Most AI first notice of loss phone intake projects are scoped as a voice problem. They are really a data problem: what the call has to write into your claims system, and what it must never decide.
Vendors sell FNOL automation on handle time. Claims leaders should judge it on how often the record is complete, matched to the right policy and routed to the right person the first time. This guide is for claims operations, IT and contact-centre leaders at carriers and MGAs.
Why first notice of loss still rings the phone
- Most homeowners do not report digitally. In J.D. Power's 2026 U.S. Property Claims Satisfaction Study, 38% of customers reported first notice of loss digitally, 49% submitted photos digitally and 45% received updates digitally (5,093 homeowners who filed a claim in the previous nine months, fielded December 2024–December 2025; J.D. Power, March 2026). The other 62% used another channel, such as a phone call or their agent.
- Ease of contact tracks satisfaction. In the 2025 edition, overall satisfaction was 777 on a 1,000-point scale when customers said it was very easy to communicate with their insurer, against 337 when it was difficult. J.D. Power named “often needing to leave voicemails” as a common failure point (5,178 customers; J.D. Power, March 2025). Cycle times are improving but still long: in the 2026 edition, the average time to final payment was 40.7 days, down 3.4 days on the previous year (J.D. Power, March 2026).
- Auto claims are getting heavier. In J.D. Power's 2026 U.S. Auto Claims Satisfaction Study, 26% of auto claims were total losses, up from 16% four years earlier. J.D. Power links the rise to higher repair costs, people keeping vehicles longer and higher deductibles that keep smaller repairs out of the claims process (8,261 customers, September 2025–August 2026; reported by Insurance Journal, October 2026).
A large share of claims still begin with a call, and the claims behind those calls are getting more complex. A 40-day cycle that starts with a voicemail and a half-empty FNOL record only gets longer.
An FNOL call that ends without a complete, matched record is not automated. It is postponed.
Start with the data contract, not the script
Before anyone writes dialogue, write down what a finished intake call must produce: the contract between the voice agent and your claims system. If a field is in it, the call cannot end without it or a recorded reason why not.
| Field | Source | How it is validated |
|---|---|---|
| Caller identity and relationship to the policy | Caller, checked against the policy record | Same verification factors your human team uses; third-party callers flagged |
| Policy and covered vehicle or property | Policy administration system (read-only) | Exact match on policy number or verified name plus address; no match means human review |
| Date, time and location of loss | Caller | Date within the policy period; address normalised; geocoded if you run catastrophe codes |
| Cause of loss | Caller, mapped to your peril codes | Pick-list mapping in code; “other” goes to a person, the model does not invent a code |
| Description in the caller's words | Caller | Stored verbatim alongside the transcript, never summarised over |
| Injuries, other parties, police or fire report | Caller | Yes/no captured explicitly; any “yes” to injury triggers the injury rule below |
| Immediate needs | Caller | Towing, emergency repairs, somewhere to stay; passed to the team that can act on them |
| Photos and documents | Upload link sent by text or email on file | Link tied to the claim number; upload status visible to the adjuster |
Write one contract per line of business. A single generic form is how intake records end up full of “N/A”.
Which FNOL calls to resolve, assist or keep human
Automate what is structured and answerable from your systems of record; keep judgement, injury and distress with people. Our AI voice agent for insurance guide covers the wider call mix; this is the split inside the claims line.
| Call | AI role | Why |
|---|---|---|
| Property-only auto or home loss, policy matched | Resolve intake | Full data contract captured, FNOL created, claim number and next steps given |
| Glass, towing or emergency-repair request | Resolve intake | Structured, time-sensitive; hand the request to your existing vendor process |
| Policy not found, or caller is a third party | Assist | Capture what is known, then route to a person to match or open a third-party claim |
| Additional information on an existing claim | Assist | Attach to the open claim; never open a duplicate |
| “Am I covered for this?” | Keep human | Coverage is a licensed, situational judgement |
| Any injury, fatality, or commercial or liability loss | Keep human | Legal and medical questions; the agent captures basics and transfers |
| Caller in danger, distressed or vulnerable | Human now | Emergency services first if anyone is at risk, then a trained person |
A reference architecture for AI first notice of loss phone intake
Five layers, each with one job, so each can be audited.
- Telephony and capacity. Sits in front of your existing phone system, not instead of it. It must hold your worst catastrophe hour and fall back to callbacks, not hold.
- Identity and policy match. Read-only lookups against the policy administration system. The model never “finds” a policy by guessing; a lookup returns one match or the call goes to a person.
- Dialogue and structured capture. The model handles the conversation; a schema handles the data. Every value is validated against the data contract, and key facts are read back to the caller.
- Rules engine. Deterministic code that runs on every turn: injury and danger routing, the coverage-question guard, duplicate-claim detection and catastrophe mode. More on this below.
- Claims system write, handoff and audit. One idempotent write that creates the FNOL (so a dropped call and a redial do not create two claims), a dated acknowledgement record, the transcript and recording attached to the claim, and a warm transfer that passes everything captured so far to the person who picks up.
Start the write layer in draft mode: the agent creates the record, a claims handler approves it. Move to direct creation per loss type once drafts stop needing correction. This “know what you don't know” behaviour is the subject of Confidently Wrong: if a behaviour matters, it lives in code, not in the prompt.
The rules that belong in code
Acknowledgement deadlines matter too. California's fair claims regulations, for example, require an insurer to acknowledge a notice of claim immediately and in no event more than 15 calendar days later, and to record the date in the claim file if the acknowledgement is not in writing (Cal. Code Regs. tit. 10, § 2695.5(e)). An intake agent that writes a timestamped acknowledgement to the claim file makes that easy to evidence. Rules differ by state; map yours with counsel. This is not legal advice.
Catastrophe night: an illustrative capacity model
The numbers below are assumptions to show the method, not a benchmark. Replace them with your own catastrophe history.
| Assumption | Value |
|---|---|
| FNOL calls in the first 3 hours after a hailstorm | 1,200 |
| Intake staff on shift | 40 |
| Human FNOL handle time | 15 minutes → 4 calls per person per hour |
| Human-only capacity in 3 hours | 480 calls; about 720 callers waiting, abandoning or redialling |
| AI resolves intake for 60% of calls (property-only, policy matched) | 720 handled by the agent; 480 reach people, which fits the capacity |
| Draft review load if every AI record is reviewed | 720 records; plan who reviews them and by when |
Two caveats. The 60% only holds if the policy match works: measure how often the lookup returns exactly one policy before trusting any containment figure, ours included. And the review queue is real work; if nobody owns it, the backlog has just moved from the phone to the claims desk. For what unanswered calls cost on a normal day, see our analysis of call center overflow cost.
Governance: where the NAIC bulletin lands
The NAIC model bulletin on insurers' use of AI systems, adopted on December 4, 2023, is a template that each state may adopt in its own form. It expects insurers to maintain a written AI program that addresses AI systems “across the insurance life cycle”, explicitly including claim administration and payment, whether the system was built in-house or by a third party. For third-party systems, it points to due diligence and, where appropriate, contract terms that give the insurer audit rights and require the vendor to cooperate with regulatory inquiries (NAIC Model Bulletin).
For an intake agent, that means evidence you can produce on request: the data contract, the rules and their test set, transcripts with actions taken, the review data behind each move to direct creation, and the vendor's audit terms. Put them in the contract, not after go-live.
A rollout that starts where the risk is lowest
- Weeks 1–2: call-flow audit. Pull a month of claims-line recordings and your last catastrophe event. Measure the property-only, policy-matched share.
- Weeks 2–4: contract and integrations. Data contracts, read-only policy lookups, draft FNOL creation, transfer targets and the rules test set.
- Weeks 4–6: after-hours and overflow first. Go live on calls your team misses today, every draft reviewed. Then extend to daytime overflow.
- Before the season: catastrophe drill. Load-test at your worst historical hour, with the policy system deliberately slowed.
Track five numbers: data-contract completeness, draft correction rate, injury-transfer accuracy (from transcript review), duplicate claims created, and answer rate during events.
Vendor checklist
- Show us your data contract for auto and for homeowners. Which fields can the agent leave empty, and why?
- Is injury and coverage routing a rule outside the model? Show the rule and the test recordings.
- What does the agent do when the policy lookup returns zero or two matches?
- How do you prevent a duplicate claim when the caller redials after a dropped call?
- What is written to our claims file, with what timestamp?
- Will you sign audit and regulatory-cooperation terms, and is our call data used to train models?
For reference, Unicall's managed service starts from $1,000 a month, billed on ticket volume rather than agent seats, with implementation quoted after a free call-flow audit (pricing). A national insurance carrier quoted on that page eliminated its overflow vendor and reported cost per ticket down 52% in the first quarter. Across our live deployments, about 60% of calls are resolved autonomously, with sample sizes and periods varying by client. Neither is an FNOL benchmark. Deployments typically go live in 4–6 weeks without replacing the existing phone system, and call data is never used to train models. To compare voice platforms on public information, see our comparison hub.
FAQ
Can an AI voice agent take a first notice of loss by phone?
Yes, for the structured part: verifying the caller, matching the policy, capturing the date, place and cause of loss, and creating the FNOL record in your claims system. Coverage, liability, injury claims and disputes stay with licensed staff and adjusters.
Should an AI agent answer coverage questions during FNOL?
No. A question such as “am I covered for this?” should be recognised and passed to a licensed person or adjuster, with the intake data already attached. That boundary belongs in code, not in a prompt.
What happens when the caller reports an injury?
If anyone is in danger, the agent tells the caller to contact emergency services first. Any reported injury then routes the call to a person, with the details already captured passed along. Injury claims carry legal and medical questions that an intake agent should not handle.
How does AI FNOL intake help during a catastrophe?
After a hailstorm or flood, claim calls arrive within hours. An AI agent answers every call, takes the property-only claims that fit your rules, offers callbacks instead of hold, and keeps adjusters for injury, complex and vulnerable-customer calls. Load-test it at catastrophe volume before the season.
Does the NAIC AI model bulletin apply to AI claims intake?
In states that have adopted it, the NAIC model bulletin expects insurers to run a written AI program covering AI systems used across the insurance life cycle, which it says includes claim administration and payment, whether the system is built in-house or bought from a vendor. Check your states' versions with counsel. This is not legal advice.
Unicall.ai builds and operates done-for-you AI voice agents that answer every call 24/7, take real actions in your systems, and hand over to your people, with full context, the moment they should.
