THE AI MINDSET
Good morning, AI leaders.
Three things this week that most published material has not caught up with yet.
The EU AI Act was amended three weeks ago. The high-risk timetable moved by sixteen months and two new prohibitions appeared, so any policy written before the end of July has the wrong dates in it.
Closer to home, the advice boundary quietly gained a third position in April. Targeted support is now a regulated activity in its own right, which means any control still testing for advice or information has nowhere to put it.
And underneath both, a structural gap: firms approve AI systems, then connect them to everything, and the register still says one entry.
Everything below is sourced to the primary documents rather than to commentary about them. Where I read the actual text, I say so.
In today’s AI Mindset
The EU AI Act Moved Its Deadlines. Most Guidance Still Has the Old Ones.
The Advice Boundary Has Three Positions Now. Most Controls Only Know Two.
Your Firm Approved One AI System. It Connected to Seventeen.
LATEST DEVELOPMENTS
AI REGULATION
🌎 The EU AI Act moved its deadlines. Your policy probably still has the old one.
If your firm has an AI policy, a compliance matrix or a vendor assessment written before the end of July, the dates in it are out of date. The Act was amended three weeks ago and almost nothing published since has caught up
Category: AI Regulation | Reading time: 7 minutes:
What changed: Regulation (EU) 2026/1744 of 8 July 2026 was published in the Official Journal on 24 July and entered into force on 27 July. It amends the AI Act in 75 places, including the whole of Article 113, which sets the timetable. High-risk systems under Article 6(2) and Annex III now apply from 2 December 2027 rather than 2 August 2026. High-risk systems under Article 6(1) and Annex I move from 2 August 2027 to 2 August 2028. Two new prohibitions were inserted into Article 5, applying from 2 December 2026.
How it affects you: The Annex III shift is sixteen months. That is long enough to change a budget cycle, a supplier decision and a board paper, and long enough that a firm working from the old date will over-invest now and under-invest later. The smaller trap is the new Article 5 prohibitions. They are frequently being reported as if all Article 5 prohibitions start in December 2026. They do not. The original prohibitions have applied since 2 February 2025 and are unchanged. What is new are two specific additions, both narrower than the coverage suggests, because paragraphs 1a and 1b qualify when they bite.
What to do: Take the AI-related documents your firm relies on, and check the dates in each against the consolidated text. Anything citing 2 August 2026 for high-risk was written before 27 July. While you are there, check the article numbers: in the final Regulation post-market monitoring is Article 72 and serious incident reporting is Article 73.
Key takeaway: A regulation that has been amended is not the regulation you downloaded. The consolidated text is the easiest way to read the current wording, but it carries no independent legal effect, so the authentic instrument remains the version published in the Official Journal. If your governance material does not record which version it was checked against, and when, it will go stale without anyone noticing.
Every reference on our mapping matrix carries the source it was checked against and the date it was verified. It is free, and it is there so you can check us
AI REGULATION
The advice boundary has three positions now, not two...
Targeted support came into force on 6 April. If your firm's controls, training or file reviews still treat the boundary as information on one side and advice on the other, they are testing against a regime that no longer describes the rules.
Category: FCA | Reading time: 7 minutes:
What changed: The FCA published PS25/22 on 11 December 2025 with near-final rules for targeted support, and the FCA Board confirmed them as final on 26 February 2026. The Handbook instrument, FCA 2026/5, runs to 122 pages, amends sixteen Handbook modules and inserts a new chapter, COBS 9B. Most of it came into force on 6 April 2026. Targeted support is a new regulated activity under article 55A of the Regulated Activities Order. It lets a firm make suggestions designed for groups of consumers with common characteristics, without meeting the full requirements that apply to personal advice. It covers pensions and retail investments only, and a firm must apply to vary its permission before using it.
How it affects you: This is a rule change that makes something possible rather than adding an obligation, which is rare enough to be worth noticing. But it breaks things quietly. Any control built on a binary boundary will now flag legitimate targeted support from a permitted firm as a potential breach, because the control has no category for it. Any record-keeping template that asks whether an interaction was advice or information cannot record what actually happened. And any training material that teaches the two-position boundary is teaching a picture that is missing a third of the map.
What to do: Take one control that tests for the advice boundary, whether that is a file review checklist, a communications sign-off, or an automated flag. Ask what it would do with a compliant targeted support interaction. If the answer is that it would treat it as a breach, or has nowhere to put it, that is the first thing to fix.
Key takeaway: Rule changes that expand what firms may do are more disruptive to controls than rule changes that restrict it, because a restriction usually maps onto an existing test and a permission does not. The boundary is now information and guidance, then targeted support, then advice. Anything in your firm that knows only two of those three is out of date. And if you are weighing whether to apply, note the prudential consequence that has had little coverage: a permission to provide targeted support carries a £500,000 permanent minimum capital requirement.
Our free Advice Boundary Checklist covers the three-way boundary in eighteen questions across five sections, with the references checked and dated.
PRACTICAL AI GOVERNANCE
Your firm approved one AI system. It connected to seventeen.
Most AI registers list models. The governance question attaches to connections, and almost nobody is counting those.
Category: Practical AI Governance | Reading time: 7 minutes
What changed: Nothing, and that is the point. This is not a rule change. It is a structural gap that has been widening while everyone watched the rulebooks. When a firm approves an AI assistant, it approves a model. What it deploys is a set of connections: to a document store, a CRM, an email system, a payments provider, an external model, and whatever else a team wired in during a pilot that never formally ended. The register records one entry. The reality is a graph.
How it affects you: Three consequences follow, and none of them is theoretical. You cannot evidence oversight of connections nobody has catalogued, which is a problem the moment anyone asks what your AI can reach. Old integrations stay live after the thing they served is replaced, so credentials persist for systems nobody owns. And personal or regulated data can flow through a service that was never assessed, because the assessment was done on the model rather than on the path the data takes. The EU AI Act anticipates this: Article 72(2) requires post-market monitoring to include, where relevant, an analysis of the interaction with other AI systems.
What to do: Take your AI register and, for the single entry you are least comfortable with, list every system it can reach, who owns each connection, and when each was last reviewed. Most firms cannot complete that list from memory, and discovering that is the useful part.
Key takeaway: Governance that stops at the model governs the least interesting part of the system. The Mills Review makes the same point in regulatory terms: firms remain responsible for the services they provide, including where they use third-party models or agentic tools. The model is procured. The connections are accumulated. It is the accumulation that nobody is watching.
Still reading? Then you already suspect there are gaps worth finding. The SAFE™ AI Assessment gives you a plain-English read on your ten governance domains, and the one most likely to catch you out. Free, under ten minutes.
Until next week