An AI agent can run insurance claims intake end to end: take the first notice of loss on chat, voice, email or SMS, verify the policyholder, collect and read the documents, validate the file, triage it, route it to the right adjuster, and keep the claimant informed through to payout. It hands over a complete file. It does not decide coverage. The gap it closes is large: J.D. Power's 2025 property claims study puts the average time from first notice of loss to final payment at more than 44 days, and Accenture reports that roughly 40% of inbound claims calls are basic status checks. Intake and communication are where an agent earns its keep. The determination stays human.
Key takeaways
One agent, one claim record: the claimant tells the story once, and every stage from FNOL to payout communication reads and writes the same file.
Decision rule: the agent may capture, verify, extract, validate, classify, route and inform. It may not decide coverage, liability or payout. Write that boundary as a guardrail before the happy path, and test it before launch.
Slow intake is measurable: property claims now take 32.4 days from filing to finished repairs and more than 44 days to final payment, the longest since J.D. Power began tracking in 2008.
Communication is the lever you control: J.D. Power scores property claim satisfaction at 777 when it is very easy to reach the insurer and 337 when it is difficult.
Regulators expect a written AI program: the NAIC model bulletin, adopted December 2023, asks for governance, risk controls and internal audit over any AI that supports regulated decisions, including vendor-built systems.
What is insurance claims intake automation?
Insurance claims intake automation is the use of an AI agent to run the front end of a claim without a human driving the conversation: capturing the loss, verifying identity and policy status, requesting and reading evidence, extracting structured data, checking completeness and consistency, classifying severity, routing to the right queue, and sending status updates until the claim closes. The output is a validated, documented claim file in your core system, ready for an adjuster.
Three terms matter here. Claims intake is everything from the first report to a file ready for adjudication. FNOL (first notice of loss) is the first stage of intake: the initial report of the incident. It is one step of seven, and a dedicated guide to AI for FNOL covers it on its own. Straight-through processing is a claim that moves to its next stage with no manual handling because it met the rules for completeness and low complexity. Intake automation raises the share of claims that qualify. It does not decide the ones that do not.
That is the one boundary worth stating plainly, once. The agent never forms an opinion about whether the claim is covered. It gathers what the coverage rules need, runs the deterministic steps, and hands the file to a person for the decision. The rest of this page is about how much useful work sits on the safe side of that line.
How does an AI agent handle the claims lifecycle from FNOL to payout communication?
The agent owns seven stages, each with a clear split between what it does and what it leaves to a human.
Stage | What the agent does | What stays with a human |
|---|---|---|
1. FNOL capture | Takes the report on any channel, asks line-specific follow-ups, checks for injuries first, opens the claim record | Injury, fatality or active emergency: a person, immediately |
2. Identity and policy lookup | Verifies the claimant, confirms active status, effective dates and named insureds | Interpreting what the policy covers |
3. Documents and extraction | Requests the right evidence per claim type, accepts uploads on every channel, reads estimates, reports and invoices into structured fields | Judging severity, valuation or medical necessity |
4. Validation | Checks required fields and documents, date logic, loss date inside the policy period, contradictions | Resolving a contradiction that looks like misrepresentation |
5. Fraud signals and triage | Applies rules: dollar thresholds, party count, duplicate claims, timeline gaps, prior-claim patterns; assigns a severity band | Any accusation, SIU investigation or liability call |
6. Routing and handoff | Packages a structured summary, extracted values with source documents, and the reason for routing; assigns the queue | Approve, deny, reserve, settle |
7. Status and payout communication | Confirms the claim number, sets expectations, sends updates when an adjuster is assigned, a document is needed, an inspection is booked or a payment is issued | Explaining the decision: amounts, coverage, fault |
Stage 3 is where most of the manual time hides and where errors propagate. The agent reads a repair estimate into line items and a total, a police report into a report number and parties, an explanation of benefits into dates of service and codes, and writes each value to the claim record with a pointer to its source document. Where confidence is low or two documents disagree, the field routes for review rather than being written as a guess.
Stage 5 is a routing decision, not a coverage decision. McKinsey's 2025 report on AI in insurance describes Aviva improving the accuracy of routing claims to the appropriate teams by 30 percent and cutting liability assessment time on complex cases by 23 days after rebuilding its claims journey around AI models. Better routing is the mechanism behind faster cycle times.
Stage 7 is the one most intake projects skip and the one claimants remember. The agent should answer "where is my claim" on any channel at any hour and push consent-first updates when the file moves. McKinsey's Claims 2030 work set the target as answering 100 percent of claims status questions digitally, with a human available when the customer prefers one. When a payment is issued, the agent confirms the payment and the timing. The adjuster's decision letter explains why.
Which core systems does a claims intake agent connect to?
Three layers, with scoped read and write access to each. The first is the claims core: Guidewire ClaimCenter, Duck Creek Claims, Majesco, Sapiens, Insurity, or an in-house system. The agent should create the claim, attach documents and write extracted fields through the system's APIs, not a screen-scraping layer that breaks on the next release. The second is policy administration, often the same vendor's PolicyCenter or Duck Creek Policy, read to verify status and named insureds. The third is the engagement layer: a CRM or service desk such as Salesforce, Zendesk or Intercom, and telephony such as Talkdesk, Amazon Connect or Twilio, which carry the conversation and the customer history.
Four integration rules keep this safe. Give the agent least-privilege tools with one job each ("create claim", "attach document", "read policy status"). Make claim creation idempotent, so a claimant who reports twice on two channels produces one record. Log every tool call with inputs and outputs; that log is your audit trail. Keep deterministic steps deterministic: a policy-period or duplicate-claim check should be code, not a model's judgement. See how to safely let AI take actions in backend systems and the current integrations.
What do the NAIC AI model bulletin and state claims-handling rules require?
The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted December 4, 2023, is the document your compliance team will hold an intake agent against. It expects every insurer to maintain a written program (an "AIS Program") for the responsible use of AI systems that make or support decisions in regulated insurance practices. That program should address governance, risk management controls and internal audit functions, and it applies whether the AI system was developed by the insurer or a third-party vendor. The bulletin also tells insurers what documentation a department may request during an investigation or examination.
Three consequences follow. First, "the vendor built it" is not a defence: your AIS Program has to cover the intake agent, its guardrails and its testing. Second, the bulletin is explicit that AI-supported decisions must still comply with unfair trade practice and unfair claims settlement practice laws, so every acknowledgement, document request and status update the agent sends is a timestamp a state examiner can check against your jurisdiction's claims-handling timelines. Third, the documentation the bulletin describes is what a well-built platform produces anyway: simulation results, guardrail configurations, transcripts, and a record of every tool call. Ask to see those artefacts before you sign, and use the practitioner's guide to AI compliance for the wider checklist. Where health data is in scope, a vendor's trust posture should show SOC 2, a BAA, PII redaction and data residency options without you having to ask.
How much does automated intake save, and what do the published numbers actually say?
Most published figures cover the whole claims function, not intake alone, so read them as the ceiling rather than the forecast. McKinsey reports that insurers taking a domain-based approach to AI have seen a 3 to 5 percent accuracy improvement in claims and a 20 to 40 percent reduction in customer onboarding costs. Its 2022 Claims 2030 analysis projected a possible reduction in employment of up to 46 percent for claims handlers, examiners and investigators: a statement of what the technology could absorb, not a first-year target.
The fraud numbers explain why triage matters even though the agent never adjudicates. The Coalition Against Insurance Fraud estimates that insurance fraud costs American consumers at least $308.6 billion a year and occurs in about 10% of property-casualty losses. Deloitte's 2026 insurance outlook estimates AI-driven, real-time fraud analytics could save P&C insurers up to US$160 billion by 2032. An intake agent contributes by flagging signals early (duplicate claims, timeline gaps, conflicting documents) and routing to special investigations with the evidence attached. The investigator decides.
The customer numbers are the ones most within reach. In the same J.D. Power property study, satisfaction for a claim completed within 10 days scores 762 on a 1,000-point scale and falls 167 points to 595 when repairs take more than 31 days. On the auto side, the 2025 U.S. Auto Claims Satisfaction Study puts average cycle time for repairable vehicles at 19.3 days, with total losses now 27% of claims. An agent cannot shorten a body shop's queue. It can remove the days a file sits waiting for a document, and it can keep the insurer easy to reach for the whole 44 days.
What this looks like in practice: a worked example
A policyholder's basement floods overnight. At 6am they open a chat on their phone, describe the water and upload two photos. A Lorikeet concierge captures the cause of loss, the date of discovery and the affected rooms, confirms no one is hurt, verifies identity, and reads the policy record: active, claimant is the named insured, administrative facts only. It opens the claim in the claims core, acknowledges the photos, lists what else it needs (a contractor estimate and photos of the water source) and texts a secure upload link. That evening the estimate arrives by email. The same agent, on the same claim record, extracts the line items and total, checks the loss date against the policy period, confirms the file is complete and routes it to the property queue with a structured summary. The claimant asks: "Is this covered, and how much will I get?" The agent acknowledges the question, explains that a licensed adjuster reviews coverage and amount, and does not answer it. Two days later the claimant calls. The voice agent picks up in under a second, recognises the claim, and reports that an adjuster was assigned that morning and an inspection is booked for Thursday.
Change one fact and the path changes. If the estimate is dated before the reported loss, the validation step fires a conflict flag, the agent stops, and the file routes to special investigations with the contradiction documented. It does not confront the claimant or pick a version to believe. That branch is a hard rule in a structured workflow, exercised in pre-launch simulations alongside a red-team set of claimants pushing for a coverage answer. In production, Coach scores every interaction against the boundary, so a near-miss surfaces as a pattern rather than a complaint, and the QA record shows what the claimant said, which value came from which document, and where the agent declined to determine coverage.
The published proof for this kind of deployment in a regulated setting comes from finance rather than insurance: Carmoola, an FCA-regulated UK car finance provider, resolves 60% of inbound conversations end to end with Lorikeet. Pricing is per resolution, about $0.80 for a chat, email or SMS resolution and about $1.20 for voice, with unresolved or unsatisfactory tickets costing nothing (see pricing). The honest limitation: Lorikeet does not estimate damage from a photo, does not replace your claims core, and asks for real setup work (mapping document types, writing escalation triggers, running the simulations) before the first live claim. Teams evaluating this usually bring their messiest intake scenarios to a demo and watch the agent decline the coverage question first.
What still needs a human
The list is short, stable, and should be hard-coded rather than left to the model. Coverage, liability, reserves and payout decisions belong to a licensed adjuster; the agent gathers and routes. Any report of injury, fatality or active emergency stops intake and goes to a person. Fraud referrals are made by an investigator; the agent supplies the flags and the evidence. Total-loss and high-severity claims warrant a human early even when intake is complete. Vulnerable or distressed claimants, complaints, and any mention of a lawyer, a regulator or the press route out immediately with the file attached. Low-confidence extractions and document conflicts go to review, not into the record. And the adjuster, not the agent, explains the decision. Lorikeet can deliver that message on the channel the claimant prefers and answer procedural questions about it, but the words that explain why come from an accountable person. Where health data is involved, an insurance deployment inherits the same handling rules a healthcare one would.
Intake automation succeeds when it is boring: the file arrives complete, the adjuster starts most of the way through, the claimant hears from you before they ask, and every step is on the record. Start with one line of business and one severity band, write the boundary before the flow, and widen scope as the audit trail proves it holds.
