Frequently Asked Questions

We offer an Exploratory Data Analysis. This EDA involves collecting anonymized data, processing and cleaning it, performing descriptive portfolio analysis, feature engineering, model building, and model evaluation. This helps us identify the best AI solutions for your business.

Yes, we ensure that all data is securely hosted in isolated containers in the Azure Cloud and only anonymized data is used during the EDA process

We require a limited set of anonymized data, which can be provided through a data dump or collected via our partners.

The EDA process typically takes one week, depending on the complexity and volume of the data provided.

The EDA is available for a fixed fee ranging from €3,000 to €8,000, depending on the scope and requirements of your analysis.

We use automated feature selection, train over 10 ML models, and conduct automated cross validations to ensure the accuracy and reliability of our predictions.

Benefits include reduced churn, increased policy density, enhanced advisor efficiency, improved straight-through processing rates, optimized operational efficiency, and ensured compliance with regulatory standards.

We use APIs to seamlessly connect with your back-end data sources and integrate predictions into your front-end applications through interactive dashboards.

'Human in the loop' refers to our approach of supporting rather than replacing your employees with AI. It ensures that automation is balanced with human oversight and interaction where necessary.

Yes, our AI Engine is designed to help maintain high standards of customer care and regulatory compliance by providing scalable, informed customer interactions.

We provide comprehensive support during the integration process and ongoing support to ensure the continuous effectiveness of our AI solutions.

Absolutely, our AI Engine is modular and can be tailored to meet the unique requirements of your business.

We strictly adhere to data privacy regulations, ensuring all data is anonymized and securely handled throughout the process.

Our exclusive focus on the insurance sector, combined with our deep industry knowledge and strategic partnerships, ensures that our AI solutions are specifically designed to meet the unique challenges and opportunities in the insurance industry.

Yes, we comply with GDPR and other international data protection regulations, ensuring that all data handling practices meet the highest standards of privacy and security.

We have strict access controls and audit trails in place to monitor data usage. Only authorized personnel have access to your data, and all activities are logged and reviewed to prevent misuse.

After the EDA process, we ensure that all anonymized data is either securely archived or deleted in accordance with your preferences and regulatory requirements.

Based on the insights from the EDA, we work with you to select specific modules that best address your business needs. These modules are then used to provide predictions as a service (PaaS), ensuring continuous and accurate predictions.

Predictions as a Service (PraaS) is a solution where Onesurance continuously provides actionable predictions based on your data, without the need for you to manage the underlying infrastructure or models. This fully managed service allows you to focus on utilizing the predictions to drive business decisions.

• Predictions as a Service (PraaS) refers to a cloud-based service that provides predictive analytics capabilities to businesses. By leveraging advanced machine learning models and algorithms, PraaS allows companies to make data-driven predictions without the need for extensive in-house data science expertise or infrastructure. Users can input their data into the service, which then processes the data and generates predictions, insights, and recommendations.
  • Benefits of PraaS for the Client are:

   

  Cost Efficiency: 

  • Reduced Infrastructure Costs: Clients do not need to invest in expensive hardware and software for predictive analytics.
  • No Need for In-House Expertise: Eliminates the need to hire and maintain a team of data scientists, which can be costly and resource-intensive.

   

  Scalabilty

  • On-Demand Resources: PraaS can scale resources up or down based on the client's needs, allowing for flexibility in handling varying data volumes and computational requirements.
  • Pay-as-You-Go: Clients only pay for the resources they use, making it a cost-effective solution for businesses of all sizes.

   

  Speed and efficiency

  • Fast Deployment: PraaS solutions can be quickly implemented, allowing clients to start generating predictions and insights in a short time frame.
  • Real-Time Predictions: we offer real-time or near-real-time predictions, enabling clients to make timely and informed decisions.

   

  Advanced analytics

  • Access to Cutting-Edge Technology: Clients benefit from the latest machine learning models and algorithms, continuously updated and maintained by Onesurance.
  • Comprehensive Insights: PraaS can provide deep insights and uncover hidden patterns in data that might be missed with traditional analytics methods.

   

  Focus on core business

  • Streamlined Operations: By outsourcing predictive analytics, clients can focus on their core business activities and strategic initiatives.
  • Enhanced Decision-Making: With accurate and reliable predictions, clients can make better-informed decisions that drive business growth and efficiency.

   

  Security and compliance

  • Secure Environment: PraaS providers typically offer robust security measures to protect sensitive data.
  • Compliance: We ensure that services comply with industry standards and regulations, which is particularly important for sectors such as finance and insurance.

Predictions are delivered through APIs directly into your existing front-end applications, enabling seamless integration and immediate usability.

Yes, as part of our service, we provide a demo or pilot phase where you can see real-world examples of how our predictions can benefit your business. This helps you make an informed decision about fully implementing our PraaS solution.

• Our strategy is to establish partnerships with clients and leading insurance firms to provide insurance AI solutions at scale. To do so, we have designed and developed a standardized data model, known as the "Onesurance data model", which forms the foundation for training our AI modules at scale.
• The minimal data requirements of the Onesurance data model are designed to use only basic policy data, because this data typically is of high quality and is always consistently available in a mature insurance company. If such basic policy data were inaccurate, it would be almost impossible for an insurance company to operationally function. This would result in incorrect policies for clients or incorrect invoicing, which is the basis of a well-functioning insurance company. On their turn, the AI modules are designed to require only these basic data, such that we can immediately add value with AI, and we need no time-consuming data cleaning project on beforehand.
• In addition to the minimal data requirements, the Onesurance data model can manage many extra data sources, such as: customer data, claim handling data, detailed product data, customer contact data, employee data, and external data sources. In general, we prefer to ingest as much data as possible, under the important constraint that this data is of sufficient quality.
• Furthermore, we have automated the processes of data cleaning, data checks, and data transformation, leveraging our deep understanding of insurance data sets to ensure accuracy and efficiency.
• When connecting a new client or partner, our ‘automated data quality check’ identifies which data sources and corresponding columns are of sufficient quality (i.e., green flag) for AI module ingestion, and which are not (i.e., yellow/red flag). Subsequently, we apply our ‘standardized data cleaning tools’ to extend the eligible dataset with the data that need cleansing. Optionally, we can give an overall conclusion of the data quality and advice for future data quality improvements.
• In the Netherlands, we have realized a successful partnership with InsuranceData, a company founded by the accountancy and audit firm SVC Groep (a member of the International PIA Group). SVC/InsuranceData serves 70%-80% of the major brokers in the Netherlands, making them a significant player in the industry. Also see Onesurance and InsuranceData Partnership
• InsuranceData integrates and standardizes data from the most widely used back-office systems into their Business Intelligence (BI) platform. This integration capability makes them a logical and preferred partner for data integration. Based on our experience and vision, we believe that fast-growing brokers in the M&A field always need a BI platform to have real-time data available. As InsuranceData focuses on descriptive and diagnostic analytics, we can achieve a synergistic partnership with such BI platforms where we add predictive and prescriptive
• While we prefer to ingest data using the Onesurance data model, we recognize that this is not always possible. To date, we have successfully integrated multiple different datasets through data transformations conducted by the Onesurance team, in collaboration with the (data provider of the) client. Although transforming the data ourselves may increase lead time, we are of course fully capable of utilizing various datasets to feed into our models.

• As demonstrated through our work with InsuranceData, we are well-equipped to manage and oversee integrations. In fact, integrating algorithms is one of our core competencies, in addition to designing, building, and operating them.
• The feasibility of any integration depends on the size of the market and the associated benefits and drawbacks. Integrations with Agency Management platforms can provide substantial advantages for both the customer and the platform, offering a competitive edge without the need for the platform to establish its own data science team. Moreover, such integrations can enhance the effectiveness of the platform's data, thereby increasing its overall value.

• Our preference is to receive data in the Onesurance data format, which is a relatively simple model consisting of basic policy data.
• However, we understand that this may not always be feasible for all clients. For common data formats, we have developed standard data connectors that allow us to automatically clean and convert the data for our clients. This capability extends to common European or American data formats, ensuring that clients do not have to undertake the work themselves.
• Furthermore, note that our team of data experts have on average more than 10 years of experience in the field of data science, machine learning, and AI. Hence, they are remarkably familiar and comfortable with the process of collecting ‘exotic’ datasets and transforming and rationalizing them into a useful dataset for training purposes, if needed.

• Providing safe and ethical solutions is one of our core values, and this commitment is reflected in our comprehensive information security policies. These policies ensure compliance with all current EU regulations, which we are legally obligated to follow.
• We are part of the Cronos Group, the largest ICT company in Belgium, employing over 11,000 people. As such, we adhere to their stringent security standards.
• Compliance is not optional for us, and to ensure the safety of our clients' data, Onesurance currently employs a Data Protection Officer (DPO). We plan to certify this individual as a Chief Information Security Officer (CISO) in the nearby future.
• Regarding ethical considerations, we collaborated with Brush-AI during the development of our predictive models. Brush-AI, also under the Cronos umbrella, focuses on the ethical implementation of AI models. They assisted us in building a framework to ensure an ethical approach. This Ethical Framework includes an analysis of the ethical risks that the use of an AI model could pose, which is integrated into the design phase.
• The founder of Brush-AI was recently elected Responsible AI Leader of the Year, and the Netherlands was recently selected as the best performing country in the world regarding Responsible AI.
• Our clients however feel most comfortable that we only retrieve the necessary anonymized data and avoid accessing Personally Identifiable Information (PII) or other sensitive data unless absolutely needed.

• Currently, we do not use generative AI because this specific branch of AI has low usability for our current use cases. Combining other branches of AI are better suited for our current applications.
• Our AI solutions are highly innovative, because they are designed to address the unique challenges and opportunities in the insurance industry. Not as a single project, but as a scalable and affordable product with unprecedented fast time to value. Our solutions leverage the availability of basic data for midsize brokers, not just large carriers, while remaining fully compliant with current and forthcoming AI regulations.
• Our team continues to explore applications of generative AI and will introduce it when it can provide additional value to our clients.
• This perspective is supported by well-known ICT consultancy firms like Gartner. We believe that specific AI holds greater value in the insurance industry because it delivers precision, reliability, and transparency—qualities that generative AI does not currently provide.

What does Gartner say:

• The hype surrounding generative AI can lead to use of the technology where it is not a good fit, increasing the risk of higher complexity and failure of projects.
• Overfocusing on GenAI can lead to ignoring the broader set of alternative and more established AI techniques, which are a better fit for the majority of potential AI use cases.

Map your use case against the relevant use case family. GenAI is:

• • Highly useful: Content generation, conversational user interfaces, knowledge discovery
  • Somewhat useful: Segmentation/classificati
  • on, recommendation systems, perception, intelligent automation, anomaly detection/monitoring
  • Hardly useful: Prediction/forecasting, planning, decision
  •  intelligence, autonomous systems
  • GenAI may also be a poor fit for your use case if the risks that come with it are unacceptable and cannot be effectively mitigated. These include unreliable outputs, data privacy, intellectual property, liability, cybersecurity and regulatory compliance, either alone or in combination with one another.

• To prove the accuracy of our predictions, we employ several rigorous methodologies. Cross-validation is a statistical method we use to evaluate the performance of machine learning models. This involves dividing the data into subsets, training the model on some subsets, and validating it on the remaining subsets. F.e. We use a set of 5 years history of policy data. This involves dividing the data into two subsets: training the model on the subset of the first 4 years, and validating it on the subset of the last year of the subset.
• This approach allows us to demonstrate that if we can accurately predict churn in hindsight on a dataset the model has not seen before, we can also make accurate future predictions. We provide clients with the predictions as a proof of the accuracy of each model.
• We employ commonly-used accuracy metrics to evaluate the performance of our machine learning models. We use measures such as precision and recall to quantify the accuracy of the predictions generated by our machine learning models.
• We complement these metrics with additional metrics to measure the business value. For example, we simulate how much churn we could have prevented if we had approached the top x% highest-risk clients by checking whether the clients we predicted would churn actually ended up doing so.

• We can make accurate predictions for clients with as few as 50,000 observations. Among our clients, we typically see an average of 200,000 to 500,000 observations, and we generally achieve better accuracy with clients who have more data. However, even starting from 50,000 observations, we can obtain useful results.
• We do not combine data from all our clients for training purposes, because each client portfolio has different market dynamics and types of customers. We find that training on specific data from each client portfolio, rather than on aggregated data from all clients, leads to better results.
• Moreover, our clients typically do not want their data to be used for the benefit of their competitors. Therefore, we maintain strict data separation between clients' datasets in the Azure cloud.
• However, we are testing with new techniques such as federated learning and the use of synthetic data to benefit small and midsize brokers. Synthetic data is artificially generated data that mimics real-world data but does not directly originate from actual events or transactions. It is created using algorithms and simulations and can be used for various purposes in machine learning, data analysis, and software testing. Federated learning is a technique that focuses on collaboratively learning from multiple decentralized datasets to create one global (pretrained) AI model. We believe this could be an interesting focus topic for joint innovation in the collaboration between MarshBerry and Onesurance.

• We have robust after-sales processes in place, where we continuously gather feedback on both technical and user experience issues. Our solutions are designed to minimize the number of interaction points that can lead to problems. When issues do arise, we have consistently been able to respond within a 24-hour window, and we expect to maintain this level of customer service in the foreseeable future.

• Most predictions we provide are not critical to the insurance operational processes, except for the AI Underwriting Agent. For such critical predictions, we have more stringent Service Level Agreements (SLAs) with our clients and instant fallback methods to ensure the continuity of these critical processes.
• If there is a significant increase in the number of end-users, we have several strategies to handle the additional workload. These include increasing staffing levels and potentially outsourcing part of the customer service to a reliable implementation partner.
• Furthermore, we are pleased to announce that we have expanded our team with an experienced Customer Success leader to further scale our onboarding and partner and client success processes. Earlier this year, she has won a SaaS Award (saasawards.nl) for her accomplishments at her current/previous position in a big fintech scale-up.

• Our AI Engine has a modular assembly, which means that our software is designed in a way that separates core functionality from customizable components. This modular architecture facilitates easier updates and maintenance without affecting custom features.
• Core Modules: These contain fundamental features and functionalities that are common across all users like churn predictions and customer lifetime value.
• Extension Modules: These are separate modules for custom features that can be added or removed as needed.
• APIs and Interfaces: Well-defined APIs and interfaces allow for integration and customization without altering the core system.

• Currently, customization is primarily achieved through configuration, which allows users to change the behavior of the software through settings and options rather than altering the codebase. We can use configuration files (e.g. YAML) to store customizable settings, provide user-friendly administrative panels where users can adjust settings and preferences, and implement feature flags to enable or disable features based on configuration settings.
• This approach ensures that all our clients use the same product (the AI Engine) but with different settings tailored to their needs (e.g., which cost components are included in calculating the Customer Lifetime Value). By limiting customization to configuration, we can easily onboard dozens of clients per month.
• Additionally, we have a scalability plan that can be implemented as needed and effective version control and branching strategies in systems like Git help us manage customizations efficiently.

• Our team consists of experienced university-educated econometricians and computer scientists. On average, they have more than 10 years experience in designing, building and implementing statistical and mathematical algorithms to economic data, and transforming these algorithms into scalable and configurable software. The core members of our current team have successfully demonstrated this in the past years for corporates like Samsung (manufacturer), Coca-Cola (European Partners), Basic-Fit (subscription), Carrefour (FMCG retail), Sligro (wholesale/retail) or Corendon (travel).
• We undertake thorough validation and testing, as outlined in question 6 (Accuracy). We conduct rigorous testing of the software's outputs against known benchmarks and real-world data to verify its accuracy. Each component of the software is tested both in isolation (unit testing) and together (integration testing) to ensure proper functionality.
• We continuously monitor the software's performance in real-world scenarios to detect and correct any inaccuracies. Robust feedback mechanisms, such as thumbs-up or thumbs-down ratings to qualify predictions, are established to support users in reporting issues and suggesting improvements.
• We keep the software updated with the latest data and algorithms to maintain its accuracy and relevance. Additionally, we monitor for data and model drift to ensure qualitative results over time. This involves regularly calibrating the software’s parameters and algorithms based on new data and industry trends.
• At all times, we can share and make available current accuracy metrics to instill trust and remove any doubt about the actual performance of the models.
• We provide comprehensive training to clients and consultants on how to use the software effectively and can develop detailed documentation for MarshBerry to assist users in understanding the software’s functionalities and outputs.
• Of course, we must take into account that a tool is meant to support the consultant and advisor. We automate where possible but maintain the human touch where necessary—a concept we call “human in the loop.”