Professional indemnity insurance and AI agents in Europe. How the market is adapting and what to look for in your policy.

Professional indemnity insurance was designed around human professional error. The insured is a person or firm; the covered conduct is the provision of professional services; the triggering event is negligence, breach of professional duty, or an omission in the course of that professional service. AI agents do not fit neatly into this structure. They are not persons, they do not exercise professional judgement in the traditional sense, and their errors do not arise from negligence in the way that a human professional's errors do. The result is a significant coverage gap that European insurers and brokers are only beginning to address systematically.

How professional indemnity insurance works and where AI creates friction

Professional indemnity insurance covers claims made against a professional for losses that a third party suffers as a result of the professional's negligent acts, errors, or omissions in the course of providing professional services. The structure of the PI policy has evolved over decades around the concept of professional judgement: a qualified person, exercising the standard of care expected of a competent practitioner in their field, makes a decision or produces work product that turns out to be wrong. The policy responds to the claim arising from that human error.

AI agents introduce friction at several points in this structure. The first friction point is the identity of the insured conduct. Most PI policies define insured conduct as professional services provided by the named insured, its employees, and its principals. An AI agent is not an employee or a principal. Its output is not the product of professional judgement in the sense the policy contemplates. Where the AI agent produces advice or analysis autonomously, without a human professional reviewing and adopting it as their own professional output, the definition of insured conduct may not be satisfied.

The second friction point is the standard of care. PI policies are interpreted in light of the standard of care applicable to the relevant profession. Courts and professional bodies have developed clear norms for what constitutes reasonable professional conduct by a human practitioner. No equivalent standard exists yet for AI-assisted professional services, either in professional regulation or in the liability case law. The absence of a clear standard creates uncertainty about whether an AI agent's output satisfied or breached the applicable standard of care, which in turn creates uncertainty about whether the PI policy is triggered.

The third friction point is causation. A standard PI claim requires a clear causal chain from the professional's error to the third party's loss. AI agent failures often involve a more complex causal chain: the model produced an output based on its training data, the deployer provided a specific operational context, the user interpreted and acted on the output, and the loss materialised as a consequence. Each step in this chain may involve a different party's conduct. The PI policy may cover only the steps that are attributable to the insured's professional conduct, leaving the AI-specific steps uninsured.

The coverage gap in practice: three scenarios

Three scenarios illustrate where the coverage gap most commonly materialises in practice.

The first scenario is an AI agent providing legal or regulatory advice directly to clients. A law firm that deploys an AI agent to respond to routine client queries on employment law faces a PI claim when the agent gives incorrect advice about notice periods, causing a client to make an underpayment that triggers a tribunal claim. The PI policy covers professional negligence by the firm's solicitors. Whether it covers an autonomous AI agent that produced the advice without partner review depends on whether the firm's engagement letter disclosed AI use, whether the AI agent's output constitutes a professional service under the policy, and whether the failure to institute human review before delivering the output constitutes a professional breach attributable to the firm. Each of these questions involves genuine coverage uncertainty that should have been resolved before deployment, not at the time of the claim.

The second scenario is an AI agent conducting financial analysis. An asset management firm uses an AI agent to generate portfolio rebalancing recommendations that are sent directly to clients without human review, branded as the firm's professional advice. A market event causes the recommendations to produce losses that the client claims would not have occurred under the firm's established investment process. The PI policy covers the firm's professional advisors. The firm argues that the AI-generated recommendations are the output of its professional process. The insurer argues that the absence of human review breaks the causal link between the professional service and the output. The coverage dispute is foreseeable and the argument is not frivolous from the insurer's perspective.

The third scenario is an AI agent providing medical triage support. A private health provider uses an AI agent as a first point of contact for patients, triaging symptoms and recommending whether to seek urgent care. A patient whose symptoms were incorrectly triaged as non-urgent suffers harm from delayed treatment. The PI policy covers the health provider's registered clinicians. Whether it covers an AI triage function that operated without clinician oversight depends on whether the clinicians are legally responsible for the triage decision and whether the health provider's professional obligations include a duty to ensure adequate human oversight of any AI system deployed in the clinical pathway.

The three exclusion categories to review

A systematic review of European professional indemnity policies renewed in 2025 and 2026 reveals three exclusion categories that most commonly affect AI agent coverage. Brokers and in-house risk teams reviewing a PI policy before AI agent deployment should seek written clarification on each of these categories from the lead insurer.

Explicit AI exclusions

An increasing number of European PI policies now contain explicit AI exclusions following the addition of AI endorsements to London and European market standard wordings. The typical formulation excludes losses arising from or connected to the use of artificial intelligence, generative AI, machine learning, neural networks, or autonomous systems, subject to carve-backs for AI used as a tool under direct human supervision. The carve-back is the operative question for most deployments. What constitutes "direct human supervision" and whether a human reviewing a summary of AI-generated analysis, rather than the underlying AI output itself, satisfies the supervision requirement, is a coverage question that should be put to the insurer in writing before an incident occurs.

Automated decision exclusions

Separate from explicit AI exclusions, many PI policies contain automated decision exclusions that predate the current AI wave. These exclusions were originally designed to address algorithmic trading, automated credit scoring, and rules-based decision systems. They typically exclude losses arising from decisions taken by automated systems without adequate human review or oversight. The definition of "adequate human oversight" in these exclusions is relevant to the EU AI Act's Article 14 human oversight requirement, and compliance with Article 14 may support an argument that the oversight standard in the exclusion is met. However, the exclusion language and the regulatory standard use different frames and the interaction has not been tested in European courts.

Regulatory fine exclusions

The regulatory fine exclusion in PI policies excludes coverage for fines, penalties, and regulatory sanctions. This exclusion is standard in most professional liability products and is legally required in many jurisdictions because indemnifying regulatory fines would undermine the deterrent effect of the regulatory regime. The exclusion is relevant to AI agent deployments because an AI-related breach that triggers both a civil liability claim and a regulatory sanction, whether under the EU AI Act, GDPR, or sector-specific rules, will produce two streams of financial exposure. The PI policy may respond to the civil liability claim but will not cover the regulatory fine. Enterprises should assess their likely regulatory fine exposure and whether their total financial reserves or D&O coverage can absorb that exposure alongside the civil liability stream.

What Armilla, Munich Re aiSure, and Lloyd's are covering

The emerging standalone AI liability products are developing in a direction that addresses the PI gap rather than simply replicating it. Three products that are shaping the European market's understanding of AI-specific coverage are worth examining in detail.

Armilla, the Canadian AI insurance specialist that raised $25 million in January 2026 and operates as a Lloyd's coverholder backed by Chaucer and Axis Capital, provides coverage for AI system performance failure. Its product covers losses arising when an AI system produces outputs that fall below the performance standards specified in the deployment agreement. This is a fundamentally different insurance structure from PI: it does not require a finding of professional negligence. It requires only that the AI system produced outputs that failed to meet the contractually specified performance standard. For enterprise AI deployments where performance specifications are defined in vendor agreements, this structure may be more appropriate than a PI extension because it avoids the "was this professional conduct" question entirely.

Munich Re's aiSure product provides parametric AI performance insurance that triggers when measurable AI system performance metrics fall below defined thresholds. The February 2026 Mosaic partnership extended aiSure to cover AI developer risk with limits of up to EUR 15 million. The parametric structure means that coverage is triggered by objective, measurable events rather than by a determination of fault. For European enterprises, the challenge with parametric AI coverage is that it requires advance definition of performance metrics and trigger thresholds, which in turn requires a level of AI system transparency and monitoring capability that many deployers do not yet have in place.

Lloyd's of London, through multiple syndicate-backed products and the CFC Underwriting and Superscript facilities, is developing AI endorsements to existing technology E&O and PI structures. The Lloyd's market approach tends to be more flexible and bespoke than parametric products, allowing manuscript wordings that address specific deployment contexts. For professional services firms deploying AI agents, a Lloyd's brokered placement that includes a specifically negotiated AI endorsement to the PI base policy may be the most appropriate structure in the current market, provided the firm has documented its AI governance and oversight procedures sufficiently to satisfy underwriting requirements.

The Product Liability Directive layer from December 2026

Directive 2024/2853, the revised Product Liability Directive applicable from 9 December 2026, reclassifies software, including AI software, as a product. This reclassification has significant implications for professional services firms that deliver AI-powered services to clients, because it creates a strict liability product liability exposure alongside the fault-based PI exposure.

Under the revised Directive, a business that delivers AI-powered advice, analysis, or recommendations can face a product liability claim if the AI software contains a defect and that defect causes damage to a person, property, or data. The claim does not require the claimant to prove that the business was negligent. The claimant must show that the product was defective, that they suffered damage, and that the defect caused the damage. The burden of proof on causation is reversed in the Directive: where a claimant can show that the damage is of a type typically caused by the defect in question, the defect is presumed to have caused it unless the defendant can prove otherwise.

For professional services firms, the Product Liability Directive creates a coverage architecture question that most have not yet addressed. The PI policy covers claims arising from professional negligence. The product liability policy covers claims arising from defective products. For AI-delivered professional services, both may be triggered by the same incident. The PI policy limit may be consumed by the professional negligence claim, leaving the product liability claim uninsured. Firms should review whether their existing product liability coverage includes software and AI systems, and whether the limits are adequate for AI-related product liability exposure.

Structuring coverage for AI agent deployments in professional services

Based on the current European market, the recommended coverage structure for a professional services firm deploying AI agents depends on the degree of AI autonomy in the service delivery and the regulatory classification of the deployment.

For deployments where AI assists a human professional who reviews and takes responsibility for every client-facing output, a reviewed PI policy with an explicit AI endorsement confirming coverage for AI-assisted services under human supervision is typically sufficient. The AI endorsement should define what constitutes adequate supervision, should confirm coverage for outputs that a supervised human professional adopts as their own professional work product, and should address the regulatory fine exclusion by confirming that the policy covers civil liability arising from the same conduct that triggers a regulatory investigation, even if the regulatory fine itself is excluded.

For deployments where an AI agent provides advice or makes recommendations to clients with limited or no human review before delivery, the PI policy alone is likely insufficient. The recommended structure is a PI policy with an explicit AI endorsement for the human professional services the firm provides, combined with a standalone AI liability product that covers the autonomous AI agent's outputs. The standalone product should cover performance failure, hallucination-driven financial loss, and the liability chain from AI output to client harm. The total coverage limit across both products should be calibrated against the maximum realistic exposure from the highest-value client engagement the AI agent handles.

For all deployments, the coverage review should include an assessment of the Product Liability Directive exposure from December 2026 and whether existing product liability policies cover AI software defects. Most standard products liability policies in Europe were written before the Directive's reclassification of software and will require explicit confirmation of scope from the lead insurer.

The documentation that supports coverage eligibility is largely the same documentation that satisfies EU AI Act Article 26 deployer obligations. Governance records, human oversight logs, AI system incident reports, and performance monitoring data are all evidence that an insurer evaluating a claim, or an underwriter evaluating a placement, will want to see. Businesses that have built the compliance documentation for regulatory purposes are simultaneously building the evidence base that supports their insurance position. The connection between compliance readiness and insurability is the subject of our analysis of the documentation-to-coverage evidence chain. For the certification pathway that produces that documentation systematically, see agentcertified.eu.