Northern Ireland's EU AI Act Plan: Which Products Are in Scope and When

For a team selling an AI-enabled device, the way software is packaged and marketed can become a compliance question. New UK guidance makes that distinction particularly relevant to products placed on the Northern Ireland market.

The government published guidance on the partial application of the EU AI Act on November 4. It describes a limited arrangement under the Windsor Framework, rather than applying the whole Act in Northern Ireland.

The timing matters. A November 4 parliamentary statement says the UK and EU will take the relevant decisions at their November 10 Joint Committee meeting. As of November 5, that formal step is still ahead. The guidance explains the planned arrangement; it is not evidence that the decision has already happened.

The boundary follows the physical product

The official guidance PDF covers AI closely integrated with physical products placed on the Northern Ireland market. It generally excludes standalone web tools, apps and APIs.

Its integration test covers AI embedded or pre-installed in a product, or separately downloaded AI where the manufacturer designed the product to depend on that system for its intended functionality. A smart speaker with a built-in assistant is one example. Merely installing a separately marketed chatbot on a laptop does not put that chatbot in scope.

Direct general-purpose-model provider obligations are excluded under this arrangement, though product manufacturers may need information from model suppliers to meet their own duties.

For a developer, the first useful document is therefore a product map: what is sold, what functionality is promised, which AI component supplies it and who controls that component. That is more informative than starting with a model's name or a broad label such as “AI-powered.”

The guidance sets several dates

For in-scope products, its timetable lists:

• November 10, 2026: AI literacy, interaction disclosure and the initial prohibited-practice rules.
• December 2, 2026: AI-output marking and additional prohibitions on generating non-consensual intimate imagery and child sexual abuse material.
• December 2, 2027: Annex III high-risk obligations.
• August 2, 2028: Annex I high-risk regulated-product obligations.

These dates concern different requirements. Being within scope does not itself make a product high-risk.

This is a useful reason to separate scope assessment from implementation planning. A team can first establish whether the arrangement covers its product, then identify the applicable duties and their deadlines. Treating the last date as a blanket grace period would miss the earlier requirements.

What product teams can prepare now

Start with a short record for each device: its intended functions, AI dependencies, supply-chain owner and target markets. Compare the technical documentation with the promises made on the product page. If those accounts differ, resolve that mismatch before using either to support a compliance assessment.

Then identify what evidence has to come from suppliers. A model service agreement and a hardware compliance file answer different questions, so assign responsibility for connecting them rather than assuming one covers the other.

The immediate next check is the November 10 decision and any accompanying official documents. Product-specific obligations need qualified legal review, especially where software delivery, intended functionality or the supply-chain role is unclear.

The broader engineering lesson is straightforward: an AI component's place inside a product can matter as much as what the model does. Teams building devices need a record of that relationship, not just a list of models they use.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn technical and product changes into clear stories their customers can understand.

Book a 30-minute call.