Your Insurer Just Wrote AI Out of the Policy. Most of Your Codebase Is Now Written by AI.

What Changed in the Insurance Market This Year?

For most of the last decade, AI risk sat inside existing policies without anyone writing it down. Lawyers call this "silent AI": a general liability, cyber or technology errors and omissions (E&O) policy that never mentioned artificial intelligence, and therefore arguably covered losses that involved it. As Fenwick put it in June, the question for 2026 renewals is whether AI-related loss has been removed or narrowed through the structure of the policy form itself.

In many cases, it has. According to Insurance Journal, the Insurance Services Office (ISO), whose standard forms most US carriers build on, now offers three generative AI endorsements for commercial general liability. CG 40 47 excludes bodily injury, property damage and personal and advertising injury arising from generative AI. CG 40 48 applies the exclusion to personal and advertising injury only. CG 35 08 applies it to the products and completed operations hazard, which is the part of the policy that responds when something you sold or delivered causes harm after the fact.

Management and professional lines moved further. W. R. Berkley's endorsement, which Hunton Andrews Kurth described as an "absolute" AI exclusion for D&O, E&O and fiduciary liability products, bars cover for claims arising from "any actual or alleged use, deployment, or development of Artificial Intelligence by any person or entity". Its definition of AI is broad enough to capture essentially every large language model and machine learning system in production today.

Not every carrier is excluding outright. The Cloud Security Alliance research note on the exclusion wave reports that Beazley and QBE have introduced AI sublimits of around 10% of overall cyber policy limits instead. But a sublimit is still a ceiling, and for a software company whose main product risk now runs through AI-assisted code and AI features, a ceiling set at a tenth of the cyber tower is a material change.

Why Should a Software Company Read the Berkley Wording Twice?

Most commentary on AI exclusions focuses on companies that use AI chatbots or hiring tools. Software companies have a more direct problem, and it is visible in the text of the Berkley endorsement itself. Sub-clause (e) excludes claims arising from "any product or service sold, distributed, performed, or utilized by an Insured incorporating Artificial Intelligence". For a SaaS vendor with an AI feature, a development agency that uses coding agents, or a product company whose roadmap now includes copilots and assistants, that is not an edge case. It describes the business.

The endorsement goes further. Sub-clause (c) reaches claims arising from "inadequate or deficient policies, practices, procedures, or training relating to Artificial Intelligence", and sub-clause (f) covers representations or promises made by a chatbot or virtual customer service agent. A second limb excludes claims based on the company's statements about AI, including its "current or anticipated business plans, capabilities, or opportunities involving Artificial Intelligence". For a company that has spent two years telling investors and customers that it is AI-first, that clause deserves careful reading.

The operative phrase across these forms is "arising out of". As the CSA note points out, carriers read that phrase broadly: AI only needs to be somewhere in the causal chain, not the proximate cause of the loss. A defect introduced by a coding agent, missed in review and shipped to a customer, arguably arises out of the use of AI. So, potentially, does a data leak through an AI feature, or a wrong answer from a support assistant that a customer relied on.

Policyholder lawyers are alarmed. Courtney Horrigan of Reed Smith told Bloomberg Law that given how widely AI is now embedded, "I don't know how anything would be left" under a broad AI exclusion. That may prove too pessimistic once courts start interpreting the wording, but no software company wants to be the test case.

How Much of Your Product Is Already Inside the Exclusion?

The honest answer for most engineering organisations is: more than leadership assumes. The 2026 Stack Overflow Developer Survey, drawn from roughly 30,900 respondents, found that 73% of developers who use AI coding assistants or agents use them daily, and that a third of daily users spend four or more hours a day with AI tools. The most-used agents were Claude Code at 66% and GitHub Copilot at 59%. AI is not a side experiment in how software gets built. It is the default workflow.

That matters for insurance because the defect profile of AI-assisted code is different, not just faster. CodeRabbit analysed 470 open-source pull requests and found that AI-co-authored changes produced about 1.7 times as many issues as human-only changes, 10.83 per pull request against 6.45, with more critical and major findings as well. CodeRabbit is a code review vendor and its sample is open-source work, so treat the multiple as directional. The direction is the point: more code, written faster, with a higher defect rate per change, is exactly the risk an underwriter wants to price or exclude.

Litigation is moving the same way. Insurance Journal reports a Gallagher study finding that AI-related lawsuits rose 978% between 2021 and 2025, and 137% in 2025 alone, with patent infringement, copyright infringement and privacy-related personal injury the largest categories. Courts have also been unsympathetic to the argument that an AI system is somebody else's problem. In Moffatt v. Air Canada, the British Columbia Civil Resolution Tribunal rejected the airline's attempt to treat its chatbot as a separate entity responsible for its own statements and held the airline liable for the wrong information it gave.

Put those together and the exposure is clear. A company that ships AI-assisted code, embeds AI features and markets itself on AI capability now carries liability that its insurance programme may have quietly stopped covering. For European companies, this lands alongside the revised Product Liability Directive, which we covered in our article on software strict liability. Strict liability for defective software with a shrinking insurance backstop is not a comfortable combination.

Isn't Technology E&O Supposed to Cover Software Defects?

Technology errors and omissions insurance is the policy most software companies and development agencies rely on when a client claims that delivered software failed, caused a loss or did not do what was promised. It is the policy customers ask for in their master services agreements, usually with a minimum limit and a requirement to produce a certificate of insurance. It is also where the AI picture is least standardised.

There is no ISO-style standard AI form for tech E&O. Fenwick notes that some carriers now decline to cover liability tied to AI-generated outputs or add exclusions for AI-related errors and decision-making, while others offer affirmative AI endorsements that can, in practice, narrow cover by listing only specific scenarios. A May 2026 arXiv study that coded 55 AI threat classes against 26 insurance products found tech E&O among the legacy lines that still carry silent AI exposure, which means cover may exist today but has not been written down and could disappear at the next renewal.

For buyers of software services, this creates a problem that is easy to miss. Your contract with a development partner probably requires them to hold tech E&O at a set limit, and your procurement team probably checked the certificate once. The certificate says nothing about AI exclusions or sublimits. If your vendor's carrier added an AI carve-out at renewal, the indemnity in your contract may now be backed by an insurer that will decline the very claim you would most likely bring.

The same applies in reverse for software vendors. Enterprise customers are starting to ask, in security questionnaires and contract negotiations, whether the vendor's E&O covers AI-related errors. A vendor that cannot answer that question in writing will lose deals, or will sign indemnities it cannot fund.

What Are Underwriters Asking For Before They Will Cover AI?

The flip side of the exclusion wave is a growing market for affirmative AI cover. The CSA note lists Munich Re, the Lloyd's-backed Armilla, Counterpart and Testudo among carriers reportedly writing it. The Artificial Intelligence Underwriting Company has built a certification standard, AIUC-1, explicitly to connect technical assurance with insurance. According to SiliconANGLE, AIUC raised a $40 million Series A in September, runs each certified agent through around 5,000 risk and attack combinations, recertifies quarterly, and counts ElevenLabs, Cursor maker Anysphere, Harvey, KPMG, Lovable and UiPath among its certified customers or partners.

What these underwriters want is revealing, because none of it is exotic. The CSA note summarises six recurring requirements from a trade source: a documented human override with a recently tested escalation path; an inventory of AI systems that make or influence decisions affecting customers, employees or regulated outcomes; a data provenance audit of training and input data; a named executive who attests to AI governance controls; out-of-band authentication for high-value transactions that resists deepfakes; and evidence, from existing logging and identity systems, that these controls are actually enforced.

Read that list as an engineering leader and it becomes a backlog. An AI inventory is a registry of models, agents and AI features with owners. A tested override is a feature flag and a runbook that someone has exercised this quarter. Provenance means knowing which code, prompts and data came from where. Enforcement evidence means review records, audit logs and access control that a third party can inspect. The companies that will get the best terms on affirmative AI cover are the ones whose delivery process already generates these artefacts as a side effect.

There is a useful parallel with SOC 2 and ISO 27001, which turned security from a promise into a set of controls an auditor can check. AI cover is going the same way. Teams already working toward an AI management system will find the overlap substantial, as we described in our piece on ISO 42001 as a procurement gate.

Which Engineering Controls Actually Reduce Your Uninsured Exposure?

Insurance is a financing decision, but the exposure underneath it is created and reduced in the codebase. Five controls do most of the work, and all of them make a company easier to insure as well as less likely to need the insurance.

First, make AI involvement traceable. Tag pull requests and commits that were substantially generated by agents, record which tools and models were used, and keep that metadata with the change. When a claim arrives two years later, "was AI involved, and who reviewed it?" should be a query, not an archaeology project. Second, make human review real and recorded. A named engineer who understood the change and approved it is your strongest evidence against an allegation of inadequate AI practices, and it is the control most often skipped when agents generate more code than reviewers can read.

Third, put a kill switch on every customer-facing AI feature. Progressive delivery, feature flags and tested rollback, which we covered in our article on release safety for AI-generated code, are exactly the documented human override underwriters ask about. Fourth, keep a living inventory of every model, agent and AI feature in production, with an owner, a data flow and a risk rating. Fifth, constrain what AI features can promise. The Air Canada case turned on a chatbot inventing a policy; output validation, retrieval from approved sources and clear limits on what an assistant may commit to are cheaper than the claim.

None of these controls require slowing down AI adoption. They require treating AI-assisted delivery as a process with evidence, rather than a productivity trick individual developers use however they see fit. That is a delivery-management question as much as a technical one, and it is where the choice of development partner starts to matter.

What Should Change in Your Outsourcing and Vendor Contracts?

If you buy software development, your contracts were almost certainly written for a world of silent AI cover. Four changes are worth making at the next renewal or statement of work.

Ask for written confirmation, not a certificate. The CSA note's first recommendation to companies is to get written confirmation from every carrier on whether AI-instrumented losses are covered, excluded or sublimited, rather than relying on broker summaries. Apply the same standard to your vendors: require them to disclose any AI exclusion or sublimit in their tech E&O and cyber policies, and to notify you if one is added at renewal. Next, require an AI usage disclosure: which tools the team uses, under what policy, with what review standard, and how AI-generated changes are marked in the repository.

Then align the indemnity with the insurance. An uncapped indemnity for defects is worth what the vendor's balance sheet and insurer will actually pay. If the policy excludes AI, either negotiate a carve-back, agree a specific AI liability cap the vendor can actually fund, or accept the risk knowingly. Finally, make the evidence deliverable. Review records, AI tagging and the AI inventory for your product should belong to you and travel with the code if you switch vendors, which also helps with the issues we described in our guide to IP and code ownership in outsourced development.

How Does a Nearshore Dedicated Team Help You Stay Insurable?

The companies that will come through this transition well are not the ones that stop using AI. They are the ones that can show, quickly and in writing, how they use it. That is much easier when the people building your software are a stable team working inside your process, rather than a rotating set of contractors whose tooling and review habits you cannot see.

StepTo builds dedicated development teams in Serbia that work as an extension of your engineering organisation: the same senior engineers sprint after sprint, working in your repositories, under your review rules, with AI usage that is disclosed, tagged and reviewed rather than improvised. Our AI engineers and security specialists can help you stand up the artefacts underwriters and enterprise customers now ask for, from an AI system inventory to feature-level kill switches and review evidence that holds up years later.

Being in Belgrade also helps in practice. StepTo works on CET, the same business hours as Berlin and Paris and one hour ahead of London, with useful overlap for US East Coast teams. When a customer, a broker or an underwriter asks how an AI-assisted change reached production, the engineers who made and reviewed it are online the same day. Engagements typically start within 2-4 weeks of the first call, with a two-week paid trial, at rates 36-51% below Western European and US equivalents.

If your renewal is coming up and you are not sure what your AI exposure looks like from the inside, talk to us. We can help you map where AI touches your code and product, and what evidence you can put in front of an underwriter.