A Claims VP at a Top 50 P&C carrier watches three numbers every quarter: the SIU caseload, the cost per investigated claim, and the loss ratio on lines where fraud is concentrated. AI claims investigation moves all three. The hard part is scoping a buy the CFO will sign, the CIO will integrate, and the SIU director will run.
This playbook is for the Claims VP who has to defend the line item. The wider stack is in our autonomous AI claims investigation pillar.
The pre-RFP diagnostic: why this is a Claims VP decision
The AI investigation buy belongs in the claims-org budget, not the IT capital plan. Three signals say so: SIU capacity is the binding constraint on loss-ratio improvement. Cost per investigated case is the lever the CFO will model. And the antifraud-plan filing under NAIC Model Act #680, adopted in 48 states, is a claims-org responsibility, not an IT one.
At most Top 50 carriers, a manual SIU investigation takes 14+ days per case, an investigator carries 200+ cases, and only about 25% of flagged claims actually get investigated. The other 75% close on adjuster judgment without an investigation file.
SIU staffing growth, per the Insurance Information Institute, slowed to 1.4% in 2021-2022 from 2.5% the prior year. The capacity gap is widening, not closing.
AI investigation is not a detection-vendor swap. FRISS, Shift Technology, and Verisk ISO ClaimSearch keep doing what they do at FNOL. Investigation is the downstream layer, covered in our piece on prevention, detection, and investigation as three distinct layers.
Budget framework: what to ask for and how to model ROI for the CFO
A defensible budget ask has three parts:
- Scope the 90-day pilot at 1,000 to 5,000 cases on one or two lines and have it quoted against that volume.
- Plan the 12-month rollout run-rate to scale with flagged-claim count rather than seats or headcount.
- Carry the ROI in basis points of loss ratio on AI-investigated lines.
Build the CFO model on coverage-corrected unit cost, not headline per-case price. Manual SIU carries a high fully loaded cost per case and reaches about 25% of flagged claims. AI investigation reaches 100% at a unit cost low enough that the marginal cost of one more investigation stops driving the triage decision.
On a carrier flagging 50,000 claims per year, the manual budget covers about 12,500 cases, and covering all 50,000 lands well under that same budget. The fraud caught in the previously uninvestigated 37,500-case tier is the loss-ratio story, not the per-case savings.
Bring two scenarios: a conservative model at a 5% confirmed-fraud rate on the incremental coverage and a base case at 10%, matching the P&C fraud rate cited by the Coalition Against Insurance Fraud.
The CFO story is the 37,500 flags that close uninvestigated
A carrier flagging 50,000 claims a year funds manual SIU for about a quarter of them; the loss-ratio upside sits in the rest.
Two pitfalls show up in CFO reviews:
- Services markup on integration. The quoted per-case rate is almost never the loaded cost, because implementation services, evidence-pack storage, and adjuster training all land in the SOW. Ask every vendor to price the fully loaded case.
- Parallel-run costs. Most carriers underfund the 60 to 90 days when SIU runs the old workflow next to the AI-assisted one for audit confidence.
The full list is in our breakdown of hidden integration costs for legacy claims AI. For ROI scenarios across three carrier sizes, see our ROI case studies for AI claims investigation.
Interactive calculator
Size your pilot and 12-month budget
Hesper does not publish a rate card, so enter the per-case price from your own quote; what it buys depends on LOB mix and integration scope. Get a quote keyed to your portfolio.
The 12-month timeline: pilot, scale, mature
Commit to phase-gate decisions, not a 12-month lock. Phase 1 is a 90-day pilot on one or two lines, typically auto and workers compensation. Phase 2 scales to more lines once the pilot signal is clean. Phase 3 retires manual triage on the covered lines.
Write three pilot exit criteria into the SOW before signature:
- Investigation cycle time measured in minutes on 95% of pilot cases.
- False-positive resolution rate above 80% on rules-flagged claims (since rules-based detection runs 60-85% false positives in industry studies).
- Zero unresolved compliance flags from the SIU director on audit-trail completeness.
If any of the three miss, the scale phase does not auto-fund.
The scale phase funds only if the pilot clears all three exit criteria
Commit to phase-gate decisions, not a 12-month lock: each phase widens the case count only after the last one proves out.
- Pilot · months 1-31,000 - 5,000Cases on one or two lines, usually auto and workers comp
- Exit gateAll threeWritten into the SOW before signature. Any miss and scale does not auto-fund
- Scale · months 4-920,000 - 50,000Adds homeowners, commercial property and bodily injury
- Mature · months 10-1250,000 - 100,000All flagged claims in scope; the run-rate joins the claims operating review
- 95%of pilot cases with investigation cycle time measured in minutes
- above 80%false-positive resolution on rules-flagged claims (rules run 60-85%)
- zerounresolved SIU compliance flags on audit-trail completeness
Stakeholder map: the question each one will ask
Run this procurement without a written stakeholder map and you will lose 60 to 90 days of cycle time to ad-hoc objection handling, mostly from CIO security review and legal review of third-party data use. Pre-empt every reviewer with the one question they actually care about.
The working documents for these conversations:
- The third-party risk review that sets an AI deployment date walks the artifacts and contract clauses reviewers ask for.
- The 12-point evaluation framework in our AI fraud investigation vendor checklist.
- For the legal reviewer, the general counsel's guide to AI fraud investigation covers discovery, privilege, and admissibility.
Compliance is the reviewer most likely to be skipped and most likely to slow the deal in week 10. Bring them in at week one.
The relevant state rule for California carriers is California 10 CCR 2698.36, which requires the SIU to investigate each credible referral of suspected fraud and to document any decision not to investigate. AI investigation makes that easier: every flagged claim gets a documented decision in the audit log, the exact artifact the state DOI asks for in market conduct exams.
Change management: deploying AI without burning out SIU adjusters
The failure mode at carrier rollouts is anxiety, not pushback. SIU adjusters who handled roughly 10 investigations per month manually do not know what a review queue no longer capped by headcount means for their job. The Claims VP needs three written commitments before the first case routes, all communicated to SIU staff in the same memo.
The investigator stops gathering data and starts deciding on evidence packs
The three written commitments that go to SIU staff before the first case routes, drawn as a decision-rights map.
- 01Document collection and review
- 02Database and index-bureau lookups
- 03Statements cross-referenced, timeline rebuilt
- 04Evidence pack with sources and recommended next actions
- Finding, referral or denial
- Accept or dispute the AI output
- Escalate a dispute to the SIU director
- 1.Investigator marks the AI output as disputed
- 2.The SIU director makes the final call
- 3.Logged and reviewed monthly for model-quality signal
- No SIU headcount reduction in year one. The freed capacity covers the 75% of flagged claims that were previously uninvestigated, where the incremental fraud recovery sits.
- The investigator becomes the decision-maker over AI-produced evidence packs. Judgment stays human; data gathering, document collection, and database lookups move to the AI layer running 15+ phases in parallel.
- A named escalation path for any AI output the investigator disputes, with the SIU director as the final authority. Disputed outputs are logged and reviewed monthly.
The director's side of that rollout is mapped week by week in the SIU director's first 90 days with AI investigation. The manager-level version - who tells the team what, and when - is in leading a claims team through AI-augmented investigation.
Named-insurer evidence helps at the SIU town hall. AXA Switzerland's deployment with Shift Technology caught a CHF 120,000 false claim involving intoxicated workers in the first week after deployment, per Samuel Klaus, Head of Fraud at AXA Switzerland, in the published Shift Technology case study.
The 12-month outcomes report: what to bring to the CEO and Board
Four metrics belong on the month-12 board report, and only four:
- Investigation coverage as a percent of flagged claims.
- Cost per investigated case.
- Cycle time per case.
- Loss-ratio delta on AI-investigated lines, isolated from rate action and other claims-handling changes.
The board narrative is a coverage story, not a per-case cost story. The carrier is now investigating four times as many flagged claims at lower total cost, with documented decisions on every claim. The loss-ratio delta lands in months 6 to 12.
The forward question for the Board is whether the carrier's antifraud plan, SIU role definitions, and reinsurance reporting reflect a 100%-coverage operating model. Carriers that lock that operating model in 2026 will price 2027 books on a different loss-cost curve than carriers still triaging at 25% coverage.
Key takeaways
- AI claims investigation is a claims-org budget item, not an IT line. The loss-ratio lever and the NAIC Model Act #680 filing both sit with the Claims VP.
- Scope a 90-day pilot at 1,000 - 5,000 cases on 1-2 lines, priced against that case count.
- The CFO math is coverage-corrected: the fraud caught in the previously-uninvestigated 75% is the loss-ratio story.
- SIU headcount does not get cut in year one. Investigators move from data gathering to decision-making.