Insights

AI in insurance requires decision-grade data

Why trusted data infrastructure determines whether AI moves from pilot to production

Rich Ward, Global Head of Insurance Practice

Insurance leaders are moving from AI experimentation to enterprise adoption. The priorities are clear: better underwriting, claims, fraud detection, service, cost, and risk selection.

But momentum is not the same as production value. In UST's 2026 Hanover study of North American insurance AI decision-makers, 43% of organizations described AI adoption as mature or optimized, while 39% said they are scaling. The market has moved beyond curiosity. The question is whether AI can operate reliably within live workflows.

That is where the pressure shows. In claims, the top barriers to scaling AI beyond pilots include unclear ROI, budget constraints, data readiness challenges, and workflow integration. In underwriting, respondents cited regulatory uncertainty, workflow integration, legacy systems, and data readiness as barriers.

The implication is clear: insurers are not short on AI ambition. They are being slowed by the foundations required to operationalize it safely and measurably.

For insurers, scaling AI is a data operating model problem before it is a model deployment problem.

DIVIDER

Why AI pilots stall in insurance

Insurers are asking where to apply AI first: underwriting assistance, claims triage, submission intake, fraud detection, customer service, policy servicing, and agent productivity. These are valid opportunities. The more important question is whether the organization can trust its data enough to make AI decisions at scale.

A pilot can succeed with curated data and manual reconciliation. Production AI has to operate within incomplete claim notes, inconsistent exposure data, disconnected policy and billing records, broker-submitted documents, product-specific rules, and regulatory constraints.

Consider commercial submission intake. AI may extract broker information, compare an account against appetite rules, recommend next steps, and preserve an audit trail. This depends on connected documents, account data, prior losses, coverage details, authority rules, and explainable recommendations.

If those inputs are inconsistent or disconnected, AI may still produce an answer, but the business cannot confidently use it.

DIVIDER

The business case is real, but it depends on readiness

The Hanover study shows that AI is already creating value where it has been implemented.

Among organizations that have achieved measurable ROI from AI in claims:

• 59% reported reduced manual processing time
• 59% improved claim accuracy
• 55% reported stronger compliance.

In underwriting:

• 60% reported higher employee productivity
• 58% improved customer satisfaction
• 54% improved risk segmentation
• 52% improved compliance.

Those results link AI to lower costs, better risk decisions, stronger compliance, faster cycle times, and an improved customer experience. But they occur only when data, workflow, governance, and operating ownership are strong enough to move AI from pilot to production.

DIVIDER

AI does not fix data debt

AI can summarize documents, detect patterns, recommend actions, and accelerate workflows. But when the underlying data is incomplete, inconsistent, poorly governed, or stripped of business context, AI can exacerbate the problem.

It can make weak data appear authoritative, generate confident recommendations from disconnected evidence, automate exceptions that should have been resolved upstream, and push operational risk deeper into underwriting, claims, service, compliance, and risk workflows.

Reporting tells leaders what happened. AI increasingly shapes what happens next. AI embedded in underwriting, claims, fraud, service, or risk selection requires decision-grade data: timely, explainable, governed, traceable, and connected to the business process where the decision is made.

What is decision-grade data in insurance?

Decision-grade data is trusted to support operational, analytical, and AI-assisted decisions in regulated insurance workflows. It is not just a cloud platform, lakehouse, data catalog, or governance program. Those may be required, but they are only part of the foundation.

Decision-grade data has five characteristics.

1. It is organized around insurance decisions, not just systems. AI needs to know which decision the data should improve: submission prioritization, claims severity, fraud detection, renewal underwriting, service routing, or risk selection.
2. It creates a common business language. AI struggles when customer, premium, exposure, claim severity, producer, or policy status have different meanings across products, regions, and platforms.
3. It connects structured and unstructured data. Submissions, loss runs, adjuster notes, broker emails, policy forms, photos, repair estimates, and reinsurance contracts all contain decision-critical information. Without a connection to structured records, AI remains a summarization tool. With it, AI becomes part of the decision infrastructure.
4. It embeds governance within the flow of work. AI requires quality rules, access controls, lineage, privacy controls, model risk management, human review, and auditability from the outset.
5. It supports continuous measurement. Leaders should know whether AI improved the underwriting cycle time, straight-through processing, claims leakage, severity prediction, producer satisfaction, employee adoption, and cost per transaction.

Usage is useful. Business impact is decisive.

DIVIDER

A practical AI data readiness checklist

Before scaling AI across underwriting, claims, service, or risk, insurers should ask:

• Is the required data available, accurate, timely, and complete?
• Are key terms consistently defined across products and platforms?
• Can structured and unstructured data be linked to the right policy, claim, customer, or exposure?
• Can the business explain which data informed the decision?
• Are privacy, lineage, human review, auditability, and impact measurement built-in?

If teams rebuild datasets for every AI use case, the issue is not model maturity but data readiness.

DIVIDER

Where insurers should begin

The path forward does not require a multi-year data program. It requires sharper prioritization instead.

Start with decisions that deliver measurable business value, such as submission prioritization, claims severity detection, renewal underwriting, portfolio monitoring, fraud detection, claims processing, policy servicing, or customer retention.

Next, assess whether the data is fit for each decision: available, accurate, timely, connected, consistently defined, governed, explainable, auditable, and measurable.

Finally, build reusable capabilities rather than one-off solutions. A claims summarization use case can improve document ingestion, entity extraction, retrieval, and claims data quality. An underwriting assistant can strengthen submission data products, appetite rules, semantic search, and producer insights.

DIVIDER

From decision-grade data to production AI

The next step is to connect AI ambition to operational and financial discipline. Every AI initiative should demonstrate how model output translates into workflow adoption, cost or risk impact, and the metrics leadership cares about.

Read UST's executive guide, From AI Pilot to Production, to learn how insurers can move from scattered pilots to production-scale AI by prioritizing use cases, identifying governance and integration gaps, aligning KPI ownership, and building a roadmap tied to impacts on the expense, loss, and combined ratios.

Get the insights