AI Errors and Omissions Insurance in 2026. The Bifurcation Between Retreating E&O and Specialist AI Coverage.

The technology errors and omissions market is splitting in two. On one side, generalist carriers are adding AI sublimits and exclusions to standard tech E&O wordings that were never designed for probabilistic model output. On the other side, specialist writers are launching affirmative AI E&O products that name hallucinations, algorithmic bias, autonomous action errors, and AI regulatory defence costs as covered perils. Understanding which side of the split your current policy sits on is the most consequential insurance decision a technology company or AI-first business faces in 2026.

What E&O traditionally covered: the standard before AI

Technology errors and omissions insurance emerged in the 1990s to address a gap in general commercial liability: software shipped to a customer could fail in ways that caused financial loss, and traditional liability policies were not built to respond. The standard tech E&O form covered the insured's legal liability for losses suffered by a client because of a negligent act, error, or omission in the performance of technology services or in the supply of a technology product. The mental model was simple. A professional or firm provides a service. The service fails. A client loses money. The insured is held liable.

Crucially, the underlying assumption was one of determinism. Software, when it fails, fails in a specific and reproducible way. A bug in a payroll calculation either appears or it does not. A database returns a wrong result either always or when certain conditions are met. An insurer pricing tech E&O could look at the insured's quality assurance processes, its testing regime, its contractual limitations of liability, and arrive at a defensible premium. The probability distribution of losses was bounded and knowable, at least in principle.

The standard E&O form also assumed a professional act was at the origin of any claim. Behind every software failure was a human developer who made a choice, a project manager who approved a specification, an engineer who signed off a deployment. That human act of commission or omission was the insured event. It connected the loss to a named party in a named relationship. The insurance contract covered that relationship.

Exclusions in standard tech E&O were correspondingly narrow. Intentional acts were excluded. Bodily injury and property damage arising outside the technology context were pushed to general liability. Contractual liability beyond what would have arisen at law was excluded. But nothing in the standard wording contemplated a piece of software that had no single author, that could not be reproduced in testing, and that operated outside the direct supervision of any professional it was connected to. Those conditions did not exist in any meaningful commercial deployment until generative AI and autonomous agents became mainstream between 2023 and 2025.

How AI broke the E&O assumption: negligence versus autonomous output

An AI system, particularly a large language model in production or an agent operating within an enterprise workflow, breaks each of the three assumptions that underpin standard E&O simultaneously.

First, AI output is not deterministic. The same prompt submitted twice may return different answers. A model may perform correctly on ten thousand test cases and fail on the ten-thousand-and-first in a way that no human reviewer predicted. This is not a defect in the engineering sense; it is an intrinsic property of probabilistic systems. Insurers cannot price a risk whose probability distribution is not bounded by a human decision-making process they can audit.

Second, the connection between human professional act and loss is severed or attenuated. When an AI agent drafts a contract clause, advises a patient on medication, approves a loan application, or executes a trade, the human professional who built or deployed the system may be several degrees of separation from the loss. The traditional E&O question, what negligent act or omission produced this claim, does not have a clean answer when the proximate cause is a model output rather than a practitioner's decision. Courts are only beginning to work through how professional liability doctrine applies in these circumstances, and the European AI Act's allocation of liability between providers and deployers adds a layer of regulatory complexity that standard wordings were not drafted to address.

Third, AI agents perform autonomous actions. An agent authorised to book travel can book the wrong flight. An agent authorised to send customer emails can send a message containing hallucinated product claims. An agent authorised to execute orders within defined parameters can execute outside those parameters if its reasoning chain reaches an incorrect conclusion about what its mandate allows. These are not software defects in the traditional sense. They are goal-directed actions taken by a system that is doing what it was designed to do, while producing an outcome the operator did not want. No standard E&O wording in existence when the first LLM-based agents reached production had language for this class of loss.

The underwriting consequence is that the carrier cannot price the risk on the existing form. The claims team cannot use the existing doctrine to settle losses. And the legal relationship between the policy and the loss is ambiguous in a way that produces expensive disputes. Carriers faced two choices: price AI silently into the premium until the losses clarified the picture, or restructure the policy to push AI risk into a category that could be separately priced. In 2025 and 2026, the market moved decisively toward restructuring.

The retreat: named carriers reducing or excluding AI risk in standard E&O

The clearest institutional signal of the retreat from AI coverage in standard E&O and professional indemnity wordings came from the International Organization for Standardization's insurance affiliate. ISO Verisk made two new endorsement forms available to carriers from January 2026: CG 40 47, a broad exclusion applying to both Coverage A and Coverage B of commercial general liability policies, and CG 40 48, a Coverage B-only exclusion. Both forms allow carriers to exclude claims tied to generative AI, including AI use, AI output, AI training, AI-based advice, and AI-driven decision-making. A large portion of the US commercial market has either adopted or begun reviewing these forms for implementation at renewal.

In the London market, Beazley, one of the largest specialist writers of technology-related professional indemnity and media liability, has been developing AI sublimit language that would cap AI-related payouts at approximately 10% of total policy limits. As of April 2026, Beazley's head of cyber underwriting management Aidan Flynn confirmed the wording was still in development and had not yet been applied to in-force policies. But the intent is clear: AI-related losses will be segregated from the main limit and subject to a sub-ceiling, not covered on the same basis as traditional technology errors.

QBE has followed a similar trajectory. Draft QBE wording circulated in the London market in early 2026 restricts LLMjacking losses, scenarios where attackers access corporate AI accounts using stolen credentials and run up inference costs at the operator's expense, to approximately GBP 188,000 on a policy with GBP 3.8 million in total limits. That is a cap of just under 5%, with the broader AI sublimit framework targeting a ceiling around 10% of total limits. The signal is the same: AI is being separated from the rest of the covered risk and given a structurally lower ceiling.

Beyond Beazley and QBE, the broader market picture from the United States, which leads European carrier movements by six to eighteen months, shows carriers including AIG, W.R. Berkley, Great American, and Hamilton Insurance Group filing AI exclusion language with state regulators. Chubb has introduced an exclusion for widespread or systemic AI events while retaining some coverage for isolated incidents, a pattern that tries to separate idiosyncratic AI failures from correlated catastrophic exposures. Hiscox, a major tech E&O writer through both its US and London market operations, has been identifying newer AI and SaaS exposures as areas of evolving appetite, without confirming affirmative coverage as a standard position.

The net effect is that the standard tech E&O or professional indemnity policy a technology company purchased in 2023 or 2024 and has not scrutinised for AI language since is very likely sitting on a coverage assumption that no longer holds. The question is not whether the market is retreating. The question is how far the retreat goes on any specific policy.

The expansion: named specialist writers building affirmative AI E&O

Against the retreat of generalist carriers, a group of specialist and specialist-adjacent writers has been building affirmative AI E&O products that name AI loss scenarios explicitly as covered perils rather than leaving them in the ambiguous middle.

Armilla is the most structurally significant. It is a Canadian MGA that operates as the world's first Lloyd's of London coverholder dedicated exclusively to AI liability insurance, with Chaucer as the lead Lloyd's syndicate. Armilla's standalone AI Liability Policy covers hallucinations, model drift, inaccurate outputs, data leakage, AI regulatory violations, harmful AI outputs, AI agent failures, non-breach privacy incidents, AI-driven property damage, and defence costs under the EU AI Act and US state AI laws. Coverage limits reach up to USD 25 million per organisation. In January 2026, Armilla raised USD 25 million in additional funding to scale its products and distribution. The policy is available on a surplus lines basis and geographic availability varies by jurisdiction.

Armilla's approach is explicitly affirmative: the policy does not attempt to retrofit AI risk onto a general professional liability or cyber form. It introduces a broad, affirmative trigger that responds to core AI performance risks including malfunctions, hallucinations, critical inaccuracies, and breakdowns in expected function. Armilla also partners with Trustible for AI governance evaluation, which means the underwriting process includes an assessment of the buyer's AI risk practices, not just the product's technical characteristics.

Vouch is a US-based insurance platform focused on technology companies that has built out AI Insurance as a product category distinct from standard tech E&O. Vouch positions AI Insurance explicitly as a response to the coverage gaps in standard E&O. Hallucinations are covered. Claims related to algorithmic bias and discriminatory model outputs are covered. IP infringement allegations involving training data or AI-generated content are covered. Defence cost coverage for AI-specific regulatory inquiries under CCPA, GDPR, and similar frameworks is included. Vouch distributes via Hiscox Corix and its own platform.

Embroker has taken the approach of including an AI coverage endorsement automatically on every Tech E&O and Cyber quote it processes for technology companies. Rather than making AI coverage an optional add-on, Embroker treats it as table stakes for a technology business, with structured protection for AI-built businesses included by default. The endorsement addresses coverage gaps in traditional policies that have struggled to address AI-related risks.

Coalition offers a Technology E&O extension that addresses AI risk categories through its active cyber platform. Coalition's product is notable for combining active cybersecurity capabilities with insurance coverage, and its technology E&O extension reduces third-party liability risk for technology companies, with AI-related claims within scope as the product has evolved through 2025 and into 2026.

At-Bay provides tech E&O alongside its InsurSec platform, which combines active cybersecurity with insurance. At-Bay's product development has tracked AI risk closely, given that AI-driven attacks have become the dominant vector in its claims data. Its tech E&O product for technology companies addresses AI risk categories including those arising from AI-generated content and AI-assisted decision-making.

The middle: bundled AI riders on existing E&O

Between the retreating generalists and the specialist-only writers, a third group of carriers is threading the needle by adding AI coverage as an explicit rider or endorsement to an existing professional liability or E&O form. This approach avoids the structural cost of a standalone product while addressing the ambiguity in the base wording.

Counterpart is the clearest example of this approach done well. The Los Angeles-based insurtech, which has placed over 28,000 policies through approximately 2,800 brokers, announced in November 2025 an expansion of affirmative AI coverage across its professional liability product lines, including miscellaneous professional liability (MPL), allied health, and a new Technology E&O insuring agreement. The affirmative AI coverage addresses claims from AI-generated errors, inaccurate AI-generated reports, biased machine learning outputs, and misclassified data. It explicitly targets the silent AI exposure that sits between what standard professional liability covers and what AI-specific policies address. Counterpart's reach into small and medium-sized businesses makes it a significant distributor of AI E&O capacity to a segment that the specialist standalone products have not fully addressed.

AIUC represents a still more integrated approach: a certification standard that feeds directly into an insurance underwriting process. AIUC (The Artificial Intelligence Underwriting Company) launched in July 2025 with USD 15 million in seed funding led by Nat Friedman, with participation from Emergence Capital, Terrain, and Anthropic co-founder Ben Mann among others. Its AIUC-1 standard covers six domains: data and privacy, security, safety, reliability, accountability, and societal risks. The standard subjects AI systems to more than 5,000 adversarial simulations. In early 2026, Schellman became the first authorised auditor of the AIUC-1 standard. The insurance product, separate from the certification, protects enterprises against AI agent failures resulting in business losses. ElevenLabs became the first company to activate a AIUC-1-backed policy, covering its AI voice agent deployments. The certification score feeds the underwriting decision, which means that investing in AIUC-1 certification is simultaneously a compliance action and a coverage action.

The bundled rider approach sits in a meaningful but potentially unstable middle position. A rider on a standard form inherits the exclusions and definitions of the base policy. If the base policy's definition of a covered professional act narrows through endorsement or judicial interpretation, the AI rider may narrow with it. Buyers who rely on an AI rider rather than a standalone product should confirm in writing that the rider's coverage grant is not subject to the base policy's AI exclusion, if one exists, and that the two instruments are not written so that the exclusion takes precedence over the extension.

How to read your E&O policy for AI risk: an annotated checklist

A technology company or professional services firm that uses AI in its work should review its current E&O policy against the following checklist before its next renewal. This checklist addresses the seven most common structural gaps discovered when an AI-related claim is submitted against a standard E&O form.

1. Find the definition of a covered professional service or professional act. Does the definition include AI-assisted activities or AI-generated outputs? If the definition describes only acts performed by named professionals or specifically qualified persons, an AI system operating without direct human supervision may fall outside it.

2. Locate any AI exclusion or AI limitation language. Search the entire policy, including endorsements and schedules, for the words artificial intelligence, generative, autonomous, automated decision, and machine learning. Any wording that follows those terms in an exclusion clause controls what the policy will not respond to.

3. Check whether an AI sublimit has been attached. A sublimit is not an exclusion, but it is a ceiling. If the policy has a USD 5 million limit and an AI sublimit of USD 500,000, the effective coverage for the most likely AI claim is USD 500,000, not USD 5 million.

4. Read the definition of a covered wrongful act or negligent act. Ask whether the definition requires the act to originate from a human professional. If so, losses caused by an autonomous agent may not qualify.

5. Examine the technology products exclusion. Many E&O policies exclude coverage for losses arising from the malfunction or failure of a technology product that the insured supplies. If the insured's AI system is categorised as a technology product under the policy, a hallucination or model error may be excluded as product failure rather than professional negligence.

6. Look at the IP and data exclusions. Claims arising from AI training data copyright disputes, or from a model reproducing copyrighted material in its output, are a significant and growing category of AI-related claim. Standard tech E&O IP exclusions vary widely. Some exclude all IP claims. Some include a carve-back for inadvertent infringement. Understanding where AI-generated IP claims sit is essential.

7. Confirm the insurer's position in writing. Before relying on any AI-related coverage under a standard E&O policy, request from the broker or carrier a written statement of coverage position specific to the loss scenario the business is most likely to face. A verbal confirmation from a broker is not sufficient. The statement should name the specific AI-related loss type and confirm it is within the scope of the insuring clause.

When you need standalone AI coverage on top of E&O: decision criteria

The need for standalone AI E&O coverage, rather than a rider on a standard form, is not universal. The decision depends on the role AI plays in the business, the nature of the outputs the AI produces, and the consequence if those outputs are wrong.

A technology company for which AI is incidental, such as a software business that uses an AI assistant for internal documentation, can likely manage AI risk through a rider on its existing tech E&O, provided the rider is explicit and the base policy does not contain an AI exclusion that overrides it. The AI activity is subordinate to the core product, and the loss scenario, while possible, is not the company's primary liability exposure.

A company for which AI output is the core deliverable needs standalone coverage. This includes companies selling AI-generated legal research, AI-generated medical information, AI-driven financial advice, AI-automated customer service at scale, or any product where the user's reliance on the AI output creates a foreseeable liability path if the output is wrong. In those cases, the risk the company is primarily running is precisely the AI risk that riders are designed to supplement, not primarily cover. A sublimit or a rider ceiling is likely to be inadequate the first time a material claim is made.

Companies operating autonomous agents with consequential execution authority need standalone coverage regardless of product type. An agent that can book, commit, approve, purchase, or communicate on behalf of the operator is creating autonomous action liability that no standard professional liability or tech E&O form was designed to absorb. The agent's wrong action is not the human professional's wrong action. The standard form's insuring clause will be contested at claims time.

Companies subject to the EU AI Act as providers or deployers of high-risk AI systems under Annex III need specialist coverage with regulatory defence cost components. The Armilla and AIUC policies both include EU AI Act defence costs. A standard tech E&O policy written before the Act came into force will not have those provisions.

The signal that a company has crossed the threshold into needing standalone coverage is when its legal or compliance counsel cannot obtain a written confirmation from its current carrier that the specific AI-related loss scenario it is most exposed to is within the scope of the existing policy. When the carrier declines to confirm, or confirms with material qualifications, the coverage gap is structural and a standalone product is the corrective instrument.

Premium impact: illustrative ranges

All figures below are illustrative guidance based on publicly available market data and broker commentary as of April 2026. They are not quoted rates. Actual premiums depend on revenue, deployment scope, AI autonomy level, sector classification, certification status, and individual carrier underwriting decisions. Contact a qualified insurance broker for a binding quote.

For a non-AI software company purchasing USD 1 million in tech E&O limits, the illustrative market rate in 2026 is approximately USD 1,000 to USD 5,000 annually, with premiums scaling by company size, revenue, and complexity of software delivered. This baseline has been broadly stable through the soft market of 2024 and into 2025.

For a technology company that uses AI as a feature of its product rather than its core, adding an explicit AI endorsement or rider typically adds 20% to 40% to the base premium, assuming the carrier can confirm the AI activity clearly and the rider covers the relevant loss scenarios without a sublimit that materially undercuts the base limit.

For an AI-native company, meaning one for which the AI system is the primary product, moving to specialist AI E&O coverage with a reputable carrier typically produces a premium in the range of 1.5x to 2.5x the equivalent non-AI tech E&O premium for similar limits. A company that would pay USD 3,000 for standard tech E&O might pay USD 4,500 to USD 7,500 for specialist AI E&O at the same limit. The range is wide because the underwriting variables are wide: a company with AIUC-1 certification, documented governance, and a narrow autonomy envelope is a fundamentally different risk from an AI company with no governance documentation and agents operating with broad execution authority.

For an enterprise buyer operating AI agents in a high-risk sector under the EU AI Act's Annex III, meaningful AI E&O limits of USD 10 million to USD 25 million command premiums in the range of USD 50,000 to USD 250,000 annually, depending on the carrier, the scope of the AI deployment, and the presence or absence of certification evidence. The autonomy loading, which the Agent Insured framework estimates at 30% to 60% over the professional indemnity base, and the sector loading for Annex III applications, which runs an additional 15% to 40%, explain much of the premium step-up from the SME baseline.

Certification evidence is the single most controllable lever available to buyers at renewal. Carriers including AIUC, Armilla, and Munich Re aiSure all reference governance documentation quality in underwriting. An organisation that has completed ISO 42001 implementation, achieved AIUC-1 certification, or completed an Agent Certified assessment is presenting a materially different risk profile from one that has not, and the premium consequence is measurable.

Frequently asked questions