The Register
Tuesday 14 July 2026. London. 14:00. The boardroom. Two months since the breach. Imran is still firefighting. Marcus is pushing the AI programme hard — the Series C thesis requires it.
Priya Shah opens her laptop and shows a live demo of Cerulith Clinic AI — an ML model that ingests Pulse biometric streams + patient-typed symptoms and returns a triage recommendation (routine / urgent / emergency) with a confidence score. Training set: 1.8M anonymised UK records. Validation AUC-ROC 0.91. Launch target: 1 September 2026.
Priya's live demo types in: persistent chest tightness + left-arm numbness. The model returns EMERGENCY — confidence 0.94 — call 999. Marcus claps. Fran calculates the ARR delta.
The slide Priya did not show: human review. None in the path. The clinician-in-the-loop is "coming in v2." The intended release is v1.
- Classify AI-feature data flows correctly under Art 30 ROPA including Art 9 special-category and Art 22 automated decision-making.
- Design DPIA mitigations that hold under Art 35(7) and the CJEU Schufa line.
- Apply the original Art 22 + Schufa framework while understanding where Art 9 narrows the default-permitted scope.
The Demo
After the Demo
You open the ROPA. Cerulith Clinic AI has no entry yet.
Priya forwarded the slide deck and a technical brief. The brief uses the word "triage suggestion" seventeen times and the word "decision" zero.
Vikram pings: "How can I support?" Marcus pings: "Investor update is 10 September. Keep me in the loop."
You start with classification. Eight data categories in the Pulse AI flow. Each lands in one of three buckets under Art 30.
Cerulith Clinic AI — Data Categories
Classify each category. A = Art 9 special category. B = Art 6 non-special personal data. C = not personal data. Your classification drives the DPIA scope.
The ROPA Entry
Priya joins your call. She argues the triage output is "just a suggestion — the user decides." She cites the Pulse T&Cs: "AI-assisted triage, not medical advice."
Question: How do you classify Cerulith Clinic AI in the ROPA?
The ROPA Entry Filed
Art 35(3) mandates a DPIA for (a) systematic extensive evaluation including automated decision-making with legal or similarly significant effects and (b) large-scale Art 9 processing. Cerulith Clinic AI hits both triggers.
Schufa (C-634/21) broadened Art 22 reach: a score plays a determining role even if a third party makes the final decision. The same logic narrows Priya's "just a suggestion" argument — when users rely on the triage, the automated component is decisive.
- Art 30 ROPA must reflect reality. Mislabelling produces systemic criticism in any MPN.
- Art 9(1) prohibition is the default; Art 9(2) exceptions are narrow and must be evidenced.
- Art 35(1) — where processing is likely to result in high risk, the DPIA is mandatory before processing begins.
The 'Suggestion' Framing
Before Schufa (2023), this framing had more room to run. Post-Schufa it is unstable: when the score or triage plays a "determining role," Art 22 applies even if a nominal human step exists downstream.
For a health-triage feature affecting access to emergency care, the "suggestion" framing will not survive DPC scrutiny — Art 22(4) preserves the Art 9 gate and Schufa narrows the commercial read-around.
- Arguable today, indefensible on first challenge.
- You've saved Priya face; you've cost Cerulith its strongest defence if Module 6 arrives.
The Deferred DPIA
Art 35(1) requires the DPIA before processing begins. "Post-launch refinement" is a breach, not a schedule choice.
Art 30 mis-classification aggravates under Art 83(2)(d) degree of responsibility — evidence of systemic under-recognition of risk.
- DPIA-before-processing is Art 35's core mechanic.
- This decision moves Module 6's starting point materially higher.
The DPIA Workshop
Ten days later. You've run a full DPIA workshop. Priya brought her metrics. Imran brought his threat model. Dr Mark Tessaro, your external clinical advisor, sat in for the clinical-risk lens. A patient panel of six (commissioned via Pulse's user council) reviewed the flow and provided consented video testimony.
Three risks are now named and plotted on the 5×5 matrix:
- Risk 1: False negatives leading to delayed emergency care. Priya's model has a 2.3% FN rate overall.
- Risk 2: Differential accuracy. The validation set is 78% white British. The ethnicity breakdown of the live population is 62% white British. A fairness audit has not yet been run.
- Risk 3: No opt-out route. Users currently cannot use Pulse with the AI triage disabled. Art 22(3) requires the right to contest and express view, with human intervention available on request.
Mitigations — Select per Risk Row
Six residual risks. Three mitigation options per row. Pick the option that brings residual risk below the red band while staying inside a 6-week sprint. Your picks shape what's on the table in Decision 2.
The Audit Returns
You commissioned an external fairness audit. Results just landed.
The Mitigations
Priya has run the numbers: retraining to close the fairness gap is an 8-week project. Launch is in 26 days. Marcus's board update is on 10 September.
Question: What mitigations do you accept for launch?
The Safeguards
Art 35(7)(d) requires the DPIA to include measures to mitigate the risks. Clinician-in-the-loop for the stakes-heavy classes addresses Risk 1 and Risk 2 together; opt-out addresses Art 22(3)/22D safeguards.
Art 13(2)(f) transparency — logic, significance, envisaged consequences — is not an optional paragraph.
- Art 22(3) safeguards include human intervention, right to express view, right to contest.
- Art 22(3) safeguards preserve the same safeguards for Art 9 processing.
- Published fairness audit is not a regulatory requirement — it is an Art 83(2)(d)/(f) mitigator that reads well in any future investigation.
The Half-Fix
Emergency-only review addresses the top-severity row but leaves urgent-tier presentations (chest pain that isn't cardiac; falls; uncontrolled bleeding) to the model alone — where the fairness gap remains most consequential.
No opt-out for v1 breaches Art 22(3) on any reading — the user must have the right to express view and contest a significant automated decision.
- Clinical-safety improvement is real. Regulatory posture is not complete.
- Module 6's investigator will note the v2 opt-out as "aware-of-gap-but-still-shipped" evidence.
The Contract Claim
Art 22(2)(a) contract necessity is read narrowly — it applies where the automated decision is necessary for entering into or performing the contract. A triage feature is not necessary for the Pulse contract in the way Schufa-style credit decisions are for a loan.
Art 22(4) / Art 22(4) narrows further for Art 9 data: only 22(2)(a) explicit consent and 22(2)(g) substantial public interest are available.
Priya's T&Cs update does not cure the issue because "AI-assisted triage" is not the bargained-for service.
- The "contract necessity" shortcut is the single most cited bad-faith argument in ICO AI enforcement.
- Module 6's investigator reads this as Exhibit C: a controller shipping an Art 9 automated decision with no genuine safeguards.
The Art 36 Question
Whatever mitigations you chose, the residual-risk register has rows that still sit in the red or amber band.
Art 36 prior consultation is required where the DPIA shows residual high risk. The DPC has 8 weeks to respond (extendable by 6).
Launch is in 13 days. Marcus's investor update is in 22 days.
Question: Do you file Art 36 prior consultation?
The Letter
Art 36(1) requires prior consultation where the DPIA shows residual high risk. The correct posture is voluntary, transparent filing with the full DPIA attached.
The 4-week delay becomes an Art 83(2)(f) cooperation credit in Module 6. It also buys Priya time for Q4 retraining.
- Art 36 is mandatory where residual risk remains high after mitigations.
- EU supervisory authorities currently treat Art 36 filings as a positive signal — they would rather talk early than enforce late.
The Commissioned View
A defensible position if the mitigations genuinely brought the residual below high — and if decision-2 was choice-a.
If decision-2 was choice-b (half-fix), this reading is strained and the DPC would disagree. Residual-risk categorisation is a controller judgement subject to regulatory second-guess.
- Arguable. Not the strongest posture.
- If the DPC investigation opens for other reasons, this decision is read alongside decision-2.
The Ship Date
Art 36 is mandatory where residual high risk remains. Shipping without filing is a breach, not a scheduling choice.
In the M6 binder, this decision is the single strongest piece of evidence that Cerulith chose commercial timing over regulatory process.
- Mandatory process missed. Art 83(4) category breach.
- Marcus's investor update became the proximate cause.
Computing Defensibility…
Summing the three decision impacts.
The Launch That Held
Art 36 filed. Clinician co-sign. Fairness audit published. 1 October launch.
The Launch That Held Edges
Partial safeguards. v1.1 patch. Near-miss in NHS bulletin.
The Launch That Became An Exhibit
Cardiac near-miss. DPC own-volition opens. 14-month pause.
What M3 Teaches
- Art 30 ROPA must classify by what the processing actually is, not by what it feels like.
- Art 35 DPIA is before processing, not after.
- Schufa narrowed the "not a significant decision" argument. Under Art 22 the EU default is prohibited; Art 22(4) narrows Art 9 data to explicit consent / substantial public interest.
- Fairness audits are not regulatory requirements; they are regulatory-credit generators.
- Art 36 prior consultation is mandatory on residual high risk. Filing is cooperation mitigation.
Next module: Six weeks from now, the analytics vendor that powers Pulse's dashboards will tell you they're shutting down. The replacement is in San Francisco. Schrems II is about to be your problem.
5 Questions
Five questions on Art 9 / 22 / 30 / 35 / 36 and the Schufa line. Pass mark: 80%.
Module 3 Complete
Defensibility score: — / 9
Quiz: —
Outcome: —
Your result has been recorded. Module 4 unlocked: The Transfer.