Venu Mathukumalli · 813.600.8927 · Venu.Mathukumalli@gmail.com · linkedin.com/in/venumathukumalli
Case study 01 Deep dive
B2B2C client app Single sign-on for client companies and their employees
An AI workflow for insurance claims triage. Every claim runs the same seven steps in the same order: the AI checks coverage, matches evidence, applies fixed fraud rules and does the payout math, and a licensed adjuster approves every payout.
At a glance
01 Problem
Starting with first-notice-of-loss (FNOL) — the initial claim filing — because it's a well-bounded, document-heavy task with clear inputs and precedent. Why FNOL is the entry point for AI agents in insurance and how a claims verification agent pipeline is typically structured both point to the same conclusion: the volume is real, but so is the risk.
But claims decisions carry real financial and regulatory weight — a wrong payout, a missed fraud signal, or an unexplained denial creates liability and erodes policyholder trust. That makes claims triage a genuine test case for a question the industry hasn't solved cleanly:
How do you let an AI system move fast on the routine 80% of cases while keeping a human demonstrably, auditably in control of every dollar that moves?
02 Impact
This is a scoped concept project, so the business impact is designed and estimated, not measured. I have marked which is which.
What I delivered
Designed, not measured
What I took from it
This concept has not been tested with adjusters, and the claims, amounts and confidence thresholds shown are illustrative. It shows hands-on depth in AI workflow UX: a confidence-gated approval flow, an escalation path and an audit-trail concept, each traced to a documented pattern.
03 Product
An adjuster works a single claim end to end. A senior adjuster or SIU investigator receives escalations, and a compliance reviewer audits the trail afterwards. Select a screen to jump to it.
The AI stages a recommendation for every claim. Nothing executes.
01 Approval queueThe adjuster sees the payout, confidence, fraud risk and the reasoning behind them.
02 Review card04 ReasoningApprove, edit the amount, or escalate to a specialist with full context.
03 Edit05 EscalationOnly an explicit human approval releases a payout.
02 ApproveEvery step is logged as it happens, for compliance and disputes.
06 Audit trailThe screens
Select any screen to view it full size.
The adjuster's entry point. Claims are sorted by confidence, lowest first, so cases that most need human judgment surface immediately. Color and status pills communicate risk without relying on color alone — each row also carries an explicit text label. The sidebar shows the AI workflow: the same seven steps for every claim, with the adjuster’s approval as the one human step.
Confidence-based prioritizationThe centerpiece screen. Before any action executes, the adjuster sees the full AI-proposed decision: payout, confidence, fraud risk, and the specific reasoning behind the number. Nothing is approved automatically, and nothing is approved blind.
Propose → review → commitReal claims rarely resolve to a clean accept or reject. This screen lets the adjuster adjust the AI's proposed payout directly, with a required reason field that becomes part of the permanent record.
Partial acceptanceAn expanded, numbered trace of exactly how the AI reached its recommendation — coverage match, evidence cross-reference, fraud screening, and payout math — each step citing its source.
Reasoning transparencyWhen a claim falls below the confidence floor or trips a fraud signal, it routes to a specialist — arriving with a full context package already attached: the reasoning trace, claim history, and submitted evidence.
Escalation with context handoffA timestamped record of every step in a claim's lifecycle — the AI's initial proposal, when a human opened it, what was changed and why, and who ultimately approved it.
Accountability record04 Experience
Open a claim, approve it, edit the payout, inspect the reasoning, or escalate. Every action is written to the audit trail. Use “Reset demo” inside the prototype to start over.
The prototype is a desktop-sized app, so it opens best in its own tab.
Open the prototype ↗How I got here
This page shows the problem, the impact, the product and the experience. The thinking behind them is in a 18-slide presentation: discovery, research materials, personas, journey maps, ideas, user flows and a validation plan.
I'm happy to walk through the thinking behind this, or my enterprise projects, in more detail.