AI Voice Agents for FNOL Intake in Insurance Claims Handling

Key Takeaways

• Most first notices of loss still arrive by phone, not through an app.
• Caller tone is a live signal, and most stacks only score it afterwards.
• Claims audio holds medical detail worth keeping inside your own boundary.

The first call after a loss is the densest personal-data moment in a claims file, and most of the industry has been tuning it for speed. We build Dograh, an open-source voice agent platform, and insurance teams ask about this one call more than any other workflow.

An AI voice agent for FNOL intake answers the first loss call, verifies the policy, captures what happened, creates the claim record, and routes the file to an adjuster. The hard part is not the workflow. It is that one short call collects medical detail alongside financial detail from someone who is upset.

What actually happens on an FNOL call

FNOL means first notice of loss, the moment a policyholder tells the carrier that something went wrong. The call has a fixed job. The agent confirms who is calling and which policy applies, takes a description of the incident, opens a claim record that returns a claim number, and hands the file to the right adjuster queue.

Look at what that collects. A vehicle claim captures injuries. A property claim captures what a family lost. Most calls touch payment details somewhere. That is medical information sitting next to financial information in one recording, gathered from someone who may still be standing at the scene.

The volume argument gets made badly, so here is the honest version. According to J.D. Power's 2026 property claims study, 38 percent of customers reported their first notice of loss digitally. The other 62 percent did not. Digital intake is growing and it is genuinely better for the people who use it. The majority of first notices still arrive as a voice call.

That same study put the average time from first notice to final payment at just over 40 days, about three and a half days faster than the year before. Cycle times are improving. The intake call is still the front door.

Why this call needs a slower agent

Callers describe a loss while they are still shaken, and an agent tuned for throughput makes that worse.

The person on the line may be at a roadside with traffic going past. They may be calling from a house that flooded overnight. They are often unsure which policy covers what happened, and they are holding a phone in one hand while dealing with everything else.

That state changes what a good agent sounds like. Pace has to drop. Questions have to arrive one at a time. A caller forced to repeat a hard detail because the agent missed it will not file that under efficiency.

Empathy here is not a voice setting. It is a sequencing decision about what the agent asks and when it stops asking. We think emotional intelligence matters more on this call than on almost any other workflow we support. A collections call can be brisk and still land well. A loss call needs the agent to register that something bad happened before it starts requesting policy numbers.

It is the same design problem we wrote about in emergency and incident intake, where the caller's state shapes the script more than the workflow does.

That has a technical consequence most deployments miss. Tone should be a runtime signal on an FNOL call, and in most stacks it is only a report-card line. The common pattern is to record the call, transcribe it, then score empathy afterwards inside a quality rubric. That is useful for training. It does nothing for the caller who was on the line this morning. By the time a rubric flags that the agent pushed through a distressed caller, the claim is open and the impression is set.

Vocal tone carries information the transcript drops. Pitch and pace and the pauses between words are the part of a sentence that tells you whether someone is holding together. A transcript of "I'm fine, go ahead" reads the same whether the person meant it or not.

So FNOL is one of the strongest candidates we know of for real-time vocal tone and emotion detection. It belongs in the runtime path rather than on an analytics dashboard. Detected distress can slow the pace, drop optional questions, disable any cross-sell logic outright, and put a human on the line earlier than the routing rules would.

Callers are increasingly willing to let AI into this moment when it clearly helps them. Insurity's 2026 consumer survey found 51 percent would feel confident filing a severe weather claim if AI helped validate the loss, up from 38 percent in 2025. That is a thirteen point move in a year. Confidence is being extended, and it can be spent badly.

Open Source Alternative to Vapi / Retell

Self-hosted voice agent platform — no per-minute fees

dograh-hq/dograh

Star on GitHub

Where the recording actually lives

Every FNOL platform conversation eventually reaches the question nobody enjoys answering, which is where the audio physically sits.

Insurance breaches are not abstract. AssuranceAmerica was breached in March 2026 and later disclosed that nearly seven million drivers were affected. The exposed categories were listed plainly, and they included names, contact information, policy and account information, driver and vehicle information, claims-related information, and driver's license numbers.

"Claims-related information" is the FNOL call. That is the artifact this whole workflow produces.

When the platform runs inside your own boundary, the audio and its transcript never transit a vendor you do not control. That is a narrower claim than it sounds, and it deserves saying honestly. Self-hosting does not make you compliant. It removes a party from the blast radius.

The version of this argument that holds up is the one we made for on-prem voice AI, which is to decide first whether the obligation you carry can be contractually delegated. Where it can, a vendor contract is a real option. Where it cannot, or where you simply do not want claims audio in a third party's storage, running the stack yourself is the arrangement that actually changes the exposure.

Dograh is open source and self-hostable, so the audio, along with the transcript and the models that process it, can sit inside your infrastructure. You can bring your own keys for commercial speech and language models, or run open models locally so nothing leaves at all.

Recording consent is a design constraint

Consent rules decide when your recorder is allowed to start, which makes them an architecture question rather than a disclaimer question.

California is the clearest text to quote directly. California Penal Code section 632 subsection (a) reads: "A person who, intentionally and without the consent of all parties to a confidential communication, uses an electronic amplifying or recording device to eavesdrop upon or record the confidential communication, whether the communication is carried on among the parties in the presence of one another or by means of a telegraph, telephone, or other device, except a radio, shall be punished by a fine not exceeding two thousand five hundred dollars ($2,500) per violation, or imprisonment in a county jail not exceeding one year, or in the state prison, or by both that fine and imprisonment."

Several other states also require every party to agree before a call is recorded. The details vary enough that they are worth checking one statute at a time rather than trusting a summary list.

Generally, that means the disclosure has to fire before captured audio reaches storage, not merely before a human reviews it. In most cases it also means the agent needs a defensible record of when consent was taken on each call. We write these lines as fixed utterances rather than generated ones, for the reason we described in intake under a confidentiality duty. A generative system will eventually paraphrase a sentence you needed said exactly.

None of this is legal advice, and positions differ by state and by line of business. Confirm your own with counsel.

Finding the sensitive calls again

The problem nobody scopes for is retrieval, which is knowing months later which recordings held the sensitive material.

A carrier taking fifty thousand loss calls a quarter has fifty thousand recordings. Some fraction contain injury descriptions. A smaller fraction contain a caller distressed enough that the interaction deserves a second look. When a regulator asks, or a complaint arrives, "we have all the audio" is not a useful answer.

Dograh now supports attaching tags to transcription results. A call can be tagged at the point of transcription for whatever matters to you, whether that is medical content, detected distress, a consent edge case, or a moment where the agent should have handed off. Those tags make the sensitive set findable afterwards, which is what turns a retention policy into something you can operate.

It also stops quality review being random sampling. Rather than pulling a hundred calls at random, you pull the calls that were flagged while they were happening. We use the same pattern for fraud-alert callbacks, where the calls needing a second look are exactly the ones a random sample misses.

What the platform roundups leave out

Published comparisons of FNOL platforms tend to rank the same surface, which is resolution rate, routing accuracy, integration coverage, audit trails. Those matter. They are not the questions that decide whether a deployment is safe.

Four questions rarely get answered. The first is where the audio physically lives and whether you can name the jurisdiction. The second is whether you can run the models yourself or are renting inference from a provider that will not show you the weights. The third is what comes with you if you leave in eighteen months. The fourth is who inside the vendor can read a transcript.

Open source answers all four by construction, which is why we build Dograh the way we do. The code is inspectable and the deployment is yours, so leaving is a migration rather than a negotiation.

Appetite for this is moving. The same Insurity survey found 42 percent believe AI could help insurers process claims more efficiently after severe weather events, up from 28 percent in 2025. The carriers that earn that confidence will be the ones who treated the intake call as a data-handling decision first.

If you are scoping FNOL automation this year, start with the call rather than the funnel. Decide where the audio lives before you decide which voice you like. Then make tone something your agent can act on while the caller is still talking, because that is the part a transcript will never give back.