Decide Who Decides
-
- $14.99
-
- $14.99
Publisher Description
Your AI is already making decisions nobody approved. Sooner or later, your board will ask who let it.
Agents now act on your systems, some under a person's credentials, without anyone approving each action first. Most approval gates were never written for an agent. And some of what you call a control is only a sentence in a prompt.
I watched a quieter version of this inside a large organisation. AI triaged 200,000 security findings over a long weekend, work that used to take months of sitting with the people who ran those systems. It made 200,000 judgements about a business it knew nothing about. It flagged as critical risks the owners had already weighed and accepted, and the results went onto a dashboard before any of them were asked. No one decided to cut them out. The slowness had been the control, and nobody noticed until it was gone.
Decide Who Decides is about taking those decisions back, on purpose: which ones a machine may close on its own, which a person must approve first, and which a person must be able to undo afterwards. Get that right and you can move fast and still stand behind what happened.
It is written for the security leader, and to be read by the CTO shipping agents, the CEO who owns the risk and the board that will ask, so all of them work from one answer to "who decided?"
By the end, you will know how to answer three questions, with evidence:
•Which AI systems are operating here, including the unsanctioned ones?
•Which decisions have moved from a person to a machine, and did anyone agree to that?
•If one of them caused harm tomorrow, who is accountable, by name?
Along the way, you will learn:
•How to give every agent its own identity and an expiry date
•Why, when a model can answer the same question differently tomorrow, the control has to be the record
•How to give a board three priorities instead of eighteen, in terms a CFO can act on
•Why cutting junior analysts to pay for AI tools deletes the senior people you will need later
•Which of ISO 42001, NIST's AI RMF, the EU AI Act and, in Australia, APRA's CPS 230 and 234 actually bind you
Six tools at the back, ready to use this week: a 20-question governance debt assessment, a ledger for the governance shortcuts you are carrying, a first-90-days plan, a board conversation guide with the five questions directors should be asking you, a three-tier model for splitting work between people and AI, and four checks before any AI workflow goes live.
This is a book about judgement, not a technical manual or a compliance guide. Every statistic is sourced, the notes say who funded each study, and any number I could not stand behind was cut.
About me: I have spent more than twenty years in technology, most of it in security. I am accountable, end to end, for the security of a large regulated organisation's products, and have served as virtual CISO to several others at once. I once stood in front of a board, was asked "Are we good enough?", and had nothing to say. I built and shipped my own AI compliance product, fine-tuning the models and building close to 10,000 control mappings across six frameworks, then watched it cite a control that did not exist, with complete confidence.
Somebody has to decide who decides. In most organisations, nobody has been asked to.