As insurance markets become more volatile and data environments grow more complex, portfolio management is shifting from retrospective analysis toward continuous monitoring and rapid decision-making. According to Pardeep Bassi (pictured), global proposition leader – data science at WTW’s Insurance Consulting and Technology division, machine learning and AI are enabling insurers to detect emerging risks earlier and act on them faster.
The change is not only about better algorithms but also about the frequency with which insurers can analyze their portfolios. Continuous monitoring, Bassi said, allows organizations to move away from traditional review cycles and toward real-time management of pricing, underwriting and claims performance.
“By having machine learning and AI models always on, scanning emerging data, you’re able to pick out subtle trends quicker, but also pick out trends and effects that you may not necessarily have been able to pick out before,” Bassi said.
Traditional portfolio management processes often relied on monthly or quarterly reviews of performance indicators. That approach created a lag between the moment when trends emerged and the point when insurers could respond.
AI-driven monitoring significantly compresses that timeline.
“Rather than performing, say, monthly or quarterly reviews, you’re reducing that cycle time by leveraging capability which is on all the time,” Bassi said. “You get automated alerts instantaneously telling you something different to expectation has occurred. So there isn’t that lag between something happening and then you having the understanding that it has happened, before you can take the required corrective action.”
Continuous analytics also enables insurers to process far greater volumes of information than traditional manual analysis. According to Bassi, machine learning systems can assess multiple models and datasets simultaneously, identifying patterns across underwriting, pricing, and claims operations.
“One reason AI is powerful is the vast amount of data it can get through,” he said. “It can get through significant volumes of data and draw insight across multiple model expectations, outcomes and time periods.”
This capability allows insurers to uncover deeper connections across their operations. Bassi said organizations can detect when predictive models begin to deteriorate or when underwriting portfolios start shifting toward segments where predictive confidence is lower.
“You can make conclusions such as: you have seen a deterioration in your ability to predict a certain outcome for a certain segment,” he said. “Then you can instantaneously tell whether you have started to write more of that mix of business where you are less certain in your ability to predict that particular outcome.”
Cross-functional insights are also becoming more accessible. For example, signals from claims operations can feed directly into underwriting and pricing strategies.
“If you’ve got machine learning predictive models running across multiple functions, you can draw insight and take conclusions from, for example, real-time predictive models embedded in claims operations to understand changing claims costs as an insurer,” Bassi said. “You can dynamically take that insight into pricing or underwriting action.”
A key advantage of AI-enabled portfolio monitoring is the ability to identify emerging performance shifts before they appear in traditional metrics.
Insurers still track familiar indicators such as loss ratio, conversion rates, and claims frequency. But AI-driven monitoring allows those metrics to be monitored continuously, insight to be automatically generated and acted upon quicker by being linked to predefined thresholds that trigger alerts.
“When those metrics change by the amount in the predefined thresholds that you set, in terms of the change in the metric over a certain time period, that gives you the ability to understand as soon as possible when a metric has changed significantly,” Bassi said.
Beyond these core metrics, insurers are increasingly focusing on leading indicators that provide earlier signals of performance changes.
“By blending external data, internal data, and alternative data, you may be able to spot a change in an early warning indicator before it materializes into something as substantial, such as a loss ratio deviation for your book of business,” he said.
These early signals can allow insurers to adjust underwriting appetite, pricing strategies, or operational processes before trends significantly impact financial performance.
While automation is accelerating decision-making, governance and explainability remain central requirements for insurers deploying AI and machine learning.
Bassi said the growing complexity of analytics environments makes transparency non-negotiable.
“Explainability is absolutely core to all of that; it’s non-negotiable,” he said. “You need to be able to explain the decisions being made, and those explanations need to be interpretable by humans – for example, the drivers behind particular recommendations or decisions being made.”
This requirement is prompting insurers to rethink how governance is embedded within analytics processes. Rather than reviewing models after deployment, organizations are increasingly integrating governance earlier in the development lifecycle.
“As you become quicker and more sophisticated in your analytics, governance needs to be fed directly, earlier in the process,” Bassi said. “Rather than having governance as something which is considered as an afterthought, you embed governance as part of the model development, validation, and deployment.”
AI itself can also support compliance and documentation processes.
“You can also make use of AI itself to help generate documentation,” he said. “So the use of AI itself can help meet your specific governance requirements. It’s a tool for governance, as well as something that you need to govern.”
As insurers expand their use of predictive analytics, the number of models in production environments is growing rapidly. According to Bassi, many organisations now operate hundreds of models across different functions.
“The ability to deploy these models at a scalable, repeatably, robustly, and in a resilient manner, giving you the confidence that you can meet the business-critical SLAs that you require, is essential,” he said.
Once models are deployed, monitoring becomes equally important. Insurers must track both model performance and the underlying data feeding those models.
“The individual performance of the outcome each model is looking to predict, and the drift of the underlying data which is being used to feed the model, are both useful indicators and safeguards to have in place,” Bassi said.
Automated monitoring platforms are increasingly being used to oversee large model estates. These systems provide centralized oversight across multiple models and business functions, generating alerts when anomalies occur.
“By having automated monitoring of business outcomes and models in place, in a single location, a single location where you have that oversight across multiple models, across multiple parts of the business, you can gain confidence that your business is performing as expected,” he said.
Machine learning is also being applied within the monitoring layer itself.
“Machine learning & AI applied at the monitoring stage will give you insight into not only that something has changed, but help you to understand the cause,” Bassi said. “That can eventually lead to capability which suggests and recommends actions to take.”
Despite the increasing role of automated models, Bassi said human judgment remains critical in a model-driven environment.
Human experience often helps shape the monitoring frameworks themselves.
“Using your previous experience as a human to say, ‘These are the sorts of things which have gone wrong in the past,’ and engineering specific monitoring assesments based off that knowledge and understanding is key,” he said.
Human expertise can also highlight areas of uncertainty that models may not immediately detect.
“You may have some context as a human which the machine learning or AI doesn’t,” Bassi said. “You can use that context and your judgment to say, ‘I’m a little uncertain about this particular segment,’ and ask the monitoring capability to pay particular attention to this.”
Looking ahead, insurers are also exploring the use of AI agents capable of making autonomous decisions within defined guardrails.
“We’re seeing an increased appetite to explore the use of AI agents, in particular across insurers,” Bassi said. “AI agents are, in fact, going to be making more and more decisions autonomously, with human-set guardrails and controls.”
That shift will introduce new operational and monitoring requirements.
“In the world of AI agents, the monitoring requirements will need to change and adapt,” he said. “You’re monitoring AI agents which autonomously make decisions. The need for control and monitoring is increased.”
