Munich Re Chief Executive Officer Christoph Jurecka used the reinsurer's July 8, 2026 investor briefing to lift the wraps on aiSure, a purpose-built cover for losses caused by artificial-intelligence errors, with Bermuda-domiciled Mosaic Insurance as primary distribution partner. It responds on a parametric basis, uses fixed model-accuracy triggers rather than proof-of-loss adjudication, and caps at a per-claim limit of €15M.

€15M
Per-Claim Parametric Limit
Jul 8
Jurecka Investor Briefing
4
Named AI Peril Categories
Feb 27
Original Mosaic MGA Announcement

Key Takeaways

  • €15M per-claim parametric limit, roughly $15M at the July 8 spot rate, against a loss universe where a single corporate deepfake fraud has already been quantified at about $25M.
  • Four named peril categories: LLM hallucinations producing commercially harmful outputs, algorithmic-discrimination findings, copyright infringement tied to training or output, and AI-specific regulatory fines.
  • The trigger replaces adjudication. Traditional tech E&O runs 18 to 36 months of contested claims process; a benchmark breach pays a pre-agreed amount with no loss quantification.
  • Calibration is the underwriting act, committed at bind and not adjustable through the claims process the way subjective loss picks are.
  • Loss ratios crystallize in months, removing the IBNR cushion that lets a long-tail book absorb a pricing error across several development years.

What aiSure Actually Covers

The four peril categories are unusually explicit for a specialty cover at launch: large language model hallucinations producing commercially harmful outputs, algorithmic-discrimination findings from automated decisioning, copyright-infringement claims tied to generative-model training or output, and regulatory fines from AI-specific supervisors. A fifth adjacent category, model-performance shortfalls against contractual benchmarks, sits in Mosaic's MGA program as a separate performance-warranty layer. The target buyer is an AI vendor, or an enterprise deploying vendor models under a service-level agreement.

ItemDetailSource
Announcement dateJuly 8, 2026Ad-Hoc News, Munich Re
Original MGA launchFebruary 27, 2026The Insurer, Royal Gazette
Per-claim limit€15M (approx $15M)Munich Re product page
Named peril categoriesLLM hallucination, algorithmic discrimination, copyright infringement, regulatory finesMunich Re, Mosaic
Trigger mechanismParametric, fixed model-accuracy benchmark thresholdsMunich Re, Reinsurance News
Insurer partnersMunich Re (reinsurer), Mosaic Insurance (MGA distribution)Mosaic press release
Munich Re unitInsure AI, led by Michael BergerMunich Re

The parametric structure is the load-bearing design choice. Traditional technology errors-and-omissions covers require the insured to demonstrate a claim, prove causation to a defined loss, and negotiate a settlement inside a contested process that can run 18 to 36 months. aiSure replaces that with a fixed trigger: if a defined model-accuracy benchmark falls below a contractual threshold, or a named regulatory finding is issued, the cover pays a pre-agreed amount.

Reinsurance News reported the parametric mechanism was the element Mosaic underwriters spent longest defending during development, precisely because it compresses claims-handling expense to almost nothing while moving the underwriter's whole problem onto trigger calibration.

Mosaic's Krishnan Ethirajan framed the intent at the February launch: "AI developers face a rapidly evolving risk landscape, and traditional insurance products are not designed to keep pace." Jurecka framed the parent-level decision the other way, telling the July briefing that Munich Re would keep prioritizing technical margin over top-line growth and treat AI as a specialty class rather than a mass-market opportunity.

Calibration Is Where the Loss Ratio Is Decided

The trigger reshapes reserving before it reshapes anything else. Traditional professional-liability and tech E&O policies produce a long-tail development pattern: notified claims mature slowly, IBNR relies on Bornhuetter-Ferguson or chain-ladder projections over sparse triangles, and an accident year's ultimate loss ratio is not credibly known for five to eight years. A parametric cover collapses that timeline. The trigger fires or it does not, the settlement is contractual, and the policy-year loss ratio crystallizes in months.

The trade-off is basis risk. The insured's actual economic loss from a hallucination event or a discrimination finding may bear little resemblance to a payout tied to a benchmark-threshold breach. Munich Re calibrates thresholds per insured, through a pre-bind model-validation exercise measuring the customer's specific model against a battery of accuracy tests, and that calibration is the underwriting act.

Both directions of miscalibration fail visibly. Set the threshold too tightly and the trigger fires on ordinary model performance variance, so the cover behaves as a stop-loss on model quality rather than insurance. Set it too loosely and it fires only on catastrophic failure, so the insured reads the cover as economically empty. Michael Berger, who heads Munich Re's Insure AI unit, has framed the book as an actuarial rather than a legal proposition on exactly this point: trigger placement, not policy wording, determines the loss ratio.

No pricing analog fits. Cyber circa 2015 had the same absence of loss history but relied on subjective proof-of-loss adjustment, which produced a decade of adverse development. Parametric weather has the same trigger design but decades of physical-observation data behind its thresholds. Tech E&O written for hyperscale cloud vendors from 2018 through 2024 shares much of the insured universe, but its triggers and adjudication patterns do not transfer.

What the parametric design buys is speed of credibility: frequency signal arrives within months of inception, faster than cyber ever built it. What it costs is reversibility. A mispriced parametric trigger cannot be corrected through the claims process the way a subjective loss pick can be tightened, because the calibration is committed at bind. A vintage set too tightly spikes the loss ratio inside the same accident year the policies were written; one set too loosely reports a loss ratio that looks unrealistically strong and gets repriced without any reserve signal to flag it. Neither pattern is familiar to the specialty-casualty reserving cadre most reinsurers rely on.

Bespoke calibration also scales badly. It sidesteps the classification problem that left the industry reluctant to attach class codes to AI-driven business, because every policy is individually priced and no class-rate table is needed. Mosaic's syndicated distribution will eventually require either a tabular schema or a machine-priced quote engine, and both need the loss-experience credibility the book does not yet have.

The Limit Does Not Reach the Loss Universe

€15M is modest against the losses already in the record. The Arup Group engineering-firm deepfake fraud alone was quantified at roughly $25M, and deepfake-fraud losses through 2025 and into 2026 have run into the tens of millions for individual corporate victims. The limit covers the mean loss for a small-to-midsize AI vendor and does not touch the tail a large enterprise deployer faces from a single model-failure event.

So aiSure will not stand alone on a serious insured's tower. It sits under a bespoke facultative program the buyer arranges elsewhere, or inside a Mosaic-led syndicated placement stacking capacity well above the €15M primary, which Mosaic's original MGA materials confirmed as the design intent. That creates an unusual dependency for follow markets: the excess layer's attachment must be defined against the primary's parametric event rather than against traditional attachment-point exhaustion, so an excess carrier is underwriting Munich Re's calibration file rather than its own view of the loss.

The regulatory frame compounds the dependency. The FTC's July 7 policy statement on AI accuracy makes it a Section 5 exposure to market an AI system as accurate, objective or neutral where the design does not deliver on that framing. It applies to the insured vendors whose losses aiSure covers, and more awkwardly to Munich Re and Mosaic when they market aiSure in accuracy-adjacent language.

The NAIC Model Bulletin, now adopted in a majority of US states, requires insurers deploying AI to document model validation, testing and vendor oversight, and that is the same evidence base an aiSure underwriter needs to calibrate a trigger. Munich Re is therefore regulated as an AI-using insurer while insuring the AI errors of companies under the same framework, with both roles generating discoverable model-governance documentation.

An excess carrier inherits that posture along with the calibration, which is a more entangled position than a follow market normally accepts. Whether the payout obligation characterizes as a loss reserve, a contingent liability or a structured settlement will also differ between Munich Re's German supervisory regime and Mosaic's Bermuda framework, and the ceded flows between them need matching characterizations before the first statutory examination asks.

Further Reading

Sources