Update, 12 September 2026: this checklist was written the week before 2 August 2026. It now works as a review: the Article 50 obligations it goes through have applied since that date, except the marking in paragraph 2 for systems placed on the market before it, whose providers have until 2 December 2026 (Article 111(4)).
If your company uses a customer-facing chatbot or publishes AI-generated content, this list is for you: the Article 50 transparency obligations come into application on 2 August 2026, except the marking under paragraph 2 for systems placed on the market before that date, whose providers have until 2 December 2026 (Article 111(4)). We've covered it calmly over the past few weeks: what it requires, who it requires what from, with which tools. This piece is different: it's the closing checklist. Five tasks that an SME can complete in the week that's left, none of them heroic, all of them demonstrable. If you've already done the work, this is your final review; if you're behind, it's your catch-up plan.
1. The short list of affected systems
You don't need to re-audit your entire AI estate this week; you need the Article 50 sub-list. From your inventory, pull out the systems that: (a) interact with people — chatbots, voice assistants, conversational schedulers —, (b) generate or modify content — image, video, audio, publishable text —, or (c) recognise emotions or categorise people biometrically: in that last case the duty to inform the people exposed is yours (Article 50(3)), and in the workplace inferring emotions is prohibited except for medical or safety reasons (Article 5(1)(f)). That's your 2 August surface. Write them down; spend the rest of the week working only on those.
2. The chatbot: verify, don't build
For every customer-facing conversational system, the deployer's task is verification: check in the actual conversation — not in the sales sheet — that the system discloses it's an AI — Article 50(1) only drops the requirement where this is obvious from the point of view of a natural person who is "reasonably well-informed, observant and circumspect" —, confirm that none of your customisations hide that disclosure, and log the check (system, date, who, result). If the disclosure isn't there, this week's call is to the provider: the design obligation is theirs; the interest in getting it fixed before the 2nd is yours.
3. Content: a decision, not a technology
If you publish content generated or retouched with AI — creatives, videos, product images —, what Article 50(4) requires is disclosing that it is artificial when the piece is a deep fake: image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to be authentic; where it forms part of an evidently artistic, creative, satirical or fictional work, disclosing its existence in a way that does not hamper the display or enjoyment of the work is enough. For the rest, the task is to make and write down a decision: which types of piece you label, with what signal (a recognisable marking, in line with the code of practice under Article 50(7), guides the "how", although adhering to the code does not grant a presumption of conformity) and where it's placed so it's visible on first contact. A one-page internal policy is enough. What isn't enough is the vague intention of "we'll label it eventually".
4. Text: activate the editorial route
For written content you publish with AI support, Article 50(4) only requires disclosure for text published with the purpose of informing the public on matters of public interest, and the duty falls away where the text has undergone human review or editorial control and a natural or legal person holds editorial responsibility for its publication. That's why the reasonable coverage for an SME is the editorial review route: human review of what's published and editorial responsibility taken on. Activating it this week costs exactly this: a one-page policy describing your publication workflow, and a per-piece log (who reviewed it, when). If your blog already works this way, the job is just to write it down; if it doesn't, this is the moment to make it work this way.
5. The record that holds it all together
Each of the four tasks above ends the same way: with a row in a log. The systems sub-list, the chatbot check, the labelling criteria, the editorial workflow. Keep it alongside the rest of your compliance evidence, because that's what it is. From 2 August onwards, any question — from an authority, a customer, a partner — gets answered by showing this. The difference between having done it and being able to demonstrate it is, once again, everything.
What you do NOT have to do this week
Equally important, so you don't waste the week on what doesn't apply: you do not have to sort out high risk (it has a different timeline: Regulation (EU) 2026/1744 deferred its obligations to 2 December 2027 for Annex III systems and to 2 August 2028 for Annex I systems); you do not have to sign the code of practice on transparency to be in order — it's voluntary and, according to recital 41 of Regulation (EU) 2026/1744, has limited legal effect and does not grant a presumption of conformity, so aligning your practices without signing is a perfectly reasonable position; and you do not have to rush into hiring anything from anyone selling last-minute panic. The five tasks above are within reach of any organisation with a week and some judgement.
On Monday 3 August, the difference between companies won't be who ran fastest on Friday. It will be who can open a folder and show five records dated July.
Content in accordance with Articles 3, 5, 50, 111 and 113 of Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744 (Official Journal of the EU, 24 July 2026), and with recital 41 of the latter.
This article is for informational purposes only and does not constitute legal advice.