Why insurers are re-underwriting cyber risk around AI
Cyber policies were built around a fairly stable set of loss scenarios: ransomware, business email compromise, a breached database, a DDoS outage. Underwriters priced those against known control baselines like MFA coverage, patch cadence, and backup isolation.
AI breaks that model in two directions. On the attack side, generative tools have cut the cost of producing convincing phishing lures and voice clones, so the volume and quality of social engineering attempts against finance and treasury teams has gone up without a matching jump in detection capability at most organisations. On the exposure side, businesses have started embedding AI features, chatbots, copilots, automated approval tools, into workflows that touch customer data and payment authorisation, often through a SaaS vendor's roadmap rather than a deliberate internal project. Insurers cannot underwrite what they cannot see, and most organisations do not have an inventory of where AI touches sensitive data or financial approval chains. That gap is what is pushing insurers to ask new questions before they will price a renewal.
What AI liability actually covers, and what it does not
AI liability is rarely sold as a standalone product in this market. It shows up as an endorsement or rider bolted onto an existing cyber policy, a technology errors and omissions policy, or occasionally a directors and officers policy when a board approved an AI deployment that went wrong.
Coverage that is genuinely available today tends to include financial loss from a flawed automated decision, intellectual property claims arising from generative output, and negligence claims when a business deployed an AI-enabled security control that failed to catch an attack it was marketed to stop. What is routinely excluded is more instructive: contractual liability the business assumed in a vendor agreement, losses from a "known vulnerability" the business had not patched, and increasingly, anything an insurer can attribute to a nation-state actor under a war or cyber-war exclusion. That last exclusion is where AI complicates things: attributing an AI-assisted attack to a specific state actor with confidence is harder than attributing a conventional intrusion, and a disputed attribution is exactly the kind of thing that stalls a claim.
The gap nobody prices correctly: silent AI
The most expensive mistake a business makes here is assuming a general cyber or professional indemnity policy automatically covers AI-related loss because AI is not explicitly excluded. That silence is not coverage. When a claim involves an AI system, and the policy was written before the underwriter considered AI risk at all, the claim moves into a dispute about intent and scope that can take months to resolve, sometimes ending in a denial rather than a payout.
The practical fix is not more insurance, it is asking the current insurer, in writing, whether AI-enabled tools, AI-generated attacks, and third-party AI vendor failures are explicitly in or out of scope. If the answer is vague, treat that as a finding, not a formality.
What underwriters actually ask for before they will price this
Before quoting AI-aware cyber cover, underwriters in this region are increasingly asking for evidence rather than a questionnaire tick-box. In practice that means:
- An inventory of AI systems in use, including embedded features inside SaaS tools, with a note on which ones touch customer data or payment approval.
- Endpoint detection and response coverage across the estate, not just antivirus. A named, monitored endpoint detection and response platform is one of the fastest ways to move a renewal conversation forward.
- A documented incident response plan that has actually been walked through in a tabletop exercise, not just filed away.
- Evidence of independent testing. Recent penetration testing results, including anything targeting AI-facing endpoints or APIs, carry more weight with underwriters than a self-assessment.
- A data governance answer to the question "what goes into which model", since that is the question a regulator will also ask after an incident.
- A change control process for AI features that get pushed by vendors rather than commissioned internally, since those are the ones nobody in the business signed off on.
Compliance and the insurance conversation are the same conversation
Underwriters and regulators are converging on the same evidence list, which is worth noticing rather than treating as coincidence. A governance, risk, and compliance programme built to satisfy a regulator's expectations, documented risk assessments, named data owners, tested response plans, produces exactly the artefacts an insurer wants to see before binding cover. Businesses that treat GRC as a paperwork exercise separate from insurance renewal end up doing the same evidence gathering twice, once badly under deadline pressure at audit time and once badly under deadline pressure at renewal time.
For a UAE or wider GCC business, that means the cyber insurance broker conversation and the compliance conversation should happen with the same evidence pack, not two different ones assembled by two different teams who have not spoken to each other.
What to check before the next renewal
Three questions decide whether an AI-related claim gets paid, and they are worth answering before a broker asks them:
- Does the current policy name AI-enabled attacks and AI system failures explicitly, or does it rely on silence to imply coverage.
- Can the business produce, within a day, a current inventory of where AI touches customer data or payment approval, and who owns that risk.
- Has the incident response plan been tested against a scenario that starts with an AI-generated lure or a manipulated automated decision, not just a generic ransomware run-book.
