The CFO ROI memo: making the case for AI claims investigation

A finance reviewer can sign off on AI claims investigation using three numbers and nothing about how the agent works: the cost per investigated case, the size of the coverage gap it closes, and the payback period. The mechanics of document forensics and OSINT are an engineering detail. The dollars are not. This is the memo a Claims VP or SIU Director hands up the chain for financial sign-off, and it is written so the finance reviewer can defend each line in a board setting.

The argument runs against a common instinct, which is to build the case on headcount savings. That is the weakest version, and it invites organizational resistance. The durable return sits somewhere else: most carriers fully investigate only about 25% of the claims they flag, and the other roughly 75% is where leakage concentrates. Closing that gap toward 100% is the financial event. Speed matters because it is what makes full coverage affordable, but coverage is the lever, not speed alone.

This piece covers the one-page summary, where the money leaks, the cost of inaction, the unit economics of an investigated case, how to model the business case, payback under three scenarios, build-versus-buy, and a pilot structure that funds proof rather than a leap. For the worked carrier scenarios behind the payback math, pair this with the ROI case studies for AI claims investigation, and for the loss-cost backdrop see the claims fraud and leakage pillar.

Hesper AI takes a flagged claim and runs the full SIU playbook end-to-end. The category is fraud resolution, not fraud detection - from fraud detection to fraud resolution is the move the memo is funding. The carrier has almost certainly already paid for detection. A finance reviewer is being asked to fund the layer that turns those flags into resolved cases.

The three numbers that carry the memo are these. First, an AI-investigated case runs about $150 versus roughly $2,500 manually, a Hesper internal benchmark. Second, coverage of flagged claims moves from about 25% to 100%, which is the line item that converts the spend into recovered and denied dollars. Third, payback follows from the gap between small incremental opex and the recoverable leakage that EY puts at 7% to 14% of total claims spend. Everything below is the supporting detail behind those three lines.

Start with the size of the pool. Insurance fraud costs the US an estimated $308 billion a year, with $45 billion in property and casualty and $34 billion in workers compensation, per the Coalition Against Insurance Fraud. The Insurance Information Institute corroborates the same figures. Roughly 10% of property-casualty claims may involve fraud, per the National Insurance Crime Bureau. That is the industry pool. The carrier-specific leak is a subset of it, and it has a precise location.

EY's claims quality assessments put leakage at approximately 7% to 14% of a carrier's total claims spend, and one of EY's four named root causes is inadequate investigation of injury causation and liability, per EY. That root cause is the one a finance reviewer can act on directly. It maps to a capacity constraint, not a detection failure: the flags exist, but most of them are never fully investigated. For the full breakdown of where leakage hides on a claim file, see how carriers reduce claims leakage.

Here is the structural mechanism. Detection vendors flag suspicious claims, and rules-based systems do so with false-positive rates of 60-85%, so the flag pile is large and noisy. A manual SIU then investigates roughly 25% of those flags, because each case takes 14+ days and an investigator carries 200+ cases. The remaining roughly 75% are paid, denied without full work, or queued indefinitely. The leakage concentrates in that uninvestigated tail. The memo is not asking finance to fund faster detection. It is asking finance to fund the investigation capacity that converts the existing flag pile into resolved cases.

Doing nothing is not a zero-dollar line. It is the leakage that compounds on every uninvestigated flag, period after period, plus the stranded value of detection spend that produces flags no one can work. The carrier is already paying a detection vendor to surface suspicious claims. If only about 25% of those flags are ever fully investigated, the carrier is buying flags it cannot act on. That is the financial waste hiding in plain sight, and it does not appear as a line item anywhere on the income statement.

Quantify it as a formula the finance reviewer fills in, not as a Hesper-published savings number. Take the carrier's annual flagged-claim volume. Multiply by the share currently uninvestigated, which is roughly 75% on a 25% coverage baseline. Apply a recovery-and-denial rate anchored to EY's 7% to 14% leakage band. The result is the illustrative annual leak sitting in the uninvestigated tail. This is illustrative arithmetic on the carrier's own inputs, not a guaranteed outcome. Each policyholder already absorbs about $900 a year in added premium from fraud, per the Coalition Against Insurance Fraud, which is the downstream cost of the same untreated leak.

The reason full coverage was never affordable is unit cost. A manual SIU investigation runs about $2,500 per case at 14+ days of investigator attention, and one investigator completes around 10 investigations per month. At that cost and throughput, investigating 100% of flags would require either a headcount bill the loss-recovery story cannot justify or a backlog that grows faster than the team clears it. So carriers ration investigation down to the roughly 25% a human team can reach. The rationing is an economic decision forced by the per-case cost, not a quality choice.

AI investigation changes the inputs. Hesper runs 15+ investigation phases in parallel - document forensics, OSINT, statement cross-referencing, timeline reconstruction, financial-pattern analysis - on every flagged claim, so per-case attention stops being the bottleneck. That compresses a case to 2-4 hours at about $150, and lifts throughput to 800+ cases per investigator per month. The order-of-magnitude drop in per-case cost is precisely what makes investigating 100% of flags affordable instead of rationing down to 25%. Coverage and cost are the same lever viewed from two sides.

A defensible memo models incremental investigated cases times the recovery-and-denial rate against incremental opex. It never leads with FTE displacement. Three inputs carry the model. First, the incremental flagged claims you can now investigate - the gap between current coverage and 100%. Second, the recovery-and-denial rate on those incremental investigations, anchored to a documented leakage range rather than an optimistic guess. Third, the incremental cost to run them at about $150 per case.

On the return side, anchor the recovery rate to EY's 7% to 14% of claims spend rather than inventing one, and apply it only to the incremental cases, not the whole book. Add the timing benefit: shorter cycle time tightens reserving and accelerates settlement and recovery, which improves when loss costs are recognized. On the cost side, include platform fees, integration, and change management, not just per-case cost. The hidden cost lines are real, and the build-versus-buy section covers them - the deeper treatment is in the hidden integration costs of legacy claims AI. For three fully worked carrier scenarios with the math laid out, the ROI case studies are the companion to this memo.

Express the loss-ratio impact in basis points on the specific lines where fraud concentrates - workers comp and auto bodily injury - rather than at the book level. A book-level number dilutes the effect and reads as noise to a board. A line-level basis-point figure on the lines a carrier already worries about reads as a targeted intervention, which is what it is.

No single payback number survives board scrutiny, so present a range. Model payback under conservative, base, and aggressive assumptions on three variables: flagged-claim volume, the recovery-and-denial rate within EY's 7% to 14% band, and how fast the pilot ramps to steady-state coverage. The mechanics are favorable in every scenario because the per-case cost drops by roughly an order of magnitude while coverage rises from about 25% to 100%, so the incremental opex stays small relative to the recoverable leakage.

The chart above is illustrative and meant to show shape, not a promised figure. The bars represent relative net year-1 return across the EY leakage band applied to a carrier's own flagged volume; fill in the absolute dollars with your inputs. The per-case delta of roughly $2,350 saved on every investigated case - about $2,500 manual less about $150 with AI - is a Hesper internal benchmark and the engine behind why even the conservative bar stays net positive. Present payback as a sensitivity table tied to your flagged volume, and treat any single figure as illustrative. The realization ramp matters too, which is why the SIU director's first 90 days with AI is worth reading alongside the finance model.

A finance reviewer will ask whether to build the investigation layer internally. The honest answer is that building it is a multi-year R&D bet, not a finance-friendly capex line. An internal team would have to build an autonomous agent that runs 15+ investigation phases in parallel, produces an audit-ready trail satisfying California 10 CCR 2698.36 and the antifraud-plan filings under NAIC Model Act 680, and integrates with Guidewire or Duck Creek. The sticker price hides services markup, customization, training, integration latency, opportunity cost, and lock-in, all detailed in the hidden integration costs of legacy claims AI.

Buying converts that uncertain capital bet into a per-case opex line that scales with flagged volume. It also keeps the spend complementary to detection tools the carrier already owns - Hesper is complementary to FRISS, Shift Technology, and Verisk, not a replacement. FRISS scores and hands off to a human investigator; Shift Technology offers handler-assist that speeds the adjuster; Verisk flags through cross-carrier data. None of them closes the downstream investigation gap. The investigation layer is the layer no other vendor occupies, which is exactly why the carrier's unrealized return sits there. For the diligence questions a finance reviewer should run before trusting any per-case figure, use the AI fraud investigation vendor checklist.

Structure the first dollar as a scoped pilot with a measured baseline, so finance funds proof rather than a leap. Capture the baseline before the pilot: current flagged-claim volume, current coverage of those flags, current cycle time, and current cost per investigated case. Run the pilot on a defined slice of flagged claims, and measure the same four metrics. Coverage and per-case cost are the primary pilot KPIs because they are the two sides of the loss-cost lever the whole memo turns on.

Frame the headcount story carefully throughout. The model does not depend on cutting investigators - it depends on coverage. With AI handling the mechanical 15+ phases, the investigator's role shifts from execution to decision-making: reviewing audit-ready output, handling exceptions, and making the determination. Headcount gets re-aimed at higher-judgment work, not eliminated, which is what keeps the case from triggering organizational resistance. The operational side of that transition belongs to the Claims VP, and the Claims VP deployment playbook is the artifact that pairs with this memo on the operations side.

The next decision after the pilot is not whether the technology works but how fast to scale coverage toward 100%. That is a finance question disguised as an operations one: each increment of coverage applies the same favorable unit economics to more of the flag pile, so the question becomes how quickly the carrier wants to convert its existing detection spend into resolved cases. A reviewer who funds a clean baseline and a measured pilot will have the data to answer that question with numbers rather than conviction, which is the position a finance function wants to be in before scaling any opex line.