Claims triage, part one: from intake to controls

A composite insurer writing health and motor business wants to triage incoming claims. The system reads the claim form and attachments, extracts structured fields, scores the claim, and routes it to fast-track settlement, standard handling or investigation. A generative component drafts a summary for the handler. This module takes that system through every artefact the course has built, in the order the work happens.

Intake: the register entry

The first artefact is the register entry, because every other obligation is downstream of it. It names the system, the business and technical owners, what it does in one sentence, built or bought, your role under each regime, the lifecycle stage and the dates.

Two things are specific to this system. It is not one object: the extraction model, the scoring model, the generative summariser and the prompt are separate components with different providers, and registering them as objects in their own right is what lets the third-party assessment from module 8 attach to the right thing. And the intake question that matters is about use rather than purchase, because the Article 25 purpose-change route is the one nobody detects. The insurer records the intended purpose in terms narrow enough to be violated: triage and routing of first-notification claims, not assessment of coverage and not pricing.

Classification: two answers in one system

The health book and the motor book give different answers, and module 6 explains why. Annex III 5(c) covers risk assessment and pricing in life and health insurance. This system triages claims rather than prices risk, so the insurer assumes nothing and records the reasoning either way, with the point at which advice was taken.

The motor line is outside Annex III 5(c) on its face, which does not end the analysis. Article 26 deployer duties attach wherever the system is high-risk on another ground, and the data protection position is independent of the AI Act. Under Article 21 revFADP, a decision based exclusively on automated processing with a legal consequence or significant effect triggers duties to inform, to allow the person to state their position, and to allow them to request review by a natural person. A routing to investigation that delays payment for weeks is capable of being a significant effect.

The classification record states the tier under each regime, the criterion relied on, why adjacent tiers were rejected, whether anything makes the insurer a provider, and the residual uncertainty. It is countersigned by the business owner, because a classification nobody in the business has read is a compliance artefact rather than a decision.

Assessment: which one is owed

Module 6 is the test. If the health triage system is within Annex III 5(c), the Article 27 FRIA applies and the six elements have to be written: the processes it is used in, the period and frequency of use, the categories of persons and groups affected, the specific risks of harm, how oversight is implemented per the instructions for use, and the measures if risks materialise including complaint mechanisms. Results are notified to the market surveillance authority.

Element (e) cannot be written without the vendor's Article 13(3) instructions for use, which module 8 made a procurement condition. The sequence is deliberate: instructions first, then the DPIA under Article 26(9), then the FRIA cross-referring to the DPIA as the Digital Omnibus permits. The DPIA is owed on its own footing, because health claims data is sensitive and Articles 22 and 23 revFADP require an assessment where there is high risk to personality or fundamental rights, with FDPIC consultation where high residual risk remains. The Colorado work item in an early draft was deleted: module 6 established that duty was repealed when SB 24-205 was replaced by SB 26-189.

Data governance: the part that cannot be done later

Module 7 supplies the structure. The training dataset record follows the Datasheets for Datasets seven sections, mapped once against Article 10(2)(a) to (g) and against the California AB 2013 disclosures so one record serves all three.

Three findings come out of doing it properly. The assumption under 10(2)(d) is that the historical routing label records claim complexity; it records what handlers had time to investigate, which varies by office. Representativeness under 10(3) is measured against the population the system will be used on, and the insurer has recently entered a thinly represented distribution channel. And the setting under 10(4) differs between the books enough that a single model across both would be indefensible.

Where a fairness test needs attributes the insurer does not hold, Article 10(5) governs: special category data may be processed for bias detection only where strictly necessary and with safeguards, so it is collected narrowly, held separately, barred from use as a feature, and given a deletion date at collection.

Controls: oversight designed against automation bias

Article 14(4) sets the standard: the oversight person must understand the system's capacities and limitations, remain aware of automation bias, correctly interpret the output, be able to disregard, override or reverse it, and be able to interrupt so the system halts in a safe state.

Here the handler sees the extracted fields alongside the source document rather than the summary alone, because module 8 established that a generated justification is not the reason for the recommendation. The recommendation is shown after the evidence. Expected decision volume and time available per decision are written into the plan, because throughput reveals whether oversight is real. And the stop mechanism has a defined safe state: what happens to claims mid-flight when routing is suspended.

The bias and robustness records follow module 6. Thresholds are set in March and recorded in a register with its own dates, so September results cannot be graded against a standard written in September. The fairness calculation follows the NYC Local Law 144 structure of selection rate, scoring rate and impact ratio including intersectional categories, with the insurer's own action threshold held separately and explicitly not treated as a 0.80 pass mark.

As of 24 September 2026.