Why Insurance IT Leaders Pick AI Layers Over Rebuilds

In 2020, insurers spent about $194 billion on IT, and forecasts put the industry on track for $271 billion in IT spend by 2025. That kind of money should translate into smoother operations, better underwriting decisions, and faster claims. Yet plenty of carriers still feel stuck—especially at the moment when AI is supposed to be paying off.

Most companies get this wrong by assuming AI success depends on buying more tools or rebuilding core platforms from scratch. The reality? The fastest wins usually come from adding an “intelligence layer” on top of what you already run—your CRM, policy admin, claims platform, data lake, and data pipelines—then putting the resulting insights directly into the hands of underwriters, adjusters, agents, and service teams.

This article is part of our AI in Insurance series, and it focuses on a practical question I hear from data and IT leaders all the time: How do we scale AI in underwriting, claims, risk pricing, and customer engagement without turning our tech stack into a science project?

The real bottleneck: the “last mile” of insurance AI

Insurance AI fails most often at the last mile: getting the right insight to the right person, inside the right workflow, at the moment of decision.

A carrier can have a solid data lake, decent feature engineering, and a portfolio of predictive models—then still struggle to move the needle because insights sit in dashboards, spreadsheets, or siloed apps. Underwriters don’t open another portal. Claims handlers won’t switch tabs ten times per file. Contact center teams can’t interpret complex model outputs while a customer is on the phone.

Here’s what “last mile” failure looks like in underwriting and claims:

• Underwriting: A risk model flags increased exposure, but the underwriter doesn’t get a clear next step. Result: default decisions, inconsistent appetite enforcement, and missed premium opportunities.
• Claims: A fraud signal exists, but it isn’t embedded into the triage workflow with supporting evidence. Result: either too many false positives (wasted SIU time) or too many missed fraud cases.
• Customer engagement: The carrier “personalizes” with generic segmentation because product recommendations aren’t consistent across agent tools, self-service, and service scripts. Result: customers buy on price, not protection.

The fix isn’t more experimentation. It’s operationalization: turning AI outputs into actions that fit the way insurance work actually happens.

Why this matters more in 2025

By late 2025, most insurance executives are past the “should we use AI?” phase. The pressure now is measurable ROI and governance—especially with responsible AI expectations rising across compliance, risk, and audit teams.

When budgets get reviewed, the easiest projects to cut are the ones that look like endless model tinkering with unclear production impact. Data and IT leaders need a path that makes AI:

• usable (embedded in workflows)
• consistent (same recommendations across channels)
• governable (privacy, explainability, controls)
• fast to iterate (without months of custom development)

Why data and IT leaders prefer “intelligence layers” to new stacks

Data and IT leaders don’t want another platform because every new system adds integration load, governance overhead, and support complexity.

Most insurers are already running multi-year, multi-million-dollar modernization programs across:

• Customer Relationship Management (CRM)
• Quote and underwriting management
• Policy administration
• Claims management

These programs matter. They create the stable, secure backbone insurers need. But they’re also slow by design—migration, testing, vendor coordination, and regulatory constraints take time.

An insurance-focused intelligence layer (like the one described in the RSS source) is popular because it:

1. reuses existing data assets (core systems, data lakes, APIs)
2. adds prescriptive decision support rather than just analytics
3. pushes insights into existing touchpoints (agent tools, service consoles, digital journeys)

Put bluntly: it’s easier to justify spending on AI when it increases the ROI of systems you already pay for.

Two high-impact investment areas insurers want to protect

Data and IT leaders typically have two major buckets of prior investment they don’t want stranded:

1. Core systems of record (policy, claims, CRM)
2. Data infrastructure (lakes, warehouses, streaming, APIs, MDM)

The intelligence-layer approach strengthens both:

• It brings insurance-specific recommendations into the core workflows.
• It turns data lake insights into front-line actions, not back-office reporting.

What “insurance-grade AI” needs to deliver (beyond automation)

Automation saves time, but prescriptive AI changes decisions. That’s where underwriting profitability, claims leakage, and retention actually move.

From the source content, the differentiators that matter to IT and data leaders cluster into six requirements. I’ll translate each into what it means operationally.

1) Insurance scores and predictions you can use

Scores aren’t valuable unless they’re tied to actions. A risk score should come with:

• a reason code or explanation a human can validate
• a suggested follow-up question
• a recommended coverage or endorsement
• a confidence indicator (so teams know when to trust it)

In underwriting, this becomes “quote faster with fewer blind spots.” In claims, it becomes “triage and assign the right handler early.”

2) Risk assessment that improves data quality

A surprisingly big AI benefit is better data capture. When AI suggests warnings and questions to refine a customer profile, it reduces missing fields and inconsistent risk details.

That matters because:

• pricing models are only as good as the inputs
• poor data quality creates compliance risk
• downstream analytics become unreliable

A practical example: if a commercial lines submission includes ambiguous building use, AI can prompt the underwriter (or agent) with a targeted clarification—before the policy is bound.

3) Multi-source connectivity without a six-month integration project

Most insurers already have the ingredients spread across:

• policy admin and claims platforms
• CRM
• document stores
• call transcripts
• third-party enrichment feeds

The value of connectors isn’t convenience—it’s speed to production and a lower maintenance burden for IT.

4) Unstructured data support (where the real signal hides)

In insurance, critical information lives in unstructured formats:

• adjuster notes
• medical reports
• repair invoices
• photos and videos
• emails
• recorded calls and chat logs

If you’re serious about AI in claims automation and fraud detection, you need strong NLP and computer vision capabilities tuned for insurance contexts (terminology, document types, edge cases).

5) Built-in enrichment that’s governed

Third-party data can improve underwriting and claims outcomes, but it also creates governance questions:

• Are we allowed to use this data for this purpose?
• How long do we retain it?
• Can we explain how it influenced a decision?

A curated insurance data catalog and clear controls help data leaders scale enrichment safely—without every team inventing its own approach.

6) Responsible AI and privacy controls as product features

Responsible AI can’t be a slide deck. It has to be operational. In practice, that means:

• role-based access controls
• data minimization and retention logic
• audit trails for model outputs
• monitoring for drift and anomalies
• consistent policy enforcement across channels

For insurers, this is particularly important because underwriting and claims decisions are high-stakes and heavily regulated.

A useful internal standard: “If we can’t explain it, monitor it, and audit it, it doesn’t ship.”

Where AI layers create ROI: underwriting, claims, and customer engagement

The strongest business case for an intelligence layer is that it supports multiple AI-driven processes with one operational backbone.

Underwriting: faster decisions, better risk selection

When prescriptive insights are embedded into underwriting workflows, you typically see gains in:

• submission triage (routing to the right underwriter)
• appetite enforcement (fewer off-strategy binds)
• coverage adequacy (less underinsurance)
• cycle time reduction (fewer back-and-forth emails)

Crucially, this doesn’t require replacing the underwriting workbench. It requires inserting the right prompts, checks, and recommendations in context.

Claims: triage, leakage control, and fraud detection

Claims is where operational friction is most visible. A well-implemented AI layer can support:

• early severity prediction (reserving and routing)
• document understanding (faster coverage verification)
• fraud propensity scoring (SIU prioritization)
• next-best-action guidance (reduce leakage and rework)

The point isn’t “AI replaces adjusters.” It’s “AI makes adjusters consistent, faster, and harder to fool.”

Customer engagement: personalization that doesn’t feel random

Many insurers talk about personalization but deliver inconsistent experiences:

• one message in marketing
• another through agents
• another in self-service
• another in the call center

An intelligence layer can orchestrate recommendations across touchpoints so the customer sees a coherent story—which matters a lot when you’re trying to close the insurance literacy gap and sell protection, not just price.

A practical checklist for data and IT leaders evaluating platforms like Zelros

If you’re considering an insurance AI platform or intelligence layer, I’d use this shortlist. It keeps the evaluation grounded in delivery, not demos.

Workflow fit

• Can recommendations appear inside the tools teams already use (CRM, policy admin, claims consoles)?
• Are next steps specific (questions to ask, actions to take), not just scores?

Governance and compliance

• Are audit logs and access controls native?
• Can you configure data retention and privacy policies per use case?
• Can the platform support consistent decisioning across channels (to reduce conduct risk)?

Data readiness and integration

• How many of your “typical” sources are supported out of the box?
• How are unstructured inputs handled (documents, images, transcripts)?

Time to measurable ROI

• Can you ship one production use case in 8–12 weeks?
• Do you have built-in monitoring and performance tracking so ROI is provable?

If the answer to these is unclear, you’re likely buying complexity.

The bigger picture: AI that helps close the protection gap

The protection gap has been widening for years, and climate-related losses keep exposing the difference between economic loss and insured loss. Technology can help, but only if it changes how people buy and use insurance.

A line I agree with: the real goal is to educate customers away from buying the lowest price and toward buying the right protection. AI can support that at scale—through clearer recommendations, better timing, and consistent messaging across every customer touchpoint.

If you’re leading data or IT in an insurer, your best move in 2026 planning may not be “another platform.” It may be an AI layer that finally makes your existing investments—core systems, data lakes, and models—show up where decisions are made.

If you could add one prescriptive insight into underwriting or claims tomorrow—something teams would actually use—what would it be?