Journal4 min

What regulators asked for in October

No new bans. The UK FCA, the Financial Stability Board and the EU all asked for the same thing in different words: evidence, control over model changes, and a way out.

An empty boardroom with a long table, light falling through tall windows

Three publications in October will shape how banks, insurers and public bodies buy AI over the next two years. None adds a new ban. All three ask for evidence and for control over what changes.

The UK FCA: hosting, notice and exit

On 19 October the Financial Conduct Authority set out good practice and areas for improvement for AI in financial services. Firms should make considered decisions on whether to host their AI models on architecture they own. Some use contract terms to get notice of significant model changes from their providers. Weak spots include over-reliance on hard-to-replace providers and no way out of a single platform. And firms can outsource AI services, but remain responsible for them, including for drift.

The FSB: twelve sound practices

On 15 October the Financial Stability Board finalised twelve non-binding practices for adopting AI. Eight follow the life of a model: data governance, explainability, performance evaluation, human oversight and the management of risks from third-party AI dependencies.

The EU: dates moved, duties did not

The Digital Omnibus on AI, in force since 27 July 2026, moved the main high-risk dates of the AI Act. Stand-alone high-risk systems, such as hiring or creditworthiness, now apply from 2 December 2027; AI inside regulated products, such as medical devices, from 2 August 2028. The next dates are closer: on 2 December 2026 the grace period ends for marking AI-generated content from systems already on the market.

In US health care, the duty moves to the buyer

On 26 October the US health IT office finalised HTI-5. Certified health records no longer have to show users how an AI tool was trained, validated and tested. A hospital that wants to know will have to ask the vendor, or run a model whose training and tests it holds itself.

What it adds up to

Read together, the asks are the same: know which model is running, know when it changes, test it on your own work, keep a record, and keep a way out. Those are the parts an owned model is built from: a version that stays the same until you decide, a release gate with a sealed test, a release record and a Model Deed.

None of this settles compliance on its own, and we never say it does. It is designed to support your compliance work: the evidence your teams and your supervisors ask for, kept on your side.

Notes

  1. FCA, AI in financial services: good practice and areas for improvement, 19 October 2026, updated 20 October 2026.
  2. FSB, sound practices for the responsible adoption of AI, 15 October 2026.
  3. Regulation (EU) 2026/1744, as consolidated by the AI Act Explorer; summary by Cobalt. The official journal page did not load when we checked.
  4. ASTP/ONC, HTI-5 final rule, 26 October 2026; most changes apply 60 days after publication in the Federal Register.

Read next

Newsletter

Release notes.

Once a month: what we released, what we held, and why. Short, dated and true, losses included.

Double opt-in. One click to leave. We never share the list.

Each issue holds

  1. One release or testWhat passed the gate, what was held, and the reason in one line.
  2. One Journal pieceAn explainer from the people who build it.
  3. One numberWith its baseline and its date, never on its own.

The first issue comes in November 2026.

Start

Bring one task. See the proof first.

In about four weeks, on your own data in a sealed setup, we show whether an Eigna expert beats your baseline. If it does not, you keep the report.