Jump to Content
Financial Services

Generative AI in insurance: How Cytora helps insurers prioritize underwriting risks

March 15, 2024
Stathis Onasoglou

CFA - Fintech Principal, Google Cloud

Julius Von Davier

AI Customer Engineer, Google Cloud

Try Gemini 1.5 models

Google's most advanced multimodal models in Vertex AI

Try it

Speed, convenience, and consistent control over risk selection and decision-making matter in the insurance industry. This is why forward-thinking insurers are investing considerably in AI and digital risk processing technologies that streamline large parts of their risk workflows. These investments help insurers to consistently identify and choose the risks they want to quote, and consequently to align the business with their risk appetite and target portfolio. They furthermore help to reduce response times to broker requests from days to, for example, hours or even minutes. This allows for a strategy that potentially outperforms their peers through better underwriting margins. Cytora, a startup that helps insurers transform their workflows with AI, is helping to make these strategic alignments easier and more powerful. 

“Cytora’s SaaS platform enables insurers to digitize, automate and streamline their core workflows, evolving from manual risk flows to digital risk flows, where human capacity is deployed only when and where required. With Cytora, insurers boost their underwriting productivity, accelerate their broker responsiveness and drive profitability through better control over the target portfolio,” explains Juan de Castro, COO at Cytora.

In this article, we'll explore how Cytora leveraged Google Cloud's generative AI foundation models to deliver cutting-edge services to their clients — and how other businesses can learn from Cytora's example.

AI and “risk digitization”

Risk submissions (i.e., requests to assess underwriting risk) to insurers arrive in a variety of disparate formats (e.g., pdf, spreadsheets, emails) and more recently, also through a variety of broker APIs with varying schemas. This variety leads to considerable technical challenges for insurers implementing digital transformation programs. In current operating models, underwriters often have to spend time searching, parsing, and scraping information from all of these sources, keying and often re-keying it into multiple systems.

“The target state for insurers is for underwriters to receive what we call decision-ready risks with streamlined quotation workflows, accelerating decision-consistency, efficiency and turnaround time,” says Sam Lewis, VP of Product at Cytora.

This is what “risk digitization” is all about: information from disparate sources and formats is automatically parsed, evaluated, and, through a risk taxonomy, mapped into a computer-readable format that machines can understand and triage.

Automation with chain-of-thought prompting

An additional challenge is that although various brokers often need the same information and ask the same questions to the end client, they phrase the questions in different manners. Thus, two questions might be after the same information semantically but may look very differently textually. For example, one broker might ask:

Do you have established processes for rapidly applying critical security patches across servers, laptops, desktops and managed mobile devices?



Another broker’s submissions might ask the same information in a slightly different way:

What processes or controls do you have in place to ensure that all endpoints in your network are updated with critical security patches?

How frequently do you install critical and high severity patches across your enterprise?

1-3 days

4-7 days

8-30 days

One month or longer

Risk workflow automation technology needs to understand what each of these questions means, and then semantically match them to the insurer's internal risk appetite criteria. 

"By leveraging Google Cloud’s Vertex AI, and in particular Gemini, PaLM 2 and Gecko text embeddings, we have been able to transform our customer’s experience in implementing digital risk processing. Through zero-shot predictions and Retrieval Augmented Generation (RAG) our platform is able to start making predictions without any training data. The fine-tuning capabilities of PaLM 2 mean that it is possible to surpass previous risk digitization performance benchmarks with significantly fewer training data,” says Aeneas Wiener, CTO of Cytora.

Cytora’s risk digitalization layer is based on three pillars:

1. Adaptable and high-accuracy natural language understanding

Cytora leverages large language models (LLMs) in the Vertex AI’s Model Garden to extract the information across all submission attachments and emails. This is achieved through a combination of Cytora’s proprietary in-prompt tuning technology and PaLM 2’s fine-tuning capabilities, which enable private and fine-tuned models to be created for each of Cytora’s customers, based on their own private data. 

The privacy and security aspects of Vertex AI were particularly important for Cytora and their customers in the highly-regulated insurance industry. At Google Cloud the customer’s data is the customer’s data, including fine-tuned customer data, and Google provides enterprise assurances for privacy and security

“As Vertex AI was enterprise-ready from the beginning, we were able to rapidly move beyond prototype to using generative AI in production workloads,” says Wiener.

Additionally, Cytora’s proprietary technology for generating synthetic insurance training data alleviates the cold start challenge and allows for PaLM 2 fine-tuning from the start. Next to PaLM 2, a range of open source and commercial models are available on Model Garden through Vertex AI, which Cytora is also leveraging to build its platform. 

2. Risk taxonomy-driven digitization

Cytora is maintaining a list of semantically meaningful (and computer readable) target fields and field values to classify commercial risks. Cytora has developed out-of-the-box baseline risk taxonomy golden schemas for a wide range of commercial lines, including cyber, commercial property, liability, fleet, aviation etc.

3. Chain-of-thought risk understanding

Chain-of-thought prompting has shown substantial performance improvements for multi-step reasoning tasks for LLMs. Cytora has adapted and extended this technique to the commercial insurance risk processing domain, leading to an improved ability for the model to reason about risk in an auditable and explainable way. For example, a model that works on a cyber insurance claim can output: 

  • “The insurer requires the applicant to have endpoint protection in place.”

  • “On page 2 of the application form there is a declaration stating “Do you require the use of anti-virus on all end user devices? [X] Yes [ ] No”.“

  • “Anti-virus is a type of endpoint protection.”

  • “Risk evaluation result: { submission_within_appetite: True }”

An end-to-end platform for risk digitization

Risk digitization is the intelligent part that provides the insurer with a competitive edge. It’s not a process in a business vacuum though and Cytora’s experience in the market has surfaced the need for additional components (see Figure 1).


Figure 1: High level view of Cytora’s risk digitization platform

At high level, Cytora’s architecture consists of the following modules:

  • Risk flow engine - A flexible, low-code platform to define and execute multi-step, human-in-the-loop risk processing workflows. At the core of the engine, BigQuery is providing real-time business analytics, e.g. to track and optimize broker turnaround time SLAs and risk flow analytics, e.g. to measure conversion rates for different segments and identify growth opportunities.
  • Processing components - digitize, evaluate, decide - These are the core components that verticalize the solution to the commercial insurance use cases by adapting data to the right formats, evaluate request profiles and suggest the next best action.
  • Underwriter console - This is where decision support information is presented to underwriters and their decisions can flow back into the digital processes.
  • Human-in-the-loop console - This layer intelligently surfaces low-confidence automated extraction outputs to human operators and enables them to review and modify them, and ultimately capture their interventions as usable training data that feeds back into the system. These data are used to “upskill” the private client adapter layer of PaLM 2 both as in-prompt examples, as well as with model fine-tuning training runs.

To see the Cytora solution in action, it will soon be available on Google Cloud Marketplace. If you are interested in learning more about generative AI on Google Cloud, you can find more about the products here.

Posted in