AI and insurance - productivity gains and emerging risks
AI is moving from experiment to everyday tool across the market. That message came through clearly at the Bermuda Risk Summit, where moderator Meg Green led a discussion with George Alayon of the Bermuda Monetary Authority, Peter Horrobin of Banyan Risk, and Devin Page of Ascot Bermuda.
The takeaway: AI can strip out hours of low-value work, but it also introduces new exposures we need to price, word, and manage with discipline.
AI is now a necessity, not a nice-to-have
"I think AI is going to be a necessity for the sector, and not a luxury any more," said George Alayon, deputy director, supervision (fintech) at the BMA. The Authority has built a dedicated data science and AI team and is upskilling supervisors - a clear signal that AI fluency is quickly becoming table stakes for regulated firms.
For carriers and brokers, that means building internal capability alongside third-party tools, and treating AI governance like any other model risk discipline.
Real productivity wins: from weeks to hours
Peter Horrobin, cofounder and co-CEO of Banyan Risk, shared a simple but telling example: an engineer built a new business intake form with AI in six hours - work that would have taken three weeks just months ago. His point wasn't headcount reduction. It was the clean removal of "horrific paper-driven cutting and pasting."
Where to point AI first:
- Submission triage and intake standardization
- Exposure summaries, bordereaux checks, and data remediation
- Wordings comparison and clause extraction
- Claims FNOL structuring and document classification
If you're building an internal playbook, start small, automate the mundane, and re-invest the freed-up time in underwriting judgment and broker relationships. For deeper how-to resources, see AI for Insurance.
New exposures on the front line
Devin Page, head of specialty insurance at Ascot Bermuda, flagged two fast-moving risk drivers. First, the surge in $10-$40 billion data centers - built as fast as power and grid access allow. Second, the risk that entry-level staff skip foundational skills because the rote work is offloaded to AI.
Practical implications to underwrite and manage:
- Data center property and BI: power reliability, water usage, heat, fire, and single-site concentration
- Supply chain: chips, transformers, backup generation, specialist contractors
- Systemic cyber: model monocultures, shared libraries, API dependencies
- AI model risk: drift, bias, IP infringement, data provenance
- Human capital: skill erosion, decision overreliance, and supervision gaps
For a useful framing on control families, see the NIST AI Risk Management Framework. For context on infrastructure demand and power constraints, the IEA's view on data centers is helpful: Data centres and data transmission networks.
Product design: stand-alone vs bolt-on
"I imagine that the amount of premium going to tech/cyber will continue to increase exponentially," Page said. Whether the market lands on stand-alone AI covers or bolt-ons, the wording work matters.
- Clarify triggers: model error vs. system outage vs. cyber event
- Define "AI system" clearly: first- vs third-party models, training data, and updates
- Exclusions and carve-backs: IP, bias/discrimination, regulatory fines
- Aggregation controls: shared cloud, model providers, and critical libraries
- Services coverage: prompt engineering, model monitoring, red-teaming as professional services
Why MGAs are built for speed
Horrobin argued MGAs will sit on the front line of emerging-risk solutions. Delegated authority plus specialist talent - engineers, digital-asset experts - creates room to experiment responsibly and underwrite what traditional structures might pass on.
Expect more insurer and reinsurer backing for skilled MGAs, paired with stronger governance around appetite, wordings, data quality, and post-bind monitoring.
How to operationalize this now
- Make AI a tool of the trade: give underwriters and claims pros approved assistants and playbooks
- Stand up model governance: inventory, testing, drift monitoring, documentation, audit trails
- Protect the junior pipeline: pair automation with training on fundamentals; mandate human review checkpoints
- Tighten vendor diligence: assess third-party models, IP indemnities, data lineage, and uptime SLAs
- Update wordings: remove silent AI exposures; add precise definitions and affirmative grants where intended
- Manage accumulations: track shared cloud/providers, critical libraries, and geographic clustering of data centers
- Scenario test: run AI-related loss scenarios across property, cyber, E&O, D&O, and BI
What to watch next
- Capacity continues to flow to tech/cyber and adjacent E&O
- Property and BI pricing adjusts around new data center clusters and grid constraints
- Regulatory focus grows on AI model governance and disclosure
- Hiring tilts toward MGAs for specialist roles; carriers partner rather than build everything in-house
- Market coalesces around clearer AI definitions and aggregation language
Final word
Insurance will never chase upside like VC, as Page reminded the room: our downside can dwarf the upside. Even so, with clear guardrails and the right partners, the sector can move faster than it gets credit for - and capture the productivity gains without taking blind risk.
Your membership also unlocks:
