Operators, not copilots.
Oxynth deploys autonomous agents into modern life & health policy administration systems to run operations end to end — claims, underwriting, servicing. Native to the platforms carriers already run. Starting with FNOL.
See a clean death claim run end to end — documents in, decision out, no adjuster required
04 Jun 2026
Ready for payment
The industry isn't short on AI. It's short on production.
Every carrier has run a pilot. Almost none have moved one into production. The gap between activity and scale is the defining fact of AI in insurance — and it is the gap Oxynth is built to close.
organizations remain stuck in pilot mode, never scaling AI across the enterprise
McKinsey, State of AI 2025 ↗of insurers have successfully scaled AI initiatives across their organizations
Microsoft Cloud, Feb 2026 ↗have taken tangible action — though 90% of insurance executives agree work must be reinvented around human-machine collaboration
Deloitte, 2026 outlook ↗The few who did scale are large enterprises with internally built copilots, or single-step point tools — not third-party operators running mid-market life & health workflows end to end on modern PAS.
The window is open now for two reasons: agentic systems crossed a production-reliability threshold in 2024–25 — not research-grade, but deployable on a real book of business — and modern PAS vendors matured their API layers in the same period, making native agent integration possible without middleware workarounds. Both conditions needed to be true. They are, as of this year.
Modern PAS are event-driven. Claims work is case-driven.
A modern policy administration system processes events. A claim filed is one event; a document received is another; a payment processed is a third. The platform handles each one correctly — and none of them is the case. An examiner needs the whole thread: every event, across time, with policy context and external evidence, ending in one operational outcome.
So carriers build intermediate layers — workbenches, claim desktops, case modules — to translate an event-driven system into case-driven work. That layer is where the manual work actually lives. It is also where the last three years of AI tools were aimed: copilots that make a human faster at operating a workaround.
A copilot speeds up one step. The workflow keeps the same queues, handoffs, and decision gates, so cycle time barely moves — it is bounded by the workflow's structure, not by typing speed. An operator is the different intervention: an agent that owns the workflow between the events, end to end, and hands a human a decision instead of a backlog.
Your workbench doesn't disappear — the manual assembly inside it does. The agent operates your existing systems, including the workbench, as a governed system actor; your examiners keep the same screens to review, override, and decide.
Today — the workbench compensates
queues · handoffs · 3–5 days for clean cases
With Oxynth — the agent operates
minutes to decision-ready · every step logged
Agents that run the workflow — not popups that suggest.
Each agent owns a bounded operation and runs it to completion. The first is FNOL. Underwriting triage and servicing exceptions follow — same architecture, different operation.
Operates the workflow
The agent owns a defined operation and runs it end to end — clean cases straight through, exceptions escalated to your team. No suggestion popups. No human-in-the-middle by default.
Lives inside your modern PAS
Native deployment into modern PAS — OIPA, Majesco, Sapiens, EIS — and the claims, document, and payment systems around them. No parallel stack. No second source of truth to reconcile.
Audited, every step
Every read, decision, and action is logged, structured, and reviewable. Override any step. Reverse any action. Built for regulated operations from day one.
Examiners · Underwriters · Operations
Set policy, approve exceptions, make the decisions, own the outcomes.
The operating layer
Reads from and writes into your systems of record, with a structured audit trail your compliance team can sign off on.
Policy admin · Claims workbench · Documents · Comms
Your PAS stays the system of record. Agents never fork the data.
The first agent: FNOL — built for life & health.
First Notice of Loss is bounded, measurable, and where most of the wait sits before adjudication even begins. And life & health FNOL is not P&C FNOL — it means claims with contestability windows, beneficiary verification, medical evidence, and regulated, compassionate communication. That is the workflow this agent is built for.
- Notification triageClassify claim type — death, last expense, critical illness, disability — and flag contestability and duplicate notifications within minutes of receipt.
- Policy verificationIn force, premiums current, riders, exclusions, contestability window — read directly from the PAS, not re-keyed.
- Document assemblyDeath certificates, physician statements, claimant ID, beneficiary forms — structured into the case file; missing items requested automatically.
- Beneficiary verificationDesignation match, dispute flags, payout instructions captured for the examiner.
- Claimant communicationAcknowledgements, status updates, and document requests in regulated, compassionate language — a coordinator, not a chatbot.
- Examiner handoffComplete case, right queue, full context. Adjudication, payout decision, and settlement remain with your examiners.
classify(type, contestability, dup) ◆ route
02 policy.verify → PAS
read(policy), check(inforce, riders, window) ◆ exception?
03 documents.assemble → case
ocr(certificates), parse(forms) ◆ complete?
04 beneficiary.verify → case
match(designation), flag(dispute) ◆ clear?
05 claimant.acknowledge → comms
status.send(), request.missing()
06 examiner.handoff → queue
assign(skill, load), attach(case)
Carriers have been offered four answers. None closed the gap.
Fixed fee. Fixed scope. Your baseline. 90 days.
A pilot is a fixed-fee engagement on a defined claim volume, measured against your own baseline — no license, no seats, no internal AI team required. In production you pay per claim processed: if the agent doesn't run the work, you don't pay for it.
Connect
Read-only integration to your PAS, document, and comms systems. Scope locked, baseline measured.
Shadow
The agent runs silently against live volume. You compare its output to your team's, case by case.
Supervised
The agent acts; your team approves each step. Thresholds tuned on real exceptions.
Autonomous
Clean cases run straight through within thresholds you set. Exceptions route to your queue.
Weeks 8–12: the agent runs at scale while the three metrics below accrue against your baseline. Day 90: a go/no-go decision on production — made on the data, not on a demo.
Zero-touch FNOL rate
Share of notifications that complete the full pipeline — triage, policy, documents, beneficiary, routing — without a human touch. Target: 65–80% of clean cases.
Time to decision-ready
Median time from notification received to a classified, verified, document-complete case in the examiner's queue.
Completeness at handoff
Share of cases handed to an examiner with no rework — complete documents, structured fields, clean policy state.
All three measure work the agent owns end to end. Adjudication and settlement stay with your examiners — we don't score ourselves on cycle time we don't control. Real numbers will be published after the first two pilots close, including the misses.
Design partners
We're selecting 3–5 life & health carriers on modern PAS for the first FNOL pilots: founder-level attention through deployment, per-claim production pricing agreed and locked before the pilot begins, and a jointly published outcome write-up.
Control is the product.
- Confidence thresholds you set, per step. Below threshold, the case routes to your team — the agent never guesses on your book.
- Override on every step. Any agent action can be reviewed and reversed by your examiners, with the full reasoning trail attached.
- Kill switch. Pause the agent instantly; work flows back to your queues unchanged. No unwind project.
- Immutable audit log. Every read, decision, and action — structured, timestamped, exportable to your compliance team.
- Shadow mode, anytime. Re-validate the agent silently after product, rule, or regulatory changes before returning it to autonomy.
- Document scope, stated plainly. The agent parses certificate format and issuer from the submitted PDF. Cross-check against a national death registry is out of scope for v0 — this boundary is explicit in every pilot agreement.
Built for regulated operations.
- Your data never trains foundation models. Carrier and claimant data is processed for your operations only.
- HIPAA-grade data handling by design for medical evidence and claimant records. Claimant PHI transits only under data-processing agreements; no PHI is used for model training. BAA available to design partners on request.
- Regional hosting options for data-sovereignty requirements, by market.
- SOC 2 Type II readiness program underway; security documentation available to design partners on request.
- Role-based access aligned to your PAS roles — the agent operates as a governed, first-class system actor, not a shared login.
The agent layer is being built by someone who has sat in the seat it takes over.
Arun Mallavarapu
Two decades inside insurance technology — not adjacent to it. Six years leading modern policy administration implementations for life & health carriers, including two end-to-end production deliveries across the Middle East and Africa on systems that run live books of business today.
Before that: operations work with major US health payors, and most recently, shipping production agentic-AI systems. The conviction behind Oxynth comes from building carrier workbench layers by hand, twice — and recognizing that the layer itself is what should be operated by an agent.
Thinking in public — sources always included.
The 7% Problem
Why AI in insurance is stuck in pilot purgatory — and what it will take to scale beyond it. The architectural argument behind Oxynth, with the data.
Tell us who you are. We'll take it from there.
We're selecting design partners and talking to investors ahead of the July 2026 launch. We reply to every message within 48 hours, from a real address.
Insurance operators. If you run claims, underwriting, or servicing at a life & health carrier on modern PAS and want to pilot an agent on a bounded workflow — this is the right form.
Investors. If you back agentic AI in regulated operations or insurance specifically, we'd like the conversation early. investors@oxynth.ai
Everyone else. Press, prospective hires, technology partners — same form, same reply window. Or write to hello@oxynth.ai.
Thanks — got it.
We'll be in touch within 48 hours, from a real address (not a no-reply).