Clicksov
Book a call

USE CASE

Claims intake in production, on European infrastructure

One workflow. One KPI. Live in about eight weeks (indicative), and kept running afterwards.

The problem for a claims department

Claims arrive by email, through the portal and via API. Each one carries attachments: forms, photos, invoices, letters. A handler opens the file, finds the policy, reads everything, decides whether the claim is straightforward or needs attention, and prepares it in the core system. Most of that time is reading and sorting, not deciding.

The result is a queue. Lead time per case grows with volume, and experienced handlers spend their day on routine files.

The workflow, in six steps

  1. 01

    Trigger

    A claim arrives by email, portal or API.

  2. 02

    Fetch

    The workflow retrieves the file and the applicable policy terms from the core system, using the handler's own access rights through a connector. It never sees more than the handler would.

  3. 03

    Read

    An open-weight model running on EU infrastructure reads, summarises and classifies the documents. It is asked for a role, "extractor" or "classifier", through an AI gateway; the model behind the role can be swapped.

  4. 04

    Rule

    A decision table applies your thresholds and routing. Within limits: fast lane. Missing document: back to the customer. Above a threshold or flagged: to a senior handler. The table is configuration, readable and changeable by the business.

  5. 05

    Human check

    The handler sees the summary, the classification, the proposed route and the evidence behind it, and approves or corrects. Nothing is actioned without this step.

  6. 06

    Action

    The file is prepared in the core system, ready for the next step in your process.

What the handler sees

A single screen with the case summary, the extracted fields, the classification, the proposed route and links to the source documents. Every correction is recorded and feeds the evaluation set, so the workflow is measured against what your handlers actually decide.

What stays with the insurer

Data

Claim files and policy terms stay in your core system and in EU infrastructure under your tenancy. No training on your data.

Keys

Encryption keys stay in your own KMS or HSM.

Rules

Thresholds and routing are a decision table you own and can change without a release.

The workflow itself

It lives in one versioned repository that is part of your exit package: spec and prompts, domain model and your mapping, model manifest, evaluation set, deployment, policy, runbook. If you leave, it goes with you.

"Sovereign by default" means exactly this: data and keys in your hands, and a documented way out.

What is measured

The KPI is lead time per case. In week one, before anything is built, the baseline is measured on your current process. After go-live the same measure is reported continuously, alongside cost per completed case and per department. Every model swap is checked against the evaluation set before it reaches production.

What the AI Act documentation covers

Each workflow comes with its own AI Act documentation, maintained as part of the repository:

  • Risk class. The assessment of where this workflow sits under the AI Act, and why. Claims intake that prepares files for a human decision is assessed per case; the classification and the reasoning are recorded.
  • Use-case register entry. What the workflow does, which data it uses, who is responsible.
  • Guardrails. What the model may and may not do, what is filtered, what triggers a stop.
  • Evaluations. The evaluation set, thresholds and the results per release and per model swap.
  • Logging and human oversight. Tracing of every step (OpenTelemetry) and the record of the handler's approval or correction.

For high-risk systems the AI Act's main obligations were scheduled for 2 August 2026. In May 2026 the EU institutions reached a provisional agreement to postpone them to late 2027 and 2028; check the current status at the time you read this. The documentation above is useful whichever date applies; DNB already expects insurers to document development choices and monitor AI after deployment.

Timeline and pricing (indicative)

  • Week 1. Process scoped, KPI chosen, baseline measured.
  • Weeks 2 to 3. Connectors with access rights, rules moved into the decision table.
  • Weeks 4 to 6. Build and test against the evaluation set.
  • Weeks 7 to 8. Go live. Day-2 starts and is included in the run fee: monitoring, evaluations at every model swap, updates, cost reporting.

Pricing is outcome-based. One-off onboarding, indicative from €25,000. Monthly result fee per completed case, indicative €4 to €8, capped. Partner infrastructure passed through at cost and itemised: capacity, sovereignty, platform. The workflow runs on a certified European partner or on your own premises.

Want to know what the baseline would be for your department? Book a conversation.

Sources

Which process do you want to take live?

Book a call