Illustrative scenarios · EU AI Act engineering

    How an Engagement Would Work

    Please read this first

    These are worked examples, not client case studies. The companies are invented, and nothing here describes real results. I wrote them to show how I would approach common EU AI Act engineering problems. For real work I've done, see the in-house chatbot case study. Whether a duty applies to your system is a question for your legal advisers.

    Each scenario follows the same shape: the situation, the steps I'd take, what you'd get and one thing to watch for.

    A support chatbot that never says it's an AI

    Illustrative scenario, not a client engagement

    Company: An invented 40-person SaaS company in Berlin with a Claude-based support chat on its website and in its app

    Relevant duties: Article 50(1): people must be told they are interacting with an AI system

    The situation

    The chat widget greets customers with a friendly name and a photo. Nothing tells them it's an AI system.

    The team also has no record of what each customer was shown. If a regulator or a customer asks, they can't prove a disclosure ever appeared.

    What I would do

    1. Step 1: Map every entry point. I list each place a person can reach the AI: the web widget, the in-app chat, email replies drafted by the model and the handover to a human agent.
    2. Step 2: Put the disclosure in the product. The disclosure appears at the start of each conversation, in the same language as the conversation, before the first AI reply. It's part of the code path, so a design change can't drop it silently.
    3. Step 3: Log that it was shown. Each conversation stores a disclosure event: the text version, the channel, the language and a timestamp. The log is append-only.
    4. Step 4: Test it. Automated tests fail the build if any entry point can return an AI reply without a disclosure event first.

    What you would get

    • Disclosure built into every entry point, with tests that keep it there
    • An append-only disclosure log the team can query
    • A one-page note for the legal team on what was built and where

    Watch for: Human handover. When a person takes over, the customer should be able to tell. When the conversation goes back to the AI, the disclosure should come back too.

    Marketing images that leave the building unmarked

    Illustrative scenario, not a client engagement

    Company: An invented e-commerce retailer that generates product images and short videos with a third-party model API

    Relevant duties: Article 50(2): machine-readable marking of synthetic content. Article 50(4): disclosure of deepfakes by deployers

    The situation

    The team generates hundreds of images a week. They are resized, compressed and uploaded to the shop and social channels.

    Whatever marking the model provider adds is often stripped by the time the image goes live. Some videos show realistic people who don't exist.

    What I would do

    1. Step 1: Trace the content pipeline. I follow one image from generation to publication and note every step that rewrites the file: editing, resizing, compression and upload.
    2. Step 2: Mark at generation time. The marking is applied as soon as the content is generated, using provenance metadata plus a watermark where the format supports it, so one stripped layer doesn't remove all of it.
    3. Step 3: Check before publishing. A publishing gate checks the marking is still present and blocks the upload if it isn't.
    4. Step 4: Add the deepfake disclosure. Content showing realistic people gets a visible label in the publishing flow, so the decision is made once, by a person, and recorded.

    What you would get

    • Marking applied at generation time, with a gate before publication
    • A visible-label step for realistic depictions of people
    • A record for each published file: when it was generated, how it was marked and who approved the label

    Watch for: Third-party tools in the middle of the pipeline. A single image editor that re-saves files can strip metadata without anyone noticing, which is why the check runs just before publishing.

    A CV-screening model nobody can explain afterwards

    Illustrative scenario, not a client engagement

    Company: An invented HR-tech provider whose model ranks job applicants for its customers' recruiters

    Relevant duties: Annex III, point 4 (employment) makes this a high-risk system. Article 12: automatic logging. Article 13: transparency to deployers. Article 14: human oversight

    The situation

    Recruiters see a ranked shortlist and a score. They don't see why a candidate scored low, and the system keeps no record of which model version produced which ranking.

    The provider has updated the model twice this year. Nobody checked whether its behaviour drifted from the documented specification after each update.

    What I would do

    1. Step 1: Define what must be reconstructable. With the team, I list what an auditor would need to rebuild one decision: the input, the model version, the score, the threshold and who acted on it.
    2. Step 2: Log events automatically. Each ranking writes a structured, tamper-evident event with those fields. This is the Article 12 layer.
    3. Step 3: Show recruiters what they need. The shortlist shows the main factors behind each score and the system's known limits, in plain language. This supports Article 13.
    4. Step 4: Add a decision point. When the model's behaviour drifts past an agreed threshold, a justification gate stops automatic ranking until a named person approves, overrides or pauses the system. This is the Article 14 layer, and it's the pattern the Spec-Drift Chronometer demonstrates.

    What you would get

    • Event logging that can rebuild any ranking after the fact
    • Plain-language explanations for recruiters
    • A drift threshold and a human decision point, with every decision recorded

    Watch for: Oversight that exists only on paper. If recruiters approve every shortlist in two seconds, the gate isn't oversight. The logs should show how long reviews take, so the team can see whether people are really reviewing.

    The decision point in the third scenario is the pattern behind the Spec-Drift Chronometer, my open-source proof of concept.

    Facing one of these?

    If your team recognises one of these situations, I'd be glad to talk it through. I usually start with a short diagnosis of where the gaps are.