Question 01
How much can AI compress First Notice of Loss processing time?
AI FNOL agents compress intake from 4 to 6 minutes per claim to 15 to 30 seconds, with validation accuracy matching or exceeding human intake on standard P&C lines. The cycle-time gain compounds downstream: faster intake means faster triage, faster reserve setting, faster customer acknowledgment. Shift Technology and Decerto benchmarks show carriers using AI-powered claims automation resolve claims 75% faster with 30 to 40% cost reductions.
Why most vendors get this wrong: they benchmark FNOL speed in isolation and skip the downstream cycle-time math. The win isn't 5 minutes per claim. The win is 4 days off the overall claim cycle, which moves loss-adjustment expense and customer NPS together.
Right answer pattern: the vendor produces a before-and-after cycle-time analysis from a comparable carrier customer (same line of business, same volume band) covering FNOL through closure, not just intake. The vendor who can't produce that comparison is selling the intake KPI in isolation.
Question 02
What's the real subrogation recovery uplift from AI?
Carriers and MGAs lose 15 to 30% of recoverable claims to deadline slippage on manual subrogation tracking. The industry misses $20 billion per year in potential recoveries. Carriers that automate subrogation identification and tracking report recovering 20% more dollars per year while cutting staff time on file management by more than half. The clean automation pattern: AI agent identifies opportunity within hours of payment, generates demand letters, tracks deadlines, escalates to legal at pre-defined thresholds.
Why most vendors get this wrong: they sell subrogation as a tracking-software upgrade rather than an end-to-end recovery agent. The legal escalation workflow is where the recovery dollars actually live, and most vendor scope-of-work stops at "we identify the opportunity." That's half the win.
Right answer pattern: the vendor's scope includes opportunity identification, demand-letter generation, deadline tracking, and legal escalation orchestration. Bonus points if the vendor has a documented partnership with subrogation legal panels rather than handing the file back to your in-house legal team to figure out.
Question 03
How does AI integrate with Guidewire ClaimCenter, PolicyCenter, or Duck Creek?
Real integration means working calls against the documented APIs for ClaimCenter, PolicyCenter, BillingCenter (Guidewire) or the equivalent Duck Creek modules, validated against the carrier's actual version (Guidewire Cloud, Guidewire on-prem, Duck Creek SaaS). Most vendor decks show the certified-partner logo without specifying which version they've shipped against. The honest test: ask for two named carrier customers running on the same version with a live integration in production.
Why most vendors get this wrong: they have a working integration with Guidewire Cloud and assume parity with on-prem Guidewire 10.x, which has 200+ carrier customizations that break the standard API contract. Duck Creek SaaS integration doesn't transfer cleanly to Duck Creek on-prem. The vendor who waves the partner logo without naming the version is going to die in integration.
Right answer pattern: two named carrier customers on the same version with live integration, a specific named integration partner if the work is done by a third party, and a working scope-of-work covering the carrier's specific customizations. If the vendor can't produce all three within the first phone call, the integration story isn't real yet.
Question 04
What's the right way to handle bias and explainability for AI in insurance underwriting?
State insurance regulators (Colorado, New York, California, Connecticut) have moved fastest on AI bias requirements in 2024 and 2025. Colorado Division of Insurance Reg 10-1-1 requires testing for unfair discrimination across protected classes for any AI used in pricing or underwriting decisions. NAIC's AI Bulletin (adopted by 20+ states by Q2 2026) requires governance frameworks for AI use across underwriting, claims, and marketing. The clean implementation: maintain a model inventory, run quarterly bias audits with documented holdout testing, preserve a complete decision audit trail per policy, ensure human-in-the-loop sign-off on every adverse decision.
Why most vendors get this wrong: they treat the audit trail as a UI feature and the bias testing as a one-time deliverable. The state DOI exam doesn't care about the one-time deliverable. It cares about the quarterly process and the model-monitoring infrastructure that flags drift before the bias surfaces.
Right answer pattern: documented bias-testing methodology, quarterly audit cadence, holdout sample preserved per audit cycle, model drift detection running continuously, and an escalation path when bias is detected mid-cycle. The vendor that can't produce a bias-testing pack on demand will not survive a state DOI exam.
Question 05
Why have only 7% of insurers scaled AI into production despite 88% planning to?
The gap between AI pilot and AI production at insurers is the largest of any regulated industry. Three reasons: legacy core-system integration (Guidewire and Duck Creek customizations take 18 to 30 months to retrofit), regulatory readiness (state insurance commissioners are moving faster than vendor compliance teams), and change management at the adjuster and underwriter level (the human-in-the-loop layer that was supposed to be temporary becomes permanent).
Why most vendors miss this: they pitch the pilot as if production is a 90-day extension. It isn't. Production at an insurer requires SOC 2 Type II, state DOI notifications in 10+ jurisdictions, integration partner work on the carrier's actual core system version, and a change-management program for 200+ adjusters or underwriters. The pilot-to-production gap is 12 to 24 months, and the vendor that pretends otherwise is going to underestimate the work.
Right answer pattern: the vendor produces a sequenced rollout plan that picks one core process (typically FNOL), gets it to production, proves the audit trail survives a state exam, then expands. The portco that tries to deploy across claims, underwriting, and customer service simultaneously stalls. The portcos that sequence work have shipped.