Insurers, and legal expenses insurers in particular, are examining how AI-assisted legal services could fit into their claims and service processes. Before a wider rollout, a bounded pilot allows an insurer to test its assumptions under controlled conditions: which matters are suitable, where human review is needed, which data has to move between which parties and how quality can be measured.
A pilot of this kind can be structured as a sequence of decisions: the question it should answer, the legal framework, the matter group, roles and checkpoints, data flows, metrics and the decision at the end. No specific insurer or project is described here, and no results are reported. The example outline further down is hypothetical and contains no performance figures.
Define the Question the Pilot Should Answer
A pilot is a decision tool. It should answer a question the insurer will actually decide on afterwards, for example whether a defined group of matters can be handled through a structured digital process with lawyer-led delivery, at a quality that the insurer and the law firm consider acceptable, and with handovers that work in daily operations.
Useful questions are specific and testable. They name the matter group, the part of the process under review and the criteria for continuing, adjusting or stopping. Setting these criteria before the pilot starts protects against reading success into whatever the data shows afterwards.
Typical questions include:
- Can the intake for a defined matter group capture the information a lawyer needs for a first assessment, without repeated follow-up questions?
- Where in the workflow do AI-generated summaries or drafts save the lawyer time, and how often do they need substantive correction?
- Do handovers between insurer, platform and law firm work reliably, including exceptions and escalations?
- How do policyholders experience the process, measured through complaints and structured feedback?
Know the Legal Framework Before Scoping
Several legal regimes shape what a pilot can look like. Legal, compliance and data protection functions should review them before the scope is fixed.
Legal services law and the role of the insurer
Under § 3 of the Legal Services Act (Rechtsdienstleistungsgesetz, RDG), out-of-court legal services may be provided independently only where a statute permits it. The legal assessment of a policyholder's matter therefore belongs with lawyers or other authorised providers. § 4 RDG prohibits legal services that can directly affect the performance of another obligation of the provider where this would endanger the proper provision of the legal service. The rule is relevant wherever the party paying for a matter would also advise on it.
Rights of the policyholder
The Insurance Contract Act (Versicherungsvertragsgesetz, VVG) contains specific rules for legal expenses insurance. Policyholders are entitled to choose freely from among the lawyers whose fees the insurer covers when they need representation in court or administrative proceedings (§ 127 VVG). A pilot can offer a recommended route to a partner law firm and must leave this right intact. If the insurer denies cover because a matter lacks sufficient prospects of success or is wanton, the contract must provide for an expert procedure or a comparable procedure, and the insurer must inform the policyholder of it (§ 128 VVG). Where an insurer writing several lines of business has assigned claims handling to an independent claims-handling company, that company is named in the policy (§ 126 VVG) and belongs among the parties to the pilot.
Professional duties of lawyers
Lawyers are bound by confidentiality covering everything they learn in the course of their work (§ 43a(2) of the Federal Lawyers' Act, BRAO). They may give service providers, such as software or platform providers, access to confidential information only to the extent necessary, must select them carefully and must bind them to confidentiality in a contract in text form (§ 43e BRAO). Within the pilot, acceptance of a matter, its legal assessment and communication with the client remain with the law firm.
Data protection
Legal matters often involve sensitive information, sometimes including health data, which is a special category of personal data under Article 9 GDPR. A pilot needs clear roles for each party (controller, joint controller or processor), data processing agreements where processors are involved (Article 28 GDPR) and an assessment of whether a data protection impact assessment is required (Article 35 GDPR). Decisions with legal or similarly significant effects that are based solely on automated processing are restricted by Article 22 GDPR, which is one reason to keep human decisions at the points where coverage and legal assessment are decided.
AI regulation
The EU AI Act (Regulation (EU) 2024/1689) classifies AI systems by their intended use. Its list of high-risk uses in Annex III includes AI systems for risk assessment and pricing in relation to natural persons in life and health insurance; obligations for these stand-alone high-risk systems were postponed to 2 December 2027 by the 2026 Digital Omnibus amendment. The transparency obligations in Article 50 apply from 2 August 2026. Among other things, providers must design systems that interact directly with people so that those people are informed that they are interacting with an AI system, unless this is obvious. Whether a specific tool in the pilot falls under any of these rules depends on its intended purpose and has to be assessed for each use case.
Outsourcing and digital operational resilience
An insurer that outsources functions or insurance activities remains responsible for compliance with all supervisory requirements (§ 32(1) of the Insurance Supervision Act, VAG). Since 17 January 2025, the Digital Operational Resilience Act (DORA, Regulation (EU) 2022/2554) has applied to insurance undertakings. Among other things, it requires a register of information on contractual arrangements with ICT third-party service providers (Article 28(3)) and minimum contractual provisions for ICT services (Article 30). Whether a legal-services platform counts as an ICT service or as outsourcing in a particular set-up should be clarified early, because the answer affects the contract and the documentation.
Choose a Bounded Matter Group
The matter group determines how much a pilot can show. Suitable groups share several characteristics:
- clear inclusion and exclusion criteria that claims staff can apply consistently;
- a recurring set of documents and facts, so that intake questions can be standardised;
- known deadlines that the process can track reliably;
- a limited need for urgent court action at the start of the matter;
- enough matters within the pilot period to observe recurring patterns, while remaining few enough to review each matter individually where needed.
Matters with short statutory deadlines need particular care. A claim against a dismissal, for example, has to be filed with the labour court within three weeks of receipt of the written notice (§ 4 KSchG). Such matters belong in a pilot only if the process guarantees immediate escalation to a lawyer. Exclusions should be documented as carefully as inclusions, together with the route each excluded matter takes instead.
Expected savings are a weak selection criterion at this stage, because they cannot yet be measured. A clear definition and the ability to review each matter are better criteria.
Assign Roles and Human Checkpoints
Each step in the workflow needs a named owner. A typical allocation looks like this; the details depend on the contractual set-up.
| Step | Responsible | Role of technology |
|---|---|---|
| Eligibility and coverage decision | Insurer | Structuring of policy and claim data; no automated coverage decision |
| Policyholder intake | Platform operator, as agreed between the parties | Guided questionnaire, document upload, completeness checks |
| Acceptance of the matter | Law firm | Case summary prepared for the lawyer |
| Legal assessment | Law firm | Extraction of facts from documents, research support, preparation of drafts |
| Approval and dispatch of letters | Law firm | Drafts prepared for review; dispatch only after approval |
| Status reporting | Platform operator and law firm | Status updates in the agreed format and data scope |
| Complaints and escalations | Party responsible for the step concerned | Logging and routing along the agreed escalation path |
Human checkpoints should be real decision points. The reviewing lawyer needs enough context to assess a draft, a simple way to correct it and the authority to stop the automated process. Corrections should be recorded, since they show where AI support works and where it does not.
Design the Data Flows
Data flows are easier to secure when they are defined narrowly from the start. Questions to settle before the first matter is processed:
- Which data fields does each party need for its step, and which does it not need?
- How is data transferred between insurer, platform and law firm, and who can access it?
- Where is data processed and stored, and which subprocessors are involved, including providers of AI models?
- Do the contractual terms of AI model providers permit the use of client data to train their models, and how is such use excluded where required?
- How long is data retained, and how is it returned or deleted at the end of the pilot?
- Which events are logged, so that decisions and changes can be traced afterwards?
Keeping the insurer's claim data separate from the law firm's client file supports both confidentiality and the separation of coverage decisions from legal advice.
Measure Without Presupposing Results
Metrics should be defined together with their data source, owner and reporting frequency before the pilot starts. Any comparison needs a baseline from the existing process; where no reliable baseline exists, the pilot can establish one and limit itself to describing the new process.
Possible measures include:
- Volume and suitability: matters referred, and matters inside and outside the defined criteria;
- Acceptance: matters accepted by the law firm, and reasons for non-acceptance;
- Time: time from referral to the first legal action and to completion;
- Quality: share of AI-assisted drafts requiring substantive correction, results of spot checks by lawyers, missed or nearly missed deadlines;
- Exceptions: escalations and process interruptions, with their causes;
- Experience: complaints and structured policyholder feedback.
Small pilots produce small samples. Differences between the pilot and the regular process may reflect the selection of matters as much as the process itself. Results should therefore be reported together with the number of matters, the selection criteria and the period, and interpreted with caution. Speed and cost figures alone say little about legal quality, which requires review by lawyers.
A Hypothetical Pilot Outline
The following outline is hypothetical. It does not describe an actual insurer or project and contains no results.
A legal expenses insurer wants to find out whether out-of-court disputes about residential service-charge statements (Betriebskostenabrechnungen) can be handled through a structured digital process with lawyer-led delivery. The pilot runs in four phases:
- Operational scoping. The insurer and its partners define inclusion criteria, such as existing tenancy cover and an out-of-court dispute about a service-charge statement, and exclusions, such as pending court proceedings or deadlines that require immediate action. They agree on expected volumes, communication routes and escalation paths.
- Governance and roles. Contracts set out data protection roles, confidentiality obligations under § 43e BRAO, data processing agreements, complaint handling and the policyholder's right to choose a lawyer. The insurer documents its outsourcing and ICT assessment.
- Integration and onboarding. The parties configure the intake questionnaire, data fields, status transitions and permissions, and test them with fictitious cases before live operation.
- Operation and review. During a pilot period fixed in advance, the parties review the agreed measures at regular intervals. At the end, they decide on the basis of the predefined criteria whether to continue, adjust or stop.
Each phase has its own approval criteria, and a phase begins only after the responsible parties have approved the previous one.
Questions for Technology and Legal-Service Partners
- Which legal entity provides the legal services, and how is it separated from the technology provider?
- Which tasks does AI support, and at which points does a lawyer review and approve the result?
- How are corrections to AI outputs recorded and used to improve the process?
- Where is data processed, which subprocessors are involved, and how are confidentiality obligations passed on to them?
- Which interfaces exist, and how are errors and outages handled?
- Which metrics can the partner report, based on which data and with which definitions?
- How are data return, deletion and an orderly exit arranged if the pilot ends?
Answers should be documented and, where relevant, reflected in the contract. References and certifications named by a partner should be verified.
Decide at the End of the Pilot
At the end of the pilot, the parties compare the results with the criteria agreed at the start. Three outcomes are possible: continuing with the same scope, adjusting the scope or process before a further phase, or stopping. Each of them is a usable result if it rests on the agreed criteria. The decision, the data behind it and the lessons learned should be documented, including findings on exclusions, escalations and corrections.
Extending the model to further matter groups works best as a new scoping exercise with its own criteria, since experience from one matter group may transfer to another only in part.
Advofleet's implementation model for insurance partnerships follows the four phases described above: operational scoping, governance and roles, integration and onboarding, and operation and review. Coverage decisions remain with the insurer, and legal decisions remain with the instructed law firm. Details are available on the page about the insurance implementation model.
This article provides general information and does not replace individual legal advice. Legal status: October 2026.