Insurance claims rarely arrive as a clean row in a database. A single claim can generate a policy document, a first notice of loss, photographs of the damage, inspection reports, repair estimates, correspondence, invoices, and structured claims data. Each source can contain a different piece of the story.

For insurance providers, the key difficulty lies not in handling larger volumes of data, but in linking relevant insights at the precise moment they are needed, a domain where multimodal AI offers compelling value.

Traditional enterprise AI and retrieval-augmented generation, or RAG, have made it easier to search and reason over large volumes of text. But insurance claims often depend on information that can’t be fully understood as text alone.

• A photograph can show the extent of property damage.
• A scanned form can contain information in a table.
• A repair estimate can combine text, numbers, and layout.
• A policy document defines the coverage that gives those pieces meaning.

Google Cloud’s Gemini Embedding 2 maps text, images, documents, audio, and video into a unified semantic space, enabling multimodal retrieval across content types. As of April 22, 2026, Gemini Embedding 2 is generally available on Google Cloud.

For insurers, the opportunity goes beyond automated image analysis. The real opportunity is claims intelligence. That means building AI systems that can connect policy, claims, evidence, and operational workflows, and help people make better, faster decisions.

A Claim is a Multimodal Story

Consider a property claim after a water leak. The insurer might have:

• A policy document describing coverage and exclusions
• A first notice of loss submitted through an app
• Photographs showing damaged rooms and equipment
• An inspection report
• A contractor’s repair estimate
• Previous correspondence with the policyholder
• Structured claims and policy data
• Invoices and receipts

Each source answers a different question:

• The policy says what may be covered.
• The photographs show what happened to the property.
• The inspection report provides an assessment.
• The repair estimate describes the expected cost.
• The claims system contains structured information such as policy number, claim status, and dates.

The problem is that these sources usually live in different systems and formats. A human claims adjuster can connect them because they can look at the whole case. Traditional software tends to see individual files and records. Multimodal AI enables AI systems to gain a richer view of the same claim.

Why Text-Only RAG Isn’t Enough for Every Claim

Text-based RAG remains useful for insurance. An AI system can retrieve relevant clauses from a policy, search claims notes, and find information in reports. But reducing every source to extracted text can remove context that matters.

Take a repair estimate. A text extraction system might capture: Roof repair, €8,500. But the document could include several line items, quantities, material descriptions, and notes explaining how the numbers were calculated. Or consider a photograph of a damaged vehicle. The image provides evidence of damage that isn’t described in the claim description.

The value of multimodal AI lies in connecting these sources rather than treating each as an isolated input.

From Image Recognition to Claims Intelligence

Computer vision has already created opportunities for analyzing insurance images.

An insurer can use AI to examine damage photographs and identify visual characteristics of a loss. But image analysis on its own doesn’t answer the broader claims question.

Suppose an image appears to show significant water damage. The important questions could include:

• Is this type of damage covered by the policy?
• Has the same property reported a related loss before?
• Does the inspection report support the visual evidence?
• Does the repair estimate appear consistent with the damage?
• Are there exclusions or conditions in the policy that need to be considered?
• Is there missing documentation?

The image is only one piece of the case. A multimodal claims system can potentially connect the image with the rest of the evidence.

Google Cloud’s insurance capabilities support multimodal document ingestion for claims evidence, such as photos, videos, and voice recordings, as well as text submissions and other documents. Google also highlights applications across claims analysis, damage assessment, and fraud analytics.

That leads to a more useful framing: The future of claims AI isn’t just about seeing damage. It’s about understanding the claim.

1. Connect the Policy to the Evidence

One of the most valuable applications is grounding claims analysis in the policy itself.

Imagine a claims handler reviewing a water damage case. Instead of manually switching among the claims system, the policy document, and uploaded evidence, an AI assistant could consolidate the relevant information.

It could surface:

• The applicable policy section
• Relevant coverage conditions
• Potential exclusions
• Key facts from the claim submission
• Relevant observations from inspection documents
• Supporting photographs
• Missing evidence that may need review

The AI shouldn’t silently decide that a claim is covered. The stronger pattern is to provide a grounded analysis that allows the claims professional to review the evidence and make the appropriate decision. That distinction becomes important when AI enters regulated and high-impact workflows.

2. Reconcile Damage Photographs with Inspection Reports

Claims evidence can contain inconsistencies. A claimant may describe limited damage. An inspection report may identify more extensive damage. Photographs may provide additional visual evidence.

Today, resolving those differences can require manual review. A multimodal system could compare the sources and flag potential discrepancies.

For example:

• Claim description: Water damage limited to the kitchen
• Inspection report: Damage identified in kitchen and adjoining hallway
• Photographs: Visible signs of water damage in both areas

The AI isn’t required to determine which source is correct. It can identify that the sources don’t fully agree and direct the case to human review.

Google Cloud describes a similar cross-document pattern for multimodal enterprise agents, where systems can compare information across different formats and flag contradictions or outdated information. For insurers, this could become an important quality-control mechanism.

3. Compare Repair Estimates with Visual Evidence

Repair estimates can vary significantly in format and detail. One contractor may submit a spreadsheet. Another might provide a PDF with tables and photographs. A third could send an invoice with scanned documentation. A claims AI system could potentially bring these sources together with the visual evidence.

Suppose photographs show damage to one area of a property, while a repair estimate includes extensive work in another area. That discrepancy doesn’t automatically indicate a problem. But it may be worth reviewing.

An AI agent could flag: “The repair estimate includes replacement of three units. The submitted photographs appear to show damage to two units. Review recommended.” That’s a very different workflow from simply asking an AI to “analyze this photograph.” The AI is helping reconcile evidence.

4. Identify Missing Evidence Earlier

Claims often progress through stages.

1. A claim is submitted.
2. Additional documents are requested.
3. An inspection occurs.
4. An estimate is received.
5. An adjuster reviews the case.
6. A decision is made.

Delays can occur when the required evidence isn’t complete. A multimodal AI assistant could help identify missing information based on the claim type and the evidence already available.

For example: “The claim includes the policy, claimant statement, and three damage photographs. The repair estimate and inspection report have not yet been received.” This kind of capability can help claims teams prioritize incomplete cases and reduce avoidable back-and-forth.

Google Cloud’s insurance claims architecture connects document ingestion, classification, extraction, damage analysis, fraud detection, and claims decision support as parts of a broader claims workflow. The larger idea is to treat evidence collection and claim analysis as a single, connected process.

5. Give Claims Adjusters an AI Research Assistant

Claims professionals spend substantial time gathering context. They may need to review policy conditions, previous correspondence, inspection reports, and supporting documents before making a decision. An AI research assistant can help compress that work.

Instead of asking: “Summarize this PDF.”, a claims adjuster could ask: “What evidence supports the reported cause of loss, what information conflicts with the claim, and which policy conditions appear relevant?” The system could retrieve evidence from multiple sources and present an answer grounded in the underlying claim records.

This is where multimodal embeddings become especially useful. Google Cloud’s Gemini Embedding 2 supports cross-modal retrieval and heterogeneous formats, enabling integration of information from documents, images, tables, and other content into a shared semantic representation.

The goal isn’t to replace the adjuster; it’s to reduce the amount of time the adjuster spends assembling the case.

6. Improve Fraud Investigation with More Connected Evidence

Fraud detection is another area where connecting multiple sources can matter.

A suspicious claim rarely reveals itself in a single field. Potential signals may exist across claim history, submitted documents, photographs, correspondence, repair estimates, and structured policy data. A multimodal system can help bring those signals together.

For example, an AI workflow could identify a combination of:

• Repeated claims involving similar damage
• Unusual similarities between submitted images
• Differences between reported circumstances and inspection evidence
• Inconsistencies between documents
• Unexpected repair costs
• Conflicting dates or descriptions

Multimodal ingestion is an input to advanced fraud analytics in insurance, alongside claims and underwriting workflows. The important role for AI here is investigation support. The system can surface patterns and explain why a claim deserves further review. Final decisions can remain with appropriately authorized teams.

7. Connect Claims with the Wider Insurance Data Platform

Multimodal AI becomes much more valuable when it isn’t isolated from the rest of the insurance data environment.

Consider the broader set of information an insurer already has:

• Policy data: Coverage, limits, conditions, dates and customer information.
• Claims data: Claims history, status, severity, payments and outcomes.
• Unstructured content: Forms, correspondence, inspection reports, invoices and estimates.
• Visual evidence: Photographs, scanned documents, videos and other media.
• External data: Weather, geospatial information, property information and other relevant datasets.

Google Cloud’s insurance reference architecture describes how data and AI capabilities can connect claims, underwriting and other insurance functions, including analytics, damage assessment and fraud monitoring.

This creates an important architectural principle: Multimodal AI shouldn’t become another data silo. Rather than creating an isolated system, it must be deployed on a framework that enables authorized AI tools to analyze and derive insights from existing organizational data.

Building a Multimodal Claims Architecture

For insurers, the architecture behind the AI matters as much as the model. A practical foundation could include several layers:

• Claims and policy systems: Policy administration, claims management, CRM, payments and other core insurance applications remain the systems of record.
• Multimodal content layer: Images, PDFs, scanned forms, inspection reports, estimates, correspondence and other evidence need to be stored, indexed and governed.
• Data platform: A governed data platform can bring together structured claims, policy, and operational data for analytics and AI applications. Google Cloud’s insurance reference architecture uses technologies such as BigQuery and BigLake as part of the data foundation for insurance workloads.
• Semantic retrieval: Multimodal embeddings can create a shared semantic representation across text, images, document pages, tables and other supported content, enabling cross-modal retrieval.
• AI and agent layer: Gemini-powered applications and agents can support research, summarisation, evidence comparison, claims analysis and other selected workflows.
• Governance and security: Identity, permissions, auditability, data lineage, monitoring and human approval need to be part of the architecture.

That last layer deserves particular attention.

Claims AI Needs Permission-Aware Context

A claims agent could have access to highly sensitive information. That doesn’t mean every agent should see everything.

• An internal claims assistant may need access to policy documents and evidence of claims.
• A customer-facing assistant may need access to a much smaller subset of information.
• A fraud investigation workflow may again require different permissions.

The AI layer needs to respect the same boundaries as the underlying insurance systems.

Gemini Enterprise Agent Platform provides capabilities for building, scaling, governing, and securing enterprise agents, including agent identity, IAM-based access controls, policy enforcement, and auditing. For insurers, this means agent identity, access control, and auditability should be treated as core architecture, not an afterthought.

Human Approval Still Matters

Claims contain decisions with financial and customer consequences. That makes the division of responsibilities important. An AI agent might be well suited to:

Action / Task	Automation Level
Retrieve policy information	Automatic
Classify submitted documents	Automatic
Summarise claim evidence	Automatic
Compare photographs with inspection reports	Automatic, with confidence thresholds and review where needed
Flag inconsistencies	Automatic
Recommend a next step	Automatic, with human review
Recommend a settlement amount	Human review / Controlled workflow
Approve a high-value claim	Explicit human approval

The exact boundaries will vary by insurer, product, and regulatory environment. The principle is straightforward: as the business impact of an action increases, the control around that action should increase as well.

From Claims Processing to Claims Intelligence

The biggest opportunity isn’t automating one step of the claims journey. It’s connecting the journey.

Today, an insurer might have one system for policy information, another for claims, another for documents, another for analytics, and another for customer communication.

Each system serves a purpose. The challenge comes when an employee needs to understand the whole claim. Multimodal AI can become the intelligence layer that connects those sources.

A claim could move from:

Submission

→ Evidence collection

→ Evidence analysis

→ Policy matching

→ Discrepancy detection

→ Fraud screening

→ Repair assessment

→ Decision support

→ Settlement

→ Continuous learning

The AI doesn’t have to control every step. It can provide context and intelligence where human teams need it most.

The Claims File Could Become an Active Knowledge Base

For decades, a claims file has been primarily a collection of records and documents. The next generation could make that file much more dynamic.

Imagine opening a claim and seeing an AI-generated evidence map:

• Policy: Relevant clauses and exclusions
• Claim: Reported cause, dates and circumstances
• Visual evidence: Damage photographs grouped by affected area
• Inspection: Key findings and supporting evidence
• Financials: Repair estimates, invoices and expected costs
• History: Related claims and previous correspondence
• Exceptions: Potential inconsistencies requiring review
• Missing information: Documents or evidence still needed

That is more than summarization. It’s a connected representation of the claim that creates a foundation on which agents can perform useful work.

A Practical Path for Insurers

The technology may be new, but insurers don’t need to rebuild their entire architecture to start. A pragmatic approach is to choose one claims process with a clear business problem. For instance, insurers could start with property damage claims where photographs, policy documents, inspection reports and repair estimates all play a role.

1. Start with the evidence: Identify the documents, images and structured records involved in the process.
2. Identify the authoritative sources: Determine which system to trust when sources disagree.
3. Create a governed retrieval layer: Make relevant content searchable across modalities while preserving permissions.
4. Add AI-assisted analysis: Start with summarization, classification, comparison, and evidence retrieval.
5. Introduce agents selectively: Once the retrieval and reasoning layer is reliable, connect it to workflow actions that have clear controls.
6. Measure the outcome: Track review time, processing time, manual effort, exception rates, and other business metrics relevant to the process.

This turns multimodal AI from an experiment into an operational capability.

The Future of Claims AI is About Connecting the Evidence

Insurance has never lacked information. It has lacked a simple way to connect all the information surrounding a claim.

• A policy document explains coverage.
• A photograph shows evidence.
• An inspection report adds professional assessment.
• A repair estimate quantifies the work.
• A claims database provides structured context.
• A customer conversation adds another layer of detail.

Each source matters. The real opportunity comes from making them understandable together.

Modern multimodal AI capabilities allow insurance platforms to synthesize diverse claims evidence, including unstructured documents, photos, videos, and voice recordings alongside structured data to streamline analysis, enhance fraud detection, and automate operational workflows with robust governance.

For insurers, that points toward a bigger shift. The future of claims AI isn’t simply a faster way to process documents or a better way to analyze damage photos. It’s an intelligence layer that can connect the entire claims story.

How Kartaca Can Help

For insurance organizations, adopting multimodal AI is only part of the challenge. The bigger opportunity is connecting AI with the policy data, claims systems, documents, evidence, and workflows that already power the business.

Kartaca helps organizations build cloud, data, AI, and analytics architectures that turn fragmented enterprise information into a foundation for intelligent applications and agentic workflows.

From data platform modernization and BigQuery architectures to AI assistants, multimodal experiences and governed AI solutions, Kartaca can help insurers create the data foundation and workflow architecture needed to put claims intelligence into practice.

The future claims platform won’t just store the evidence. It will understand how the evidence connects.

Contact us today to start building your data and AI foundation.