EU AI Act Employment High-Risk Classification for AI Task Allocation to Claims Adjusters
It depends on how the allocation logic works. Recital 57 of the EU AI Act classes AI that allocates tasks on the basis of individual behaviour, personal traits or characteristics, and AI that monitors or evaluates people in work-related relationships, as high-risk. Routing on claim attributes alone is analytically different from routing that draws on individual adjuster data, but borderline designs are a legal judgment for counsel. If a system is in scope, deployers need runtime controls and evidence showing what the system was permitted to do and what it did.
- Inventory every AI system that routes, prioritises or assigns claims, recording intended purpose, input fields and outputs.
- Mark each system for use of adjuster-level performance, behaviour or characteristic data, and for per-adjuster profiling.
- Check whether any allocation output is used by managers to monitor or evaluate adjusters.
- Document classification reasoning, including any reliance on Article 6 conditions, and obtain counsel's review.
- Define which fields each routing model may receive and enforce that policy at runtime.
- Place routing inputs under change control with re-assessment before release.
Claim attributes only (complexity, line of business, severity)
Inputs do not refer to an individual adjuster. Least likely to engage the individual behaviour or characteristics language.
Adjuster performance history as an input
Draws on individual behaviour. Engages the allocation wording directly.
Adjuster personal characteristics as an input
Engages the 'personal traits or characteristics' wording directly.
Workload balancing (queue depth, capacity)
Ambiguous middle case. Whether capacity signals count as individual behaviour or characteristics needs counsel's view.
Per-adjuster profiles built by the system
Article 6 states that an Annex III system performing profiling of natural persons is always considered high-risk, so exemption arguments are weakened.
Routing outputs reused as performance metrics
Raises the separate monitoring or evaluation category in recital 57, even if allocation alone was out of scope.
Mapping deployer obligations to runtime controls and evidence
Input data relevance
Restrict which fields reach the routing model or agent. If the design intends claim-only routing, block adjuster-level performance fields from allocation prompts and tool calls. Evidence: the policy version and the fields passed or removed. Whether this keeps a system out of scope is still a legal judgment.
Human oversight
Where a recommendation affects an individual adjuster, require human approval before the assignment tool call executes. Evidence: approver identity, decision and timestamp.
Record-keeping
Record each allow, redact, block or escalate decision with policy version, inputs and outcome, retained for the period counsel confirms.
What the system was permitted to do
Keep the policy configuration history so a reviewer can see which rules were in force on a given date.
Change control
Tie policy changes to approvals so that adding or removing a routing input leaves a reviewable trail.
What the employment category says about task allocation
Recital 57 states that AI systems used in employment and workers management should be classified as high-risk. It names recruitment and selection, decisions on terms of work-related relationships, promotion and termination, allocating tasks on the basis of individual behaviour, personal traits or characteristics, and monitoring or evaluation of persons in work-related contractual relationships. The stated rationale is the appreciable impact these systems may have on career prospects, livelihoods and workers' rights.
The Commission states that classification follows the intended purpose of the system. A system is high-risk if it is intended for a use case listed in Annex III, which includes employment. For a claims team, that means the question is not whether the product is called a routing tool. It is whether the system as designed is intended to allocate work using individual behaviour, traits or characteristics, or to monitor or evaluate the people doing that work.
The practical method is to split the system into input signals and decision outputs, then test each against that wording. Nothing in the Act's text speaks to claims routing specifically, so this is an analytical exercise that counsel should confirm.
Where classification remains a legal judgment
Several points cannot be settled by an engineering review. Article 6 lists two conditions that can support treating an Annex III system as not high-risk: that it performs a narrow procedural task, and that it performs a preparatory task to an assessment relevant to Annex III use cases. Reliance on either should be documented with reasoning, and it does not help where the system profiles individuals. The Commission has published guidelines with practical examples to help providers and deployers assess whether a system is high-risk, and these should be applied to each design.
Coverage is another question. Recital 57 refers to persons in work-related contractual relationships, not only employees, so carriers and brokerages that use independent or contract adjusters should confirm how that wording applies to them.
Design changes matter as well. Adding an adjuster performance field to a routing model can change the classification. Put routing inputs under change control, require re-assessment before release, and name an accountable owner for the classification decision.
Recital 96 says deployers of certain high-risk systems, such as banking or insurance entities, should carry out a fundamental rights impact assessment before use. Whether that applies to claims routing is a question for counsel.
Where runtime enforcement fits, and where it stops
Trussed enforces policy in the execution path on individual interactions and tool calls: allow, redact, block, or escalate for human approval. It produces audit-ready evidence mapped to the EU AI Act, and also to NAIC, NYDFS, HIPAA and NIST. That is enforced at runtime, not reviewed after the fact. For a claims-allocation design, an auditor can ask for a given assignment and see the policy in force, the data the system saw, the action taken and any human approval.
The limits are as important. Trussed does not determine legal classification, and it does not provide transaction-level or workflow-level governance across tasks and agents. Enforcement at the interaction and tool-call level supports a classification decision once it is made. It does not replace the decision, the counsel review or the organisational duties that sit outside the execution path.
Insurers should expect a regulator to ask two things: what the system was allowed to do, and what it actually did. Policy history answers the first. Per-decision evidence answers the second.
Claims allocation tested in four steps
Label: Input signals
Description: Claim attributes, adjuster-level data, or both.
Label: Decision output
Description: A queue, a priority or a named adjuster assignment.
Label: Downstream use
Description: Whether outputs feed performance monitoring or evaluation.
Label: Runtime evidence
Description: Policy in force, data seen, action taken, approver.
Evidence for what your claims AI was allowed to do and what it did
Talk through how policy enforcement on interactions and tool calls can support your classification decisions and deployer obligations.
Talk to an Expert