---
title: "Automated claims investigation: the step that grows as everything else speeds up"
description: "New York DFS logged 41,686 suspected healthcare fraud reports in 2024 and 89% arrived unresolved. Automated claims investigation is the layer that absorbs what every upstream automation creates."
date: "2026-09-04"
lastModified: "2026-11-16"
author: "Nitish Badu"
tags: ["Guides"]
canonical: "https://gethesperai.com/blog/automated-claims-investigation-explained/"
---

# Automated claims investigation: the step that grows as everything else speeds up

> **TL;DR** Automated claims investigation is software that takes a flagged claim and runs the investigation itself: selecting procedures, gathering and reconciling evidence across independent sources, and producing a documented finding a human signs. It is the step that absorbs the risk every other claims automation creates, because faster straight-through processing pushes more claims past the last person who would have looked.
>
> - 41,686 NY healthcare fraud referrals in 2024, and 89% left unresolved
> - Soft fraud is 60% of incidents at 20-40% detection rates
> - Investigation execution is the one stack layer with no incumbent

- **89%** - Suspected healthcare fraud reports filed with NY DFS in 2024 that arrived unresolved (NY Department of Financial Services, March 2025 report)
- **$45B** - Annual cost of property-casualty insurance fraud in the United States (Coalition Against Insurance Fraud, 2022 economic impact study)
- **-6%** - Projected change in US claims adjuster, appraiser, examiner and investigator employment, 2025-2035 (Bureau of Labor Statistics, Occupational Outlook Handbook)
- **~25%** - Share of flagged claims that receive a full manual investigation (Hesper internal benchmarks)

Referrals are not the scarce resource. Resolution is. A carrier can add detection capacity on a procurement cycle and, until recently, could not buy investigative capacity at any price, because no vendor sold it. Automating everything upstream is what made the investigation gap expensive.

This post defines automated claims investigation as a capability test and shows why the flagged lane grows as the clean lane accelerates. For the lifecycle view, see our [guide to claims automation](/blog/claims-automation-pillar/). Two companion posts cover [which layer of the claims stack has which vendor](/blog/insurance-claims-automation-2026-whats-automated/) and [what full investigation autonomy requires architecturally](/blog/autonomous-ai-claims-investigation-pillar/).

## Claims automation worked, and that is what made investigation expensive

Claims automation has worked at every stage that reduces to a repeatable operation: intake, triage, coverage verification, damage estimating, payment. Investigation does not reduce that way. So the pipeline got faster on both sides of investigation and stayed the same in the middle, and the investigation step inherits a bigger queue.

The clearest evidence comes from the US government, not a vendor. The [Bureau of Labor Statistics](https://www.bls.gov/ooh/business-and-financial/claims-adjusters-appraisers-examiners-and-investigators.htm) projects a 6 percent decline in employment of claims adjusters, appraisers, examiners and investigators from 2025 to 2035. Its stated reason: "Technology is expected to automate some of the tasks that these workers currently perform."

Meanwhile the mix is moving toward claims a photograph cannot settle. [CCC Intelligent Solutions reports in Crash Course 2026](https://www.cccis.com/reports/crash-course-2026), which is vendor-published industry data, that repairable claim volume fell 9.7% in 2025. Bodily injury paid claim frequency is up 11% over two years, and BI now accounts for 52.4% of total liability dollars paid.

## What automated claims investigation actually means

The term is loose enough to mean almost anything, so the definition here is a capability test. Each item is anchored to something a regulator already requires of a human investigation, so the bar is not one a vendor gets to set.

### The eight capabilities that make it count

1. It decides what the specific flagged claim requires. A fixed script run on every referral is a checklist, and [California 10 CCR § 2698.36(a)(3)](https://www.law.cornell.edu/regulations/california/10-CCR-2698.36) contemplates procedure selection rather than a uniform script.
2. It acquires evidence across independent sources without a human commissioning each one.
3. It reconciles conflicts between sources rather than concatenating them.
4. It produces a finding with an explicit basis. California 10 CCR § 2698.36(a)(5)(A)-(F) is the practical template, down to whether the investigation is complete.
5. It emits a reconstructable record. NAIC Model Regulation #902 § 4.B requires detailed documentation in each claim file sufficient to permit reconstruction of the insurer's activities relative to each claim.
6. It runs on the whole flagged population, not the top of the queue.
7. It finishes inside the settlement clock. NAIC Model #902 sets 21-day and 45-day checkpoints on first-party claims.
8. It hands adjudication to a human. The pay, deny or refer call stays with the adjuster and the SIU lead who signs it.

Capability four is the tell. A system that cannot say that the investigation is not complete, and name what is missing, is not investigating. It is summarising.

**Figure 1 (The capability test) - It counts as investigation only if it ends in a finding a human can sign.** The definition in this article as a pipeline: every stage is something a regulator already requires of a human investigation.

Flagged claim in: Claim file and loss history; Policy and underwriting record; Cross-carrier data and court records; Provider and billing history; Imagery, metadata and statements

The investigation itself · the automated layer's job

1. **Procedure selection** - Decides what this claim requires, not one script for every referral
2. **Evidence gathering** - Pulls each independent source without a human commissioning it
3. **Conflict resolution** - Reconciles contradictions between sources instead of listing them
4. **Finding with basis** - States the facts, the misrepresentation, materiality and completeness

Reconstructable: every source, tool call, timestamp, inference

Escalation from stage 4: Not complete? The finding says so and names what is missing

**Documented finding - Adjuster and SIU lead decide**: Pay, Deny, Refer

**Not this:** A fraud score, a red flag, an OCR pass, a case queue or a summary of a file a human already built.

Source: capability test as defined in this article, anchored to California 10 CCR § 2698.36(a)(3) and (a)(5)(A)-(F), NAIC Model Regulation #902 § 4.B and NAIC Model #900 § 4.F.

### What automated claims investigation is not

- Not a fraud score. A score is a prior about a claim, not evidence gathered for that specific claim.
- Not a red flag firing. Under [California 10 CCR § 2698.36(c)](https://www.law.cornell.edu/regulations/california/10-CCR-2698.36), a red flag is the input that obligates an investigation. It is definitionally not the output.
- Not an OCR or document processing pass. Extracting the fields on a document is not testing whether they are true.
- Not a summary of a file a human already assembled. That compresses reporting time, not investigation time.

## Faster settlement moves more claims past the only person who would have looked

Straight-through processing does not reduce investigation load. It concentrates it. Paying the clean lane without a human touch works, but the residue left behind is denser in exactly the fraud type that is hardest to catch, because a fast lane selects for claims that look ordinary.

Detection vendors describe this architecture plainly. [FRISS](https://www.friss.com/), one of the established detection platforms, describes its own product this way:

> Anything that crosses your risk threshold gets flagged for a human to review, everything else moves through untouched
>
> - FRISS, product description on friss.com

That is the correct design, and a complete statement of the problem. The flagged lane terminates in a person, and its throughput is set by headcount.

[Deloitte data reported by Insurance Journal in May 2025](https://www.insurancejournal.com/news/national/2025/05/14/823588.htm) splits fraud in two. Soft fraud - inflation, buildup, exaggeration - is about 60% of all incidents and is detected at rates of 20% to 40%. Hard fraud is about 40% of claims fraud and is detected at 40% to 80%.

The majority category is the least-detected one, and it is the category a fast lane pays. Nothing is categorically wrong with the claim, only with its size.

**Figure 2 (What survives the fast lane) - The biggest category of fraud is the one detected least often.** Deloitte's split of claims fraud: each type's share, and the range of detection rates reported for it.

**Soft fraud**

- Share of all fraud incidents: 60% (Inflation, buildup, exaggeration)
- Detection rate: 20% to 40% (Reported range)

**Hard fraud**

- Share of claims fraud: 40% (The minority category)
- Detection rate: 40% to 80% (Reported range)

**Takeaway:** A fast lane selects for claims that look ordinary. Nothing is wrong with a soft-fraud claim except its size.

Source: Deloitte data reported by Insurance Journal, May 2025. Soft fraud is a share of all incidents, hard fraud a share of claims fraud; detection rates are reported as ranges.

Upstream automation raises throughput into the flagged lane, and detection improvements raise the flag rate. Neither touches the number of claims one investigator can work in a month. The lane model is in our [claims triage framework](/blog/claims-triage-decision-framework/).

## The flag volume is compounding, and the referral data says so

Fraud referral volume is growing faster than the capacity to resolve it, and one state regulator publishes the series that proves it. Suspected healthcare fraud reports filed with the New York Insurance Frauds Bureau went from 21,015 in 2020 to 41,686 in 2024, a 98% increase over four years.

| Suspected healthcare fraud reports received by the New York Insurance Frauds Bureau (NY DFS, March 2025 report) | Value | Share |
| --- | --- | --- |
| 2020 | 21,015 | 50% |
| 2022 | 30,079 | 72% |
| 2024 | 41,686 | 100% |

These are healthcare-related reports, New York only, and 93% of them are no-fault auto, which is property-casualty business. What happens to them is the finding. [The report](https://www.dfs.ny.gov/system/files/documents/2025/03/nysdfs-2024-health-fraud-annual-rpt-20250315.pdf) states it in the Department's own words:

> Eighty nine percent of these reports were designated by the insurer as 'for intelligence purposes' only or that the insurer is continuing to investigate the claim with the remaining 4,585 reports recommended for further review by the Department.
>
> - NY Department of Financial Services, Investigating and Combating Health Insurance Fraud, March 15, 2025

That is not a carrier SIU coverage rate. What it does establish is that the overwhelming majority of fraud referrals arrive at a regulator in an unresolved state, and that the insurers filing them said so themselves.

Carriers describe the same dynamic. In the Coalition Against Insurance Fraud's [2024 State of Insurance Fraud Technology Study](https://insurancefraud.org/wp-content/uploads/SHIFT-State-of-Insurance-Fraud-Technology-Study.pdf), a small-sample survey of 35 carriers, respondents list "More referrals" among the benefits of anti-fraud technology. On the same slide, among the challenges, they list "SIU cannot handle the volume of potentially fraudulent claims."

Against Hesper's internal benchmarks, a manual SIU investigation runs 14+ days per case, an investigator carries 200+ open cases, and throughput lands near 10 investigations per investigator per month. That is roughly 20 months to work a standing caseload once.

So about 25% of flagged claims get a full manual investigation. The rest are paid, denied without complete work, or queued until they age out. The mechanics are in [why most flagged claims are never investigated](/blog/why-flagged-insurance-claims-never-investigated/).

**Figure 3 (The referral pile) - Referrals nearly doubled. One caseload takes about 20 months to clear once.** The regulator's series shows the pile growing; the caseload arithmetic shows why the people working it cannot catch up.

1. **Reports, 2020: 21,015** - Suspected healthcare fraud reports to the New York Insurance Frauds Bureau
2. **Reports, 2024: 41,686** - Up 98% in four years, 93% of them no-fault auto
3. **Unresolved: 89%** - Intelligence-only or still under investigation, by the insurer's own designation
4. **Further review: 4,585** - Recommended for further review by the Department

200+ divided by 10 equals 20

- Open cases per investigator: 200+ (Hesper internal benchmark)
- Investigations a month: about 10 (Manual, 14+ days per case)
- Months to work the caseload once: roughly 20 (Before a single new referral arrives)

**Result:** About 25% of flagged claims get a full manual investigation. The rest are paid, denied without complete work or age out.

Sources: New York Department of Financial Services, Insurance Frauds Bureau report, March 2025 (healthcare-related reports, New York only; the 89% is not a carrier SIU coverage rate). Caseload arithmetic: Hesper internal benchmark, derived.

## The regulation prices the gap

Claims regulation puts a price on the investigation gap in a way an operations dashboard does not. Settlement clocks run on the promptness duty, and the fraud carve-out from those clocks is conditional on producing investigative output.

[NAIC Model #902](https://content.naic.org/sites/default/files/model-law-902.pdf) is where the clocks live. Section 7.A requires that a first party claimant be advised of acceptance or denial within twenty-one days of properly executed proofs of loss. Section 7.B requires a letter at forty-five days, and every forty-five days after that, setting out the reasons more time is needed for investigation.

The stage-by-stage timeline is in [the claims process from FNOL to settlement](/blog/claims-process-end-to-end-fnol-to-settlement/). Then the carve-out:

> Where there is a reasonable basis supported by specific information available for review by the insurance regulatory authority that the first party claimant has fraudulently caused or contributed to the loss, the insurer is relieved from the requirements of this subsection; provided, however, that the claimant shall be advised of the acceptance or denial of the claim within a reasonable time for full investigation after receipt by the insurer of a properly executed proof of loss.
>
> - NAIC Model Regulation #902, Unfair Property/Casualty Claims Settlement Practices, Section 7.A

The relief is real and it is conditional. "Specific information available for review by the insurance regulatory authority" describes investigative output. A carrier investigating roughly a quarter of what it flags holds the carve-out on paper for all of those claims and in fact for a quarter of them.

> **California extends the duty to investigate to system-generated referrals**
>
> [California 10 CCR § 2698.36(c)](https://www.law.cornell.edu/regulations/california/10-CCR-2698.36) requires the SIU to investigate each credible referral of suspected insurance fraud it receives from integral anti-fraud personnel, including automated or system-generated referrals. A credible referral is one that includes a red flag. So a detection engine in California produces more legally obligated investigations, not just more flags. The state walkthrough is in our [guide to 10 CCR 2698 SIU compliance](/blog/california-10-ccr-2698-siu-compliance/).

Florida turns the gap into a filed statistic. [Fla. Stat. § 626.9891](https://www.flsenate.gov/Laws/Statutes/2024/626.9891) requires each insurer to report annually, for each line of business, the number of claims received, the number referred to the anti-fraud investigative unit, and the number investigated or accepted by that unit. Coverage is a number the state collects.

None of this forbids automation. The practical bar is one sentence long: an automated investigation has to be able to explain, per claim, what it did, what it found, what it could not resolve, and why.

*Figure: Speed up the clean lane and the flagged lane does not shrink. It gets denser, and its throughput is still set by headcount.*

## Which layer each vendor actually occupies

Every layer of the claims stack has an incumbent except one. Prevention, detection, estimating, payments, case management and claims administration all have named vendors with real products behind them. Investigation execution has manual SIU teams.

| Dimension | Manual SIU investigation | Detection score only | Automated claims investigation |
| --- | --- | --- | --- |
| Input | A referral in a queue | A claim at or after FNOL | A flagged claim plus the full file |
| Output | A written investigation summary | A score, a rank or a red-flag list | An evidence-backed finding with its basis |
| Coverage of flagged claims | ~25% (Hesper benchmark) | 100% scored, 0% resolved | 100% investigated |
| Cycle time per case | 14+ days (Hesper benchmark) | Real time to score | Minutes, not weeks |
| Evidence produced | Yes, by hand, one phase at a time | None, because a score is not evidence | Yes, 15+ phases in parallel |
| Conflicts between sources | Resolved by the investigator | Not surfaced | Resolved and shown with the reasoning |
| Auditability | Depends on the investigator's file discipline | Model transparency only, no per-claim evidence chain | Reconstructable per Model #902 § 4.B |
| Satisfies CA 10 CCR § 2698.36(a)(5)(A)-(F) | Yes, when done | No, a score answers none of the six questions | Yes, as the standard output format |
| What the investigator does | Executes every phase | Triages the score, then executes every phase | Reviews, challenges, overrides, signs |
| Scales with flag volume | No, bounded by attention | Yes | Yes |
| Vendor incumbents | None, the SIU team is the incumbent | FRISS, Shift, Verisk, BAE NetReveal, SAS | Hesper AI |

The table above compares the three ways a flagged claim gets handled today. The row that decides procurement is coverage: a faster manual process is still bounded by the number of investigators, so the ~25% to 100% shift, not the day count, is the loss-cost lever.

The agent wave has the same shape. As we found in [a survey of shipped claims agents](/blog/agentic-ai-insurance-claims-beyond-fnol/), deployments cluster in intake, communication and summarisation, where a language model has an obvious bounded job. Investigation is where the work is unbounded, which is why it comes last.

The 2024 CAIF study makes the same point on capability. These are the printed "not incorporated" percentages from the 35 responding carriers:

| Anti-fraud capabilities not incorporated, share of 35 responding carriers (CAIF, 2024 State of Insurance Fraud Technology Study) | Value | Share |
| --- | --- | --- |
| Geographic data mapping | 54.3% | 54.3% |
| Exception reporting / anomaly detect | 48.6% | 48.6% |
| Text mining | 37.1% | 37.1% |
| Entity alert capability | 31.4% | 31.4% |
| Case management | 25.7% | 25.7% |
| Automated red flags / business rules | 11.4% | 11.4% |

Automated red flags are close to universal. Asked what they plan to invest in over the next 12 to 24 months, every option respondents named is detection, analytics or case management. Investigation execution does not appear, because at the time of the survey it was not something a carrier could buy.

That is the layer Hesper AI occupies. It takes a flagged claim, runs 15+ investigation phases in parallel - document forensics, OSINT, statement cross-reference, timeline reconstruction, financial pattern analysis - and returns an audit-ready finding in minutes rather than weeks. It has built-in detection, so it runs standalone, and it is complementary to FRISS, Shift Technology and Verisk.

## What to do with the next automation dollar

If FNOL through payment is already automated and a detection platform is already in place, the marginal loss-cost lever is not more front-end speed. It is coverage of the flagged population, the one number in the stack a decade of fraud-tech investment has not moved.

This is not a headcount argument. SIU staffing does not shrink in this model, it gets re-aimed. The investigator's role shifts from execution to decision-making: reviewing findings, challenging them and owning the outcome on the record.

Make every flagged claim investigable. The test is not a demo. It is whether the number of claims flagged in a month still predicts the number investigated that month. If it does, the investigation layer is still manual, whatever else has been automated around it.

## Key takeaways

- Automated claims investigation absorbs the risk every other layer of claims automation creates, because faster straight-through processing moves more claims past the last person who would have looked at them.
- New York's Insurance Frauds Bureau received 41,686 suspected healthcare fraud reports in 2024, up 98% from 21,015 in 2020, and 89% were logged by the filing insurer as intelligence-only or as still under investigation.
- NAIC Model #902 § 7.A relieves an insurer of the 21-day settlement clock only where there is a reasonable basis supported by specific information available for review by the regulator. That information is investigative output.
- Every anti-fraud technology the 35 carriers in the 2024 CAIF study said they plan to buy in the next 12 to 24 months is detection, analytics or case management. Investigation execution is the one layer with no software incumbent.

## Frequently asked questions

### What is automated claims investigation?

Software that takes a claim already flagged as suspicious and runs the investigation itself: selecting which procedures the claim needs, gathering evidence across independent sources, resolving the contradictions between them, and producing a documented finding a human signs. The bar is set by what regulators already require of a human investigation. A system that cannot say whether the investigation is complete is doing something upstream of investigation.

### How is automated claims investigation different from fraud detection?

Detection decides which claims are worth looking at. Investigation is the looking. A detection platform scores or flags a claim, and its output is a signal. An investigation produces evidence: documents tested, histories pulled, statements cross-referenced, contradictions resolved. Under California 10 CCR § 2698.36(c) a system-generated flag is the event that obligates an investigation, not a substitute for one. They are different layers, not competing products.

### Is AI allowed to investigate insurance claims?

Yes, subject to constraints that shape what the system has to produce. The NAIC adopted its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers in December 2023. The operative expectation is that AI-supported decisions comply with existing insurance law and that regulators can examine the system. Automation is permitted. Unreconstructable automation is not.

### Can an insurer deny a claim based on an automated investigation?

No insurer should treat an automated finding as the denial decision. NAIC Model #900 § 4.F defines refusing to pay claims without conducting a reasonable investigation as an unfair claims practice. An automated investigation is built to produce the basis, meaning the evidence, the reasoning and the sources, but the pay, deny or refer decision stays with the human adjuster and the SIU lead who signs it.
