Two different duties get mixed up in this question, and only one of them is yours.
The machine-readable marking duty in Article 50(2), embedding watermarks or provenance metadata in generated output, falls on the provider of the generative system: Midjourney, OpenAI, Stability, and so on. Using their tools doesn't make that duty yours. Your side of that duty is only negative: don't strip whatever provenance metadata the tool embedded when you crop, compress, or re-export the file, and keep originals.
Your own duty, as the company publishing the images, is the visible labelling rule in Article 50(4), and it applies only to the subset of AI images that are realistic: content depicting real-seeming persons, places, objects, or events that a viewer could take for an authentic photograph or recording. A photorealistic AI team photo, a realistic office scene on your About page, an AI headshot: label with the fixed-position "AI" icon at every exposure. Stylised blog-header art, abstract illustrations, obviously synthetic graphics: no labelling duty.
Deceptive intent doesn't matter; an innocuous decorative image that reads as a real photo is still in scope. Conversely, "made with AI" doesn't itself trigger anything: the Act has no general duty to disclose AI use in imagery that no one would mistake for reality.
Practical workflow for a small marketing team: sort your AI image inventory into "could be mistaken for a photo" and "obviously synthetic"; label the first pile per the Code of Practice convention; and record the sort decision in a dated note. That note is cheap, and it's exactly what turns an honest judgment call into defensible evidence if anyone ever asks.
Not sure how this applies to your specific setup? The free 2-minute check tells you which of the four Article 50 duty areas likely apply to your company. No email required.
Take the free check Generate disclosure text