Preparing Insurance Data for Agentic AI

Carriers and MGAs need data strategies that preserve context, support specific use cases and keep human judgment in the workflow.

(Image credit: IIR/Gemini.)

As insurance enters its agentic AI era, data quality is top of mind for CIOs, and for good reason. Asking AI agents to handle complex tasks, such as underwriting or claims processing, requires pinpoint data accuracy. Yet carriers and MGAs that wait for their data to be “clean enough” before deploying AI could be waiting forever.

Deploying agentic AI with confidence requires a different kind of data integration and hygiene strategy than what we saw during the Big Data era. The old instinct to clean everything once, store it centrally and serve it to whoever needs it is giving way to a new paradigm that preserves raw data and cleans it for specific use cases.

Getting Past the ETL Hangover

Many carriers’ AI data strategies are rooted in traditional Extract-Transform-Load (ETL) frameworks, which sanitize data into a one-size-fits-all version. Transforming data before it is used, however, can strip away the context and signals AI models need to make trustworthy decisions. The result can be data that is clean enough for everyone and therefore ideal for no one.

A more effective AI-ready data strategy removes the “transform” phase and preserves raw data, whether or not it appears valuable in the moment. Doing so enables AI teams to look at raw data retrospectively, then clean and tailor it for specific AI use cases.

Skilled underwriters already do a version of this instinctively, filtering out outliers that don’t reflect sound risk selection and pricing what remains. AI doesn’t change that principle. What it adds is the ability to apply that same discipline consistently across thousands of risks simultaneously, without the same cognitive load, and with clear guardrails for human intervention and judgment.

Working Within Legacy Constraints

An AI data strategy can only be effective if AI is embedded into existing systems and workflows. Legacy systems remain a persistent obstacle to AI ROI, but the question is not only one of integration. It is also about the data model underneath: whether it was designed around sound underwriting philosophy or around a vendor’s product roadmap.

Point solutions can deliver incremental gains, such as efficiency improvements in document-heavy claims intake workflows, but they often struggle to make an impact at the organizational level.

One reason is that AI added to existing systems can create data integration challenges, leading to uncertainty about who owns data and where it goes. That uncertainty can create compliance concerns. When carriers or MGAs load Protected Health Information (PHI) into a third-party tool such as ChatGPT, for example, that tool may become a sub-processor under the California Consumer Privacy Act.

Architectures that use APIs to connect with legacy systems can reduce these risks by keeping data closer to core workflows. New solutions also continue to enter the market, changing how insurers approach pricing, coverage and portfolio growth.

Keeping Human Judgment in the Workflow

As AI is embedded into complex workflows such as underwriting, human oversight becomes even more important. With simpler tools, such as AI-powered Optical Character Recognition (OCR), verification is relatively low-stakes. A person can compare the original document to the AI output and confirm the response. Once agents begin automating tasks step by step, however, confirming output accuracy becomes more difficult.

Human-in-the-loop checkpoints for agentic AI are essential, but carriers and MGAs should first ask a deeper question: Does the process being automated provide AI with enough data to make a logical, accurate decision? Most underwriters are already data-rich. The gap is in how dynamically that data can be interrogated and acted on at the point of decision. An AI agent operating on a static data picture hits the same ceiling, only faster.

Closing that gap requires a willingness to change how data is accessed, interacted with and applied.

Learning from AI Experiments

CIOs should also rethink how they measure the cost of AI failure. Traditional software investments required a solid business case, budget approval, vendor interviews, and a multi-month—or even multi-year—implementation that could run millions of dollars. In those cases, failure was not an option.

AI is different because it is relatively cheaper and lower risk, enabling IT teams to run multiple experiments. The successful ones may offset the costs of the unsuccessful ones. Failed experiments are not necessarily wasted, either. They can show IT teams what works and what does not, enabling subsequent test-and-learn cycles that improve the outcome.

AI success for insurance CIOs will be less about the sophistication of the algorithm than about data integrity and quality. Those who evolve their data pipelines beyond ETL, clean and prepare data for specific use cases, and embed AI into core systems will be better positioned to improve underwriting, claims and operational workflows without losing control of data quality or governance.

Agentic AI Blurring Core System Boundaries—Datos Insights’ Wein

Kevin Gaut // Kevin Gaut is chief technology officer at INSTANDA, an AI-augmented, cloud-native intelligent policy administration platform that enables carriers and MGAs to build, manage and distribute products at scale. His diverse career has evolved from COBOL programmer to the Head of Architecture at a large U.K. financial services organization and now to CTO. Along the way, he has implemented life and general insurance solutions inside major carriers and prominent InsurTechs.

Leave a Comment

(required)