Hiscox Cuts London Market Quote Cycle 99% With Gemini AI

Hiscox London Market has cut specialty quote turnaround from roughly three days to roughly three minutes using a generative AI underwriting model built on Google Cloud's Gemini and run through its internal Hailo platform. WTW wrote the first risk through it.

The headline is a 99.4% cycle-time reduction. The number that qualifies it is the scope: sabotage and terrorism renewals in the US and Canada, excluding the New York and Chicago metro areas.

Key Takeaways

• Three days to three minutes is a 99.4% cycle-time reduction on a live production workflow, not a proof of concept, running since the August 2024 launch.
• 15 or more data points are extracted autonomously from unstructured broker email attachments, then fed straight into Hiscox pricing models as inputs.
• 92 to 94% is the industry submission-extraction accuracy hyperexponential reports, leaving a 6 to 8% residual error rate sitting upstream of the indicative price.
• Renewals only, in one line, in two countries, with two metro areas carved out. Every piece of that scope is a model validation choice rather than a rollout convenience.
• 14% of specialty carriers run agentic AI in underwriting today against a projected 70% by 2028, which is the window in which this is a differentiator rather than table stakes.

What Actually Runs, and What Does Not

The system automates three steps and stops. Gemini reads broker submission emails and extracts 15 or more fields including insured name, location, coverage type, limits and prior loss history, geocoding addresses and normalizing formats that differ by broker. Hailo scores the risk against Hiscox pricing models and returns an indicative price. The system then drafts a pre-filled response email.

Final pricing authority and the bind decision stay with the underwriter. Hiscox also kept the AI out of customer-facing interaction entirely. Kate Markham, Hiscox London Market CEO, put the intent as freeing underwriters "from manual tasks and allowing them to focus on more complex risks where human expertise is critical."

Metric	Before AI	After Gemini	Change
Quote turnaround	~3 days	~3 minutes	-99.4%
Data points extracted	Manual entry	15+ automated	Full automation
Underwriter role	End-to-end execution	Review and approve	Shift to oversight
Engineering productivity	Baseline	+50%	Across tech teams

The three days it replaced were not three days of thinking. Day one was extraction and data entry, plus a follow-up email if the submission was incomplete. Day two was pricing and an aggregation check by geography and industry sector. Day three was drafting and reviewing the response. Two of those three days were handling, which is why the compression ratio is as large as it is and why it does not generalize to lines where the bottleneck is judgment.

Scope explains the rest. Hiscox picked a line it described as involving considerable manual data extraction and analysis, and started with renewals rather than new business, where prior policy data exists to check the output against.

The Extraction Layer Is Now a Pricing Input

Once extraction feeds pricing directly, extraction accuracy stops being an operations metric and becomes a rating one.

Hyperexponential's research puts industry submission data extraction accuracy at 92 to 94%. In a specialty line a single field can move the price by double-digit percentages: a misread limit, a misclassified industry code, or a geocode that drops a risk into the wrong aggregation zone. So the 6 to 8% residual is not a throughput loss, it is a distribution of pricing errors with a fat tail, and the exception-handling protocol is what determines its shape.

That is also why the renewal-only scope matters more than it first appears. On renewals the extracted values can be reconciled against the expiring policy, which gives the accuracy figure a benchmark. New business submissions are less complete, more varied in format and harder to classify, and there is nothing to reconcile them against. The validated accuracy today is accuracy on the easier half of the problem.

The prize is sized in the same research: loss ratio improvement of 3 to 5 percentage points for carriers implementing agentic AI, which on a $1 billion premium portfolio is roughly $40 million of annual underwriting profit, alongside quote-to-bind reductions of 60 to 99%. Hiscox's own baseline gives that room to show up: a 2025 group combined ratio of 87.8% and a London Market segment combined ratio of 85.9% on $1.25 billion of premium. Ki Insurance, also on Google Cloud, ran its algorithmic follow syndicate to $1.11 billion of premium at a 91.3% combined ratio, which is the comparison that matters more than any vendor projection.

The Aggregate the Quote Cannot See

Terrorism and sabotage is an accumulation line, and accumulation is the one thing a three-minute quote is structurally poor at.

The system generates an indicative price from the risk's own characteristics. It does not necessarily hold real-time visibility of the carrier's aggregate position across every terrorism exposure it has already written. Submissions for risks in the same geographic zone can be processed in parallel, each priced correctly on its own terms and collectively wrong on the portfolio's. The human review step is the control for that, which means the review is not a formality even though the automation has removed most of the reasons to linger over it. Excluding the New York and Chicago metros reads as the same concern handled by scope rather than by system.

Nothing market-wide is coming to help with the underlying data problem either. Lloyd's Council approved sunsetting Blueprint Two in March 2026 after six years, with re-platforming pushed to 2028 at the earliest and heritage systems supported until at least 2030. CEO Patrick Tiernan said the project "has not yielded the benefits that were originally envisioned." So submissions will keep arriving as inconsistent PDFs, spreadsheets and scans, and each carrier will keep solving extraction privately. Guidewire's 2026 barometer found 86% of respondents advancing their own technology strategies regardless, with 42% naming submission intake and extraction as AI's most valuable use case.

That leaves the accuracy question unauditable across the market. Every carrier's extraction layer is calibrated on its own submission flow, in its own line, against its own renewals, with no shared benchmark to compare against. The 92 to 94% figure is an industry estimate, not a measurement anyone can reproduce from outside.