Start with a claims task that creates the same manual work over and over, such as checking incoming documents against policy records. Map the inputs and the exceptions. Give an adjuster the authority to review flagged issues. Compare the results with the current process. FNOL triage can be another candidate when the intake data is available and the escalation rules are clear. Spej guidance: the right first task depends on the carrier's process and its records.
By Sean Blair, Co-Founder & Chief AI Officer, Spej | November 19, 2026
Choose a first claims workflow
A first workflow needs inputs, an output, an owner, and a handoff. The table compares three candidates a claims team can scope. Treat each row as an illustration. The published claims engagement is described later, in its own section.
| Workflow | Inputs | Output | Owner | Handoff |
|---|---|---|---|---|
| FNOL triage | First notice details and attachments | A suggested priority or route for review | The person who owns intake | Incomplete notices and escalations go to a person |
| Document intake | Incoming files and the claim record | A check for missing information | The person who receives new files | Missing items go back for follow-up |
| Document review | Claim documents, policy terms, and previous records | Flagged inconsistencies for a person to review | An adjuster with authority to review | Flagged issues go to that adjuster, who records the disposition |
Spej guidance: pick a task that happens often, stays inside a clear boundary, and uses records the team can reach. Name an owner. Record a baseline you can measure. FNOL prioritization and pulling details out of incoming documents are examples you might test. They are illustrations, and they are separate from the published engagement below.
Map the work before choosing a tool
Follow a claim from intake to review
Here is an illustrative flow for a claim file:
- Capture the FNOL and the attachments.
- Check for missing information.
- Compare the relevant records.
- Send exceptions to an adjuster.
- Record the outcome.
Ask where the data lives, who owns it, and how the team handles a missing record or two records that disagree. Coverage stays with the people who already have authority to decide it. This flow is a way to prepare the file for that review. It is an illustration, and it is separate from any client engagement.
Set the human approval point
Spej guidance: define the data the workflow may use, who can access it, what the reviewer is allowed to decide, and how an item escalates. Keep a record of the source information, the output, and the disposition. A person reviews flagged issues and approves actions that affect the claim. Review privacy, security, and the requirements that apply to your carrier with your own specialists.
The NIST AI RMF Core organizes AI risk management into four functions: govern, map, measure, and manage. Under map, processes for human oversight are defined, assessed, and documented. The page is an excerpt from the NIST AI RMF 1.0 (2023), and it notes that a revised version is in progress. NIST describes the framework, and its playbook, as voluntary. Treat this as general guidance. It does not set insurance law. NIST AI RMF Core
For how a carrier can write down decision rights, see AI strategy and governance.
What a regional insurer's claims review showed
A regional property and casualty insurer used a system that checks claim documents against policy terms and previous records, then flags inconsistencies for a person to review. The case reports: Over $100K in potential fraud flagged within weeks of deployment. Faster, more consistent claims reviews, with a person checking flagged issues. The figure is potential fraud flagged for review, which is different from confirmed fraud or money recovered. Read the claims review case.
A separate case covers a regional insurer's AI operating model. The first initiative was automated competitive intelligence. The case reports 60 days to first working value, and 12 AI initiatives prioritized and roadmapped. In the first two weeks, Spej conducted stakeholder interviews across seven departments. The governance work included an AI acceptable use policy, a model risk management process aligned to existing enterprise risk standards, and a monthly AI steering committee with the C-suite. These are results from that case. Read the operating model case.
A wider discussion of records and review for insurance AI is in trusted data before agentic AI.
Measure a pilot and decide what comes next
Spej guidance: baseline handling time, rework, and review effort before the pilot changes the work. Test representative claims, and test the exceptions you already see. Check missed issues and incorrect flags. Track reviewer time and the cost of running the pilot. Agree who can pause the workflow. Agree what results justify continuing, what results justify a change, and what results justify stopping.
AI implementation is the build work that follows once the records, the owner, and the approval point are clear enough to scope.
Where AI falls short
- Documents and policy records may be incomplete, outdated, or inconsistent.
- A check can miss an issue. It can also flag a claim that has no inconsistency.
- Review needs time, the evidence behind the flag, and the authority to act.
- Pause automation that changes a claim when the inputs are unreliable or nobody owns the result.
- A person reviewing the output can still reach a wrong outcome.
FAQ
Where should a P&C carrier start with AI in claims?
Spej guidance: start with a repeatable claims task, such as checking documents against policy records. Set a baseline for the current process, check which records you can reach, and name the person who reviews flagged issues. Test representative claims and exceptions before you expand. The right task depends on the carrier's process and its data.
Can AI help with FNOL triage?
It can help organize incoming information and route items for a person to review. Define the inputs, the routing rules, and the escalation path before you test. Run the test on representative cases, including notices with missing details. Keep a person responsible for the decision. This is a proposed use, separate from Spej's published claims case.
Does AI decide whether to pay a claim?
In Spej's published claims case, the system checks claim documents against policy terms and previous records. It flags inconsistencies for a person to review, and a person checks those flagged issues. The case describes that review. Autonomous coverage decisions and payment decisions sit outside what the published case shows.
How should claims leaders measure a pilot?
Compare the current process with the pilot on handling time, rework, review effort, and operating cost. Check missed issues and incorrect flags on representative claims and exceptions. Agree what results justify continuing, changing, or stopping, and name who can pause the workflow. Spej guidance: measure against your own baseline.
Start with one claims workflow
Bring a claims workflow, and the gaps in its data, to a conversation with Spej.
AI Roadmap is starting at $10K. Design & Build is starting at $20K. AI Office is starting at $10K a month. Scope determines the price.