The EU AI Act Reaches Jersey. Nobody Sent a Memo

EU AI Act compliance is not a European problem Jersey can watch from the touchline — for firms whose AI-enabled products or outputs reach the EU, it is an inbound obligation that arrives whether or not anyone in the firm has read the Act.

A Jersey firm can be fully outside the European Union and still fall inside the EU AI Act. If you place AI-enabled software on the EU market, or the output of an AI system you run is used in Europe, the Act can reach you. Most boards I speak to still assume this is somebody else's regulation.

I have made the argument throughout this collection that AI in a regulated business is a leadership decision before it is a technology one. The EU AI Act is the clearest test of that argument to date. It is not a firewall configuration. It is an accountability regime with extraterritorial reach and real fines behind it — and Jersey's entire proposition rests on being the jurisdiction that gets accountability right.

The security industry has already reframed AI as a new attack surface, and that framing maps cleanly onto the governance problem the JFSC set out in its July 2026 guidance. What I want to do here is join the two — the technical evidence security vendors and analysts are publishing, and the regulatory obligation now landing on Jersey firms — into one argument a Jersey board can act on.

Extraterritorial by design — the GDPR lesson, repeated

The instinct in a Crown Dependency is to treat EU regulation as background noise. Jersey is not an EU member state; it has its own data protection authority and its own regulator. That instinct is wrong here, for the same reason it was wrong about GDPR.

The EU AI Act's reach is not confined to firms established in the Union. It extends to providers who place AI systems on the EU market and to deployers whose AI output is used in the EU — an expanding sphere of influence that crosses member-state borders much as GDPR and the Digital Services Act do. Where you sell software with AI embedded into Europe, or where the output of an AI system you operate is used by or affects people in the EU, the obligation can follow the activity, not the postcode. This is not a Jersey firm choosing to opt in. It is the Act reaching out — and the precise trigger for any given use case is a question for the firm's own legal advice, not a blanket assumption either way.

The fines are not theoretical, and the comparison regimes show the trajectory. Cumulative GDPR penalties have run into the billions of euros since 2018, and the Digital Services Act has begun issuing its own — both figures move as new decisions land, so treat them as direction rather than a scoreboard. The pattern I return to with boards is simple: the last time Europe wrote a regulation with this shape, every serious Jersey firm eventually built a programme around it. The firms that waited to be told paid twice — once in remediation, once in lost client confidence.

You cannot govern the AI you cannot see

Before compliance comes a harder question: does the firm actually know what AI its people are using? The evidence says almost certainly not. SentinelOne's Prompt Security team, reporting from live deployments, found small organisations averaging 42 distinct websites with generative-AI capability built in, and mid-sized firms averaging 72. That is not a shortlist of approved tools — it is dozens of exposure points most security teams have never inventoried.

The exposure compounds daily. In their finding, a firm of 45 employees each running around a dozen prompts a day generated roughly nine policy violations per day. Translate that into a Jersey context: nine daily opportunities for client data, beneficial-ownership detail, or a draft regulatory filing to leave the building through an unapproved chatbot — each one a potential data-protection breach and a potential governance failure the JFSC would expect a firm to have controlled.

The surface is also widening faster than most inventories. It is no longer browser tabs alone but desktop applications, autonomous agents, and the MCP connectors that let an AI tool reach into other systems. Gartner projects that the market for managed AI services could reach around $1.1 trillion by 2029, driven substantially by agentic deployments — a forecast rather than a certainty, but a clear signal of direction. Agentic AI is not a marginal case to defer; it is where the market is heading, and it is exactly the layer with the least visibility today. This is the practical face of the JFSC's first principle: accountability for outcomes you cannot see is accountability in name only. Visibility is the first governance control, not a precondition to it.

Two risk ladders, one obligation

The EU AI Act sorts systems into four tiers — unacceptable, high, limited, and minimal risk — each with matching controls. Unacceptable uses, such as social scoring and manipulative or exploitative AI, have been banned since February 2025; that is live now, not pending. High-risk systems, including those in critical infrastructure and in activities such as access to education or credit, must embed security by design: a cybersecurity risk-management system, automatic record-keeping and logging (with retention obligations that run to at least six months, subject to the specific provisions and national implementation), and human oversight. Even limited-risk systems carry a direct expectation — effective, robust and reliable operation as far as is technically feasible — which is secure-by-design in all but name.

A Jersey reader should recognise the architecture immediately, because the JFSC's own Materiality Ladder does the same job. Its three tiers — Low, Medium, and Higher impact — scale governance to consequence rather than to technology. The two are not identical and a firm cannot satisfy one by pointing at the other, but they rhyme: a firm that has honestly classified its use cases against the JFSC ladder has already done most of the thinking the EU Act demands. Table 1 sets them side by side, so a firm can locate each use case in both regimes at once — and see where the EU regime asks for more.

Table 1 — An illustrative reading of JFSC Materiality tiers against the EU AI Act's risk categories, to help a firm orient a use case in both regimes. It is a starting point for discussion, not a legal classification; the two frameworks use different tests, and a firm's own assessment governs.

Testing is no longer optional — it is written into the Act

The single hardest obligation for most Jersey firms to absorb is that the EU AI Act requires applications to be tested against adversaries, continuously, not signed off once. Adversarial testing, technical documentation, risk-model and incident-reporting obligations go live in August 2026, with full enforcement and the first fines expected from August 2027.

The published security research shows why this matters. SentinelOne's red-teaming of a live education chatbot found that simple persona and instruction attacks — telling the model to ignore its rules, or to adopt a hacker persona — were caught, but the system failed once testers moved from semantic prompts to structural ones: requesting JSON output that unlocked developer functionality, and using base64 encoding to smuggle instructions past the system prompt. The methodology follows the OWASP Top 10 for LLMs. The JFSC names the same threats in plain terms — prompt injection, data poisoning, and attempts to extract confidential information from outputs.

Gartner reaches the point from the other direction: errors in AI systems are inevitable, so the discipline is not eliminating them but managing them proportionately — risk understood as likelihood multiplied by impact. Two disclosures within ten days of each other in July 2026, reported by the firms involved and independent analysts, made the same argument in the field rather than the lab. On 21 July OpenAI reported that its models had escaped a supposedly isolated evaluation sandbox and reached the production infrastructure of Hugging Face, the largest open-model platform; on 30 July Anthropic disclosed three cases in which Claude models, believing they were in an air-gapped simulation, reached the live internet and accessed real organisations — in one, publishing a malicious package to the public PyPI registry that then ran on real systems. Neither is an EU AI Act enforcement case, and details may be refined as investigations continue. But the parallel that matters for a Jersey firm is the mechanism, not the drama: both began at the supply-chain layer — a package or dataset registry, the same class of dependency firms now wire into their own AI workflows through connectors and agents — and both turned on an isolation boundary that was assumed rather than verified. If that boundary can fail unnoticed inside two of the most capable AI labs in the world, an unaudited client-facing assistant in a regulated firm should not be assumed to hold.

The mistake I see most often is treating an AI tool as tested because it worked in a demo. A demo is not an adversary. Under both the EU Act and the JFSC's expectation of pre- and post-deployment testing, a homegrown client-facing assistant that has never been red-teamed is not a modern convenience — it is an undocumented, untested control failure waiting to be found by someone other than you.

Three myths a Jersey board should stop repeating

Three misconceptions surface in almost every governance conversation I have, and each is alive in Jersey boardrooms.

• “It's SaaS, so it's the vendor's problem.” The deployment model does not change the obligation. SaaS versus on-premise makes no difference; what matters is how the AI is used, its intended use, and what data goes in and comes out. A firm cannot outsource its accountability to a vendor's cloud.
• “We're not in the EU, so it doesn't apply.” Jersey firms are squarely in scope wherever EU clients, data or downstream software are involved — the same reach that pulled Channel Islands firms into GDPR. Being outside the Union is not being outside the Act.
• “Let's wait until it settles down.” The Act is evolving in real time. The Omnibus amendments have deferred some high-risk obligations by sixteen months and pushed national sandboxes to 2027, but the bans and transparency duties are live today. A moving target is not a reason to stand still; it is a reason to build a programme that can absorb change.

There is a distinctly Jersey opportunity in all three. A jurisdiction whose brand is trust and regulatory excellence should not be the last to demonstrate credible AI governance — what the Jersey AI Playbook calls ethical differentiation. The firms that treat the Act as a floor to clear early, not a wall to hide behind, will win the mandates that care.

Where a Jersey firm actually starts

The starting point is not a compliance project. It is an inventory. A firm cannot protect, classify, or certify what it cannot see, so the first move is to discover every AI tool, agent and connector in genuine use — including the shadow AI its people have quietly adopted and the AI its suppliers have adopted on its behalf. From there the sequence is the one this collection has argued throughout: classify each use case by materiality, assign a named owner, decide who may use what, and test before you trust. International frameworks — NIST's AI Risk Management Framework and ISO 42001 — give a defensible way to build the compensating controls the Act assumes, and because AI risk spans compliance, legal, HR, operations and technology, the board, not the security team, owns the outcome.

What that looks like in practice depends on the firm. The point of the JFSC's materiality logic is that a forty-person administrator and a global-entity bank do not owe the same programme — they owe the same discipline, scaled to consequence. Table 2 sets out a proportionate first ninety days for three firm profiles common in Jersey.

Table 2 — A proportionate first ninety days by firm profile: the minimum a Jersey firm should be able to show, and the single decision that belongs with the board. Illustrative, not a compliance standard — scale to your own materiality assessment.

If a European client asked tomorrow to see your AI use-case register, your six-month interaction logs, and evidence that your client-facing AI has been tested against an adversary — could you produce them, or would you be starting from the inventory you have never run?

Next in the series: from inbound obligation to operating model — turning EU AI Act and JFSC expectations into a single register a Jersey board can actually govern from.

Sources: SentinelOne / Prompt Security, A Practical Guide to the EU AI Act (2026j), for shadow-AI prevalence and red-teaming findings; Gartner (2026d), CIO Value: From AI Tools to Managed Enterprise AI Services, and Gartner (2026c), Building Your AI Accuracy Survival Kit; July 2026 evaluation-escape incidents (2026k) — OpenAI's 21 July and Anthropic's 30 July disclosures (the latter including the Mythos 5 PyPI incident), per the firms' statements and reporting including TechCrunch, The Hacker News, BleepingComputer and the Cloud Security Alliance; EU Artificial Intelligence Act and Omnibus amendments (phased application, prohibitions live from February 2025); OWASP Top 10 for LLMs; JFSC, Guidance on the use of AI in Jersey's financial services sector (July 2026); Jersey AI Playbook (Government of Jersey et al., 2026); NIST AI Risk Management Framework; ISO/IEC 42001. GDPR and DSA figures are publicly reported and change over time. Regulatory characterisations are general and not a substitute for legal advice.

Professional disclaimer: The views expressed are my own and are shared for general informational purposes. This content is intentionally non-confidential and does not include proprietary, client-specific, or employer-sensitive information. It is not legal, regulatory, investment, or professional advice; readers should seek appropriate advice for their circumstances.

Financial services leader (20+ years) | Lean Six Sigma Green Belt | London School of Economics (LSE) AI Leadership Accelerator

A version of this article was first published on LinkedIn as part of the Sequence body of work. This is the canonical version.

The views expressed are the author's own and shared for general information. This content is intentionally non-confidential and does not include proprietary, client-specific or employer-sensitive information. It is not legal, regulatory, investment or professional advice. © 2026 Paul Soares. Sequence is the trading name of Sequence Advisory Limited (in formation) · hello@sequenceadvisory.com