaluated. As a feature added to a contract you signed for something else, from a supplier you cannot name, on terms you were never shown. This is a blueprint for shared AI infrastructure, owned by the people who use it — built by groups that already have something in common: a profession, a district, a sector, a shared purpose. Each paper can be read alone. Free to copy, adapt and build from.
For the whole series: All 10 decks · All 10 PDFs
What is actually happening, and why it is not simply that software costs more. The mechanism is wider than price, and what it costs you in the end is not money. Two papers: what is documented, and six questions you can put to your own organisation this week.
Your renewal quote went up. The plan you were on is being retired. Moving would mean re-entering four years of records. So you sign — and everyone signs, because signing is the rational decision. What is documented about how that position was reached, and what was actually lost when it was.
Read the essay · Questions · Glossary · Sources · Slides · PDF
Six questions about any AI arrangement — starting with whether you can even find out what you are already using. Apply them to your own organisation, or to a supplier's claims. New Zealand currently fails four of the six, and each finding says whether it is evidenced or whether we searched and did not find.
Read the essay · Questions · Glossary · Sources · Slides · PDF
Shared AI infrastructure, and what it takes to build some. Not a better vendor — the same relationship at a better price is still the same relationship. Something the organisations using it own a share of: services suited to their own work, capacity that earns rather than sits idle, redundancy that actually holds, and terms they set themselves. Three papers: what it is, what it does for you, and what it costs.
Twelve things it has to be able to do — each checkable by watching it work, not by reading a description. Eight decisions that are yours and cannot be made for you. How to start with one region, and the four ways this fails, all of them documented. The legal form is a co-operative; this is where that is set out and why.
Read the essay · Questions · Glossary · Sources · Slides · PDF
A three-partner practice hits a question nobody has dealt with since 2021. What happens next: answered by a system that has read their own files, escalated to a person only when it is unsure, and the answer kept so next time is easier. The service described as itself.
Read the essay · Sources · Slides · PDF
We cannot tell you, and nor can anyone who has not built one. What we can give you is the structure for working it out: four cost pools, the one number everything ladders to, and worked figures that are illustrative placeholders, not estimates.
Read the essay · Sources · Slides · PDF
How you run an organisation when most of the work is done by software that acts on its own. You are still accountable for all of it, and conventional management assumes that finding out what happened is expensive. Two papers on what replaces that — and on what actually stops an AI doing the wrong thing, which is not what most people think.
Read this one even if the rest is not for you. Who is responsible for what an agent did at three in the afternoon, when someone asks? How do you keep your own claims straight, stop your diagrams lying, and decide what to do deliberately badly so the whole thing survives a shock? Includes where our own method failed, because one that reports only successes is not reporting.
Read the essay · Sources · Slides · PDF
Most AI safety is instructions given to the agent, and instructions are requests. One test sorts every safeguard into two piles: if the agent were entirely mistaken about what it was doing, would this still hold? Almost everything on the market fails it, and the reason is not negligence.
Read the essay · Slides · PDF
What a government could change, most of it costing nothing. Not a new AI Act — amendments to instruments already in force, one small piece of infrastructure that nobody owns, and an argument about why most technology policy changes nothing at all.
New Zealand's Privacy Act has a complete enforcement chain — investigation, notice, tribunal, order, offence — ending in a fine of up to ten thousand dollars. It does not fail because it is badly written. Six questions to walk any proposed obligation through, and the three roles a government actually has.
Read the essay · Slides · PDF
Five of them cost nothing. One is a single undefined phrase already sitting in government cloud policy which, defined, would become the strongest sovereignty provision in New Zealand procurement. Written for an official with a procurement budget and no appetite for a new programme.
Read the essay · Slides · PDF
Everyone is planning to migrate their cryptography. Almost nobody is dealing with the records already signed — which do not become insecure, they become forgeable, backwards. And they cannot be re-signed, because the people who signed them may no longer exist.
Read the essay · Slides · PDF
Drafted with AI assistance, then checked and revised by the author. The judgements and the errors are the author’s — declaration.
Version 0.1, September 2026 · licensed CC BY 4.0 — reuse is free, including commercially, provided the author is credited and the reuse does not suggest that the author endorses it. agenticgovernance.digital · all papers