Agentic AI in Insurance: What VivaTech Proves

A lot of “AI in insurance” demos still look like smart search bars. Useful, sure—but not the point.

At VivaTech 2025 in Paris, the more interesting shift is agentic AI: systems that don’t just answer questions, but carry out multi-step work across underwriting, claims, fraud, and customer operations. Zelros’ VivaTech presence (hosted in a dedicated area inside the Capgemini booth during the June 11–14 event) is a good signal of where the market is heading: insurers and banks want AI that can act, not just chat.

This matters beyond insurance, too. In our AI in Payments & Fintech Infrastructure series, we keep coming back to the same reality: the highest ROI AI projects are the ones that reduce operational friction in regulated workflows—especially where identity, documents, risk decisions, and money movement intersect.

Why agentic AI is showing up everywhere (including insurance)

Agentic AI is rising because the work is procedural, high-volume, and time-sensitive. Insurance operations are packed with “do this, then that” processes: gather documents, validate identity, read policy wording, check coverage, assess risk, decide next best action, message the customer, log compliance notes.

Traditional automation (RPA, scripts, workflow engines) handles rigid steps well, but breaks when inputs vary. Traditional generative AI handles language well, but often stops at “here’s an answer.” Agentic AI bridges the gap by combining:

• Reasoning over context (policy text, claim history, KYC/KYB files)
• Tool use (retrieving data, generating letters, updating systems)
• Guardrails (approved actions, escalation rules, audit logs)

The punchline: agentic AI turns a conversational interface into an execution layer for operations.

The myth most teams still believe

Most companies get this wrong: they buy a chatbot and expect transformation.

The real value comes when AI can:

1. Understand the request (intent + constraints)
2. Collect missing info (documents, clarifications)
3. Perform checks (eligibility, fraud signals, policy rules)
4. Propose a decision (with rationale)
5. Execute the approved steps (updates, notifications, handoffs)

That’s agentic AI in practice—and it’s why events like VivaTech are becoming the proving ground for real workflows, not slides.

What “agentic AI for insurance” looks like in real operations

Agentic AI shows up as a digital coworker embedded in underwriting, claims, and servicing—not as a separate toy app. Here are the three insurance workflows where it’s already easiest to justify.

Underwriting: faster triage, cleaner files, better decisions

Underwriting isn’t one decision. It’s a chain: intake → document review → risk factors → pricing inputs → exceptions → referral.

Agentic AI helps by:

• Extracting structured data from messy submissions (PDFs, emails, forms)
• Flagging missing fields before the file hits an underwriter’s queue
• Summarizing risk drivers in plain language (what changed, what’s unusual)
• Routing cases based on rules (straight-through vs. refer)

If you’re in commercial lines, the impact can be dramatic because submissions are inconsistent and narrative-heavy. If you’re in personal lines, the value often comes from exception handling—the 10–20% of cases that break straight-through processing.

Claims: the shortest path from FNOL to resolution

Claims operations win when they reduce cycle time without increasing leakage. Agentic AI supports that by orchestrating the “middle” of the claim:

• Intake the first notice of loss (FNOL) across channels
• Validate policy status and coverage basics
• Request missing documentation with customer-friendly messaging
• Summarize adjuster notes and next actions
• Create draft communications and settlement letters (approved templates)

Here’s what works: agentic AI that prepares the file so humans decide faster. Claims leaders don’t need AI to “decide” everything. They need AI to remove the dead time between steps.

Fraud detection: move from alerts to investigations

Fraud models have existed for years. The bottleneck is what happens after a flag.

Agentic AI can:

• Compile a case narrative (timeline, inconsistencies, prior claims)
• Pull supporting evidence from internal systems
• Generate investigator checklists and outreach scripts
• Recommend escalation paths (SIU referral vs. monitor)

The best fraud teams treat agentic AI as an investigation accelerator, not a verdict machine.

The payments and fintech infrastructure angle insurers can’t ignore

Insurance workflows increasingly touch fintech infrastructure. Premium collection, refunds, claim payouts, chargebacks, identity verification, and sanctions screening aren’t “adjacent”—they’re embedded.

That’s why VivaTech’s AI theme matters to insurance leaders: the same building blocks powering modern payments operations (identity, routing, fraud signals, messaging, case management) are now being repackaged into insurance operations.

Where agentic AI overlaps with payments operations

If you run payments, these patterns will feel familiar:

• Disputes and chargebacks resemble claims triage: documents, timelines, rules, decisioning
• Transaction monitoring resembles fraud: signals, case creation, investigator workload
• Customer servicing is the same pain: long handle times, inconsistent notes, compliance risk

Agentic AI is becoming the shared “automation layer” across both industries.

One practical takeaway: when you evaluate agentic AI for insurance, ask whether the platform can also support money movement workflows (payout status checks, refund rules, reconciliation notes). If it can’t, you’ll end up with siloed automation.

What to look for in an agentic AI demo (and what to challenge)

A good demo shows controlled autonomy with measurable outcomes. A weak demo shows fluent text.

If you’re meeting vendors at an event like VivaTech (or running your own internal bake-off), use this checklist.

1) Does it handle the messy middle?

Ask to see:

• A claim file with missing documents
• A policy with endorsements and exceptions
• A customer who changes their story mid-process

If the system only works on perfect inputs, it’s not ready.

2) Can it cite what it used?

For regulated industries, “because the model said so” is unacceptable.

Look for:

• Source citations to policy clauses, documents, or knowledge articles
• Decision rationale that maps to your rules
• Versioning (which policy wording, which guideline)

3) How are actions governed?

Agentic AI must have permissioning.

Push on:

• What the AI can do without approval (draft, suggest, classify)
• What requires human approval (customer comms, payouts, denials)
• What is blocked entirely (pricing overrides, irreversible actions)

4) Where’s the audit trail?

If you can’t audit it, you can’t scale it.

Require:

• Full logs of prompts, retrieved sources, and actions taken
• Clear separation of system instructions vs. user inputs
• Ability to reproduce outcomes for QA

5) What’s the integration story?

Agentic AI that can’t connect to core systems becomes a sidecar tool.

Ask specifically about:

• CRM/contact center integration
• Claims management system integration
• Document management and email ingestion
• Secure APIs and role-based access

A 90-day plan to turn “cool demo” into production value

The fastest path is a narrow workflow with clear metrics and strict guardrails. Here’s a practical approach I’ve seen work when teams want results without a year-long transformation program.

Days 0–30: pick one workflow, define success

Choose a use case with:

• High volume
• Repetitive steps
• Clear quality checks

Examples:

• Claims document chase + summarization
• Underwriting submission triage + referral packaging
• Fraud case narrative generation

Define 3–5 metrics:

• Average handle time (AHT)
• Cycle time (FNOL → first adjuster action)
• Reopen rate or rework rate
• Customer satisfaction (CSAT) where applicable
• Compliance QA pass rate

Days 31–60: implement guardrails and QA

Focus on control:

• Approved templates for customer communications
• Escalation rules for edge cases
• Human approval checkpoints
• Red-team testing (prompt injection, hallucinations, data leakage)

Days 61–90: scale to a second adjacent workflow

Once you prove the first use case, expand sideways—don’t jump to an unrelated department.

Good expansions:

• From claims intake to claims servicing
• From underwriting triage to underwriting communications
• From fraud narratives to SIU referral packaging

Scaling works when the data sources and compliance patterns are similar.

A reliable rule: if you can’t explain how the AI’s output will be audited, you’re not ready to deploy it.

Why VivaTech matters for insurance leaders right now

VivaTech isn’t just a tech expo. It’s where enterprise buyers pressure-test what’s real, what’s compliant, and what can survive contact with production systems.

Zelros’ presence at VivaTech 2025, centered on agentic AI for financial services, is a useful cue: this category is moving from “innovation theater” into operational budgets. The most forward-leaning insurers aren’t asking whether AI can write a better email. They’re asking whether AI can close the loop—from customer intent to completed task—without increasing risk.

If you’re responsible for claims, underwriting, fraud, contact center, or digital operations, your next step is simple: treat agentic AI like you’d treat a payments platform or core system change. Demand governance, integrations, measurement, and auditability.

Where do you see the biggest opportunity for agentic AI in your organization—underwriting triage, claims cycle time, or fraud investigation throughput?