AI for Insurance: Deploying Video Agents for Claims and Policy Conversations

At 2 a.m., a burst pipe sends water across a policyholder's kitchen floor, and the policy that would tell her what is covered is buried in an email folder. She calls her insurer, and how that call goes will shape how she feels about the company for years.

Most AI for insurance works behind the scenes or handles the easy calls. The hard calls still reach adjusters because they depend on presence, the sense that someone is paying attention to what the caller means as well as what she says. Real-time video agents are built for those conversations.

What AI does in insurance today

AI for insurance covers the machine learning, language models, and automation that carriers use across the policy lifecycle. Pricing models are the most mature, while systems that talk to policyholders are the newest.

Underwriting and pricing

Pricing is the most established use. Models score applications against historical loss data and set rates by risk, and WTW's 2026 survey found that close to 80% of property and casualty insurers rely on advanced rating and pricing models.

AI that assists underwriters directly is newer. It can pull together application data and surface the risks worth a closer look, but a person owns the decision on complex cases.

Fraud detection

Fraud is a costly problem, and Deloitte estimates that about 10% of property and casualty claims are fraudulent. Models look for patterns that people miss across thousands of claims, such as repeated losses, inconsistent documents, or timelines that don't add up.

Soft fraud, such as inflating a legitimate claim, accounts for most incidents and is harder to catch than outright fabrication. A flag sends the claim to an investigator instead of denying it, so honest claimants keep moving.

Claims handling

Document extraction and image analysis read estimates, invoices, and photos, which speeds up routine claims. A water-damage claim with clear photos and a standard repair estimate can move quickly, while one with a disputed valuation cannot.

Complex or disputed losses still depend on an adjuster's judgment about coverage and fairness. The claimant often has to explain the situation again when the file changes hands.

Customer conversations

Chatbots and voice bots handle simple requests such as ID cards, payment dates, and claim status, and language models have made written and spoken answers easier to automate.

A first notice of loss or a coverage question is different, because a hesitant description needs room to finish and a scripted answer can cost a policyholder money. These conversations have mostly stayed with people. Real-time video agents are a newer attempt at them, adding the facial and vocal signals that text and voice channels drop while licensed staff keep the judgment calls.

Where insurance conversations break down

Automated channels struggle most when they meet someone who is stressed or unsure. Surveys of claimants and policyholders point to three pain points.

Claimants repeat the same details

First notice of loss is where the claim file is built, and a stressed claimant rarely knows what to say first. A voice bot that answers after a fixed silence cuts into a hesitant description, and a text form can't tell a calm answer about injuries from a shaken one.

In a 2024 Hi Marley survey of 800 claimants, 54% worked with two to three people and had to repeat basic incident details more than once. Among those who told their story only once, 53% were very satisfied.

Claimants chase answers

After intake, the claim turns into a long wait. In J.D. Power's 2026 study of property claims, customers waited an average of 40.7 days for final payment, down 3.4 days from the year before.

The study lists a lack of explanations or the chance to discuss the estimate or settlement, and frequent customer-initiated contacts, among the common issues.

Policyholders don't know what they bought

Policyholders often buy coverage they can't describe. In Hippo's 2026 survey of 1,047 US homeowners, 37% could not confidently explain what their policy covers for weather damage, and 55% didn't know their deductible without looking it up.

Nearly four in ten assumed something was covered and were wrong, and for about one in five of all respondents that cost them money out of pocket. A scripted answer at renewal leaves that misunderstanding in place until a claim exposes it.

What video agents have to get right in insurance conversations

Video adds signals that voice and text drop, such as a tightening expression or a glance at a damaged ceiling. Using them well takes four capabilities working together.

Perception lets the agent notice when a calm answer and a strained face disagree and adjust before the claimant has to say she is struggling. Timing lets a hesitant description finish, and memory and grounding let the agent recall what the claimant said last week and answer coverage questions from the actual policy wording. The fourth is handoffs, which means recognizing which decisions belong to a licensed person.

A claim runs for weeks, so these capabilities have to hold across every call, from the first report to the last follow-up.

Compliance, governance, and where human judgment belongs

The National Association of Insurance Commissioners (NAIC) adopted its NAIC Model Bulletin on December 4, 2023. It says decisions made or supported by AI systems must comply with existing insurance laws, including those on unfair trade practices and claims settlement. Insurers are also expected to document their governance, oversee third-party vendors, and notify consumers.

Regulators are not the only source of pressure, because the insurance market is adding AI exclusions too. Insurance Business reported that state filings showed 2,369 AI exclusion records in force across 49 states by mid-2026, and Berkley applies an absolute AI exclusion across its D&O and E&O products. A startup called Harbor raised $10 million in October 2026 to insure the AI-agent risk those exclusions leave behind.

Together, the bulletin and the exclusions leave accountability with the carrier. The safest design settles who answers in advance by deciding which conversations an agent never closes on its own, and three situations should route to a person on any platform.

• Claim disputes: When a policyholder contests a valuation or settlement offer, the agent records the objection and transfers to an adjuster with the file attached.
• Coverage denials: Denial decisions and denial language stay with licensed adjusters, and any request for a coverage determination routes to them.
• Fraud flags: When the carrier's fraud signals fire, the conversation moves to a human investigator and the agent makes no determination.

Instructions and configuration steer an agent but don't guarantee its behavior, so adjuster review of flagged sessions belongs in the design from the start.

What a PAL for insurance does differently

Assembling perception, timing, memory, and handoffs from separate tools means integrating and maintaining each one on its own. Tavus is the human computing company, building PALs that see, hear, understand, and respond in real-time conversations.

A Personified Application Layer (PAL) is a real-time application a policyholder talks to face to face, one that sees, hears, remembers, and responds. Teams build one on the Conversational Video Interface (CVI), an API that plugs into existing claims and policy apps. Four components run as one closed loop inside it.

• Timing: Sparrow-2 predicts who holds the floor straight from raw audio, so the PAL responds when a human listener would. In preliminary results on TurnBench, it posts 92.4% end-of-turn recall and 97.4% interruption recall.
• Perception: Raven-1 fuses the claimant's emotional and attentional signals.
• Reasoning: The large language model (LLM) layer reasons about what to say and do next, including when to bring in an adjuster.
• Rendering: Phoenix-4.5 renders the PAL's face, head, and upper body in real time.

Beyond the loop, Knowledge Base, Memories, and Function Calling connect the PAL to the carrier's data. Knowledge Base grounds answers in policy documents, Memories carry context from one session to the next, and Function Calling triggers actions in other systems, such as paging an adjuster or pulling a claim status.

Objectives and Guardrails turn the three handoff situations into configuration. Objectives move a conversation toward completion criteria, such as confirming the policy number at intake. Guardrails steer the PAL away from coverage determinations and fire a webhook for the audit log.

How a PAL could handle a claim, step by step

Connected to the carrier's claims system, paging tool, and policy documents, a PAL could handle a single water-damage claim. The first call comes right after the loss, questions about the payment follow, and a renewal conversation arrives months later.

1. First notice. Water is coming through a policyholder's kitchen ceiling. She says "I'm fine" while her voice tightens, so Raven-1 flags the mismatch, the LLM layer asks again whether anyone is hurt, and Sparrow-2 leaves room for her answer.
2. Intake. An Objectives sequence collects her policy number and loss details while Phoenix-4.5 keeps an attentive face on screen.
3. Escalation. When she mentions a sagging ceiling, Function Calling alerts the on-call adjuster through the carrier's paging system and passes along what she has said, so the adjuster starts from there.
4. Settlement. Days later, she asks why her payment is lower than expected. Function Calling retrieves the settlement details from the claims system, Knowledge Base supplies the relevant terms from her policy documents, and Memories bring back her earlier questions, so she gets an explanation without retelling the loss.
5. Renewal. Months later, a higher premium has her weighing a bigger deductible. She says "that sounds fine" while her answers get shorter, Raven-1 catches the hesitation, and an Objective written for teach-back can keep the conversation going until she restates the deductible and what it covers.

The adjuster keeps every coverage decision, and the carrier ends up with a transcript of each conversation for review.

Start with the conversation policyholders remember

At 2 a.m., a policyholder watches water come through her kitchen ceiling and calls her insurer. Years later, what she remembers is whether someone noticed she was shaken, let her finish a hesitant sentence, and didn't make her start over when she called back days later.

Presence is the sense that someone is paying attention to what she means as well as what she says. It has always been what policyholders want from their insurer. Tavus builds PALs so claims and renewal conversations can offer it at any hour, while adjusters keep the judgment calls.

See it for yourself. Book a demo.