Some EU AI Act obligations were deferred. Others are already in force. Does your board know which apply to you?
In July 2026 the EU amended the AI Act. The amendment was widely reported as a delay, and for one category of obligation it was.
Regulation (EU) 2026/1744 was published in the Official Journal on 24 July 2026 and entered into force on 27 July. It moved the obligations for high risk AI systems to 2 December 2027 and 2 August 2028. It did not move the transparency obligations, which applied from 2 August 2026. It did not reduce penalties. It added new prohibitions.
It also removed something boards should note. The original proposal made the new high risk dates conditional on the Commission confirming that standards and compliance tools were ready. The adopted text dropped that condition. The dates are now fixed.
Separately, in July 2026 the Monetary Authority of Singapore published SAFR, a framework for governing autonomous AI agents in financial services. It is not regulation and binds no one.
Most boards we speak to are unclear on which of this reaches them, and by when.
Midships offers a complimentary private briefing for boards and senior executives at large enterprises. The purpose is clarity: what applies, what does not, where exposure actually sits, and what management should do next.
No product presentation. No proposal at the end.
High risk, standalone
moved → 2 Dec 2027
High risk, embedded
moved → 2 Aug 2028
Transparency (Art. 50)
NOT moved — live 2 Aug 2026
Penalties
not reduced for large enterprises
New prohibitions
added
Readiness condition
dropped — dates now fixed
MAS SAFR (Jul 2026)
not regulation; binds no one
01
What applies, and when
Prohibited practices
Feb 2025
Banned in the EU since February 2025. These include social scoring, certain manipulative techniques, and emotion recognition in workplaces and educational settings. Highest penalty tier: up to €35 million or 7% of global annual turnover.
AI literacy
Feb 2025
Since February 2025, organisations in scope must ensure staff dealing with AI systems have a sufficient level of understanding.
General purpose AI model obligations
Aug 2025
In force since 2 August 2025 and untouched by the amendment.
Transparency
2 Aug 2026
From 2 August 2026, Article 50 introduces disclosure and marking obligations for specified AI systems and outputs. Depending on the use case and on whether the organisation is a provider or a deployer, these may include informing people that they are interacting with an AI system, machine readable marking of synthetic content, and disclosure of deepfakes and certain AI generated material. Systems already on the market before 2 August 2026 have until 2 December 2026 to meet the machine readable marking requirement.
Coming
New prohibitions
AI systems generating child sexual abuse material or non consensual intimate imagery.
High risk, standalone systems
This category captures much material use in financial services, insurance and employment, including creditworthiness assessment and risk pricing.
High risk, embedded systems
Penalties
€15m / 3%
Most other breaches, including high risk and transparency failures: up to €15 million or 3%.
€35m / 7%
Prohibited practices: up to €35 million or 7% of global annual turnover.
02
Does it reach us?
Scope cannot safely be inferred from where the organisation is headquartered.
Reach — not where you sit
The Act reaches organisations established outside the European Union in two principal circumstances: where they place an AI system on the EU market, and where the output produced by their AI system is used in the EU. An organisation headquartered in Asia with European customers, European operations, or products distributed into Europe may be in scope without having deployed anything in Europe.
Role — provider, deployer, or both
Obligations also differ by role. A provider develops or places a system on the market. A deployer uses one under its own authority. A large enterprise may act as both, particularly where it substantially modifies, rebrands or places a system on the market under its own name. Organisations that assume they are only deployers, and that their vendors carry the obligations, may find that assumption does not hold.
The practical implication is that scope cannot be settled from the top down. It requires an inventory.
03
What SAFR is, and whether you should care
Safeguards for Agentic Finance at Runtime was published on 3 July 2026 by the Monetary Authority of Singapore under its BuildFin.ai initiative, together with a group of financial institutions and FinTechs. The paper states explicitly that it does not constitute regulatory guidance or supervisory expectations.
If you are not deploying autonomous AI agents, SAFR is not your concern. It is worth understanding, not acting on.
The problem it addresses
If you are, it addresses a specific problem. Identity and access management answers who is acting and what they may reach. An agent can satisfy both, holding valid credentials and calling only approved systems, and still take an action inconsistent with the purpose it was given. SAFR proposes that a proposed action be evaluated against an explicit machine readable mandate before execution, resolving each to permit, deny, escalate or observe, with a record of what was proposed, what was decided and why.
Its significance
Its significance is not legal force. It provides financial institutions with an industry developed reference model for considering how autonomous agents could be governed while they operate. How widely it is adopted, including by auditors, counterparties or supervisors, is not yet clear.
Three questions for your next board meeting
Do we know where AI is being used, including inside products we bought for other reasons?
Inventories frequently miss AI capability embedded in existing vendor software, and AI tools adopted by staff without procurement. Both can carry obligations.
Can management demonstrate that the controls operate, not just that they exist?
Policies, committees and approval gates are necessary and insufficient. The question a regulator asks is whether the control worked at the moment the system acted.
If challenged on a specific decision, could we explain what happened, who was accountable, and why it was permitted?
Without records created at the time of the action, reconstruction is often possible but incomplete. Logs will usually show what happened. They frequently cannot show why it was allowed.
If any of these is unclear, that is a board level accountability gap rather than a technical one.
The briefing — complimentary · 60 minutes · private
A complimentary 60 minute private session for boards, audit and risk committees, executive committees, and senior legal, risk, technology and governance leaders at large enterprises.
You leave with:
Clarity
Which framework reaches your organisation and which does not, in plain terms, with the reasoning shown so your own advisors can check it.
Priorities
An initial view of where material exposure sits and what management should be able to demonstrate.
A plan
A 90 day action plan with suggested owners, outputs and sequencing. Yours to keep and execute with your existing teams and advisors. There is no requirement to engage Midships afterwards.
When this briefing is not worth your time
We would rather say so now than in the meeting.
No EU exposure, no agents
If you have no European market exposure, no European customers, no products reaching Europe, and no autonomous agents in or near production, the EU AI Act may have limited immediate relevance to you and SAFR is unlikely to warrant active assessment.
Already done the work
If you have already completed an AI inventory with risk classification, named an accountable executive, and can evidence your controls, you do not need this. You need your auditors.
Need a legal opinion
If you want a legal opinion on scope or classification, that is your counsel's work and not ours.
07
What we do, and what we do not
What we do not
Midships does not replace your legal counsel, auditors or regulatory advisors.
We do not provide: legal opinions, formal regulatory interpretations, risk classification, conformity assessments, or certification. Those belong with appropriately qualified legal, risk and audit specialists.
What we do
Our work is operational. We spend our time on identity and security infrastructure for large regulated enterprises. That gives us a specific and limited perspective: whether governance written on paper can be made to function in production, whether the controls can be operated, and whether the evidence they produce would survive inspection.
That is the layer between policy and production. It is narrower than what your advisors do, and it is the part we can speak to honestly.
A regulator may ask what your AI did. An auditor may ask who approved it. A customer may ask why it was allowed.
The board does not need to operate the controls. It does need to know the organisation can answer.
For boards and senior decision makers at large enterprises