AI agents that do the work — and leave an audit trail a regulator can read.
Claims intake, submission triage, guideline questions: the document-heavy work that buries adjusters and underwriters. I build agents that handle it inside your own cloud, stop for a named person before anything consequential, and record why every decision was made.
The difference
A copilot waits to be asked. An agent does the work.
Most insurance AI today is a chat window that drafts or summarizes when someone opens it. That saves minutes. The hours are in the queue itself: the FNOL packet nobody has read yet, the broker submission sitting unscored, the guideline question that waits for the one underwriter who knows the answer.
Assisted AI
Answers when prompted. Forgets between sessions. Keeps no record of why. Stops at the chat window.
Forward-deployed agents
Watch the queue and start the work. Run against your real systems. Stop for a named approver. Log every step.
What I build
Three agents, each a fixed-scope pilot
Each agent is a defined build, not an open-ended retainer. It runs on your data, in your tenant, and it is measured against your own team's numbers before anyone scales it.
Claims Intake Agent
Reads the FNOL packet, medical records or loss documents as they arrive. It extracts the facts, checks coverage against the policy, flags missing items and fraud signals, and routes a complete file to the adjuster.
Adjuster approves before any claimant contact or reserve change.
Submission Triage Agent
Turns broker emails, ACORD forms, SOVs and loss runs into a structured submission. It scores the submission against your appetite and guidelines and moves the best-fit risks to the top of the underwriter's queue.
Underwriter approves before any quote or decline goes out.
Guidelines & Policy Q&A
Answers questions about underwriting guidelines, policy forms and procedures, and cites the exact clause for each answer. Tests catch a wrong answer before a user does.
Every answer cites its source. No citation, no answer.
- 01WEEK 1Discover
Choose one queue, one metric and one approver. Get access to a sample of real documents in your tenant.
- 02WEEKS 2–5Pilot
Build the agent with its approval gates and audit log. Run it beside your team, then compare its results with theirs.
- 03AFTERProduction
Harden it, hand it to your team, and add the next queue. The governance layer carries over, so the next agent is configuration, not a rebuild.
Governance
Autonomous work. Human authority.
Regulators now expect insurers to govern the AI they use, including AI from vendors. The NAIC model bulletin, which many states have adopted, asks for documented controls, testing and oversight. These controls are part of the design from the first day, not something added before an audit.
Named-person approval
The agent prepares the work and a specific, role-based person approves it. The platform enforces every gate, so the agent cannot skip one.
Tamper-evident log
Every model call, retrieved document, recommendation and approval is recorded. The log is hash-chained, so any edit to a past entry can be detected.
Your tenant, your keys
Everything runs in your Azure, AWS or GCP account under your identity model. Policyholder data does not leave your cloud.
Measured, not assumed
Before launch, a test set built from your own cases measures accuracy. After launch, the same tests run again, so drift shows up in a chart before it becomes a complaint.
The stack
Built on the enterprise stack you already approved
No proprietary black box. Standard, supported components mean fewer questions in your security review and a system your team can maintain.
- REASONING
Anthropic Claude and Azure OpenAI
Larger models for judgment steps, smaller ones for high-volume classification, called through your cloud's own endpoints.
- ORCHESTRATION
LangGraph · Azure AI Foundry · MCP
Multi-step workflows with saved state, failure recovery and approval checkpoints built into the graph.
- RETRIEVAL
Azure AI Search · GraphRAG · lakehouse SQL
Retrieval I have run over a 120 GB index and a 374,000-node ontology.
- PLATFORM
Terraform · Kubernetes · GitOps · Vault
The infrastructure-as-code discipline of a FedRAMP High environment. Every change reviewed, repeatable, reversible.
Also serving
The same pattern works wherever decisions need evidence
Public sector & defense
FedRAMP High delivery experience. Available as a subcontractor to primes. US citizen, clearance-eligible.
Manufacturing & R&D
Knowledge agents over formulation, materials and lab data for global manufacturers, delivered in the customer's own tenant.
Life sciences
Platform and security engineering for a medical device manufacturer, where regulated data rules decide the architecture.
Why regulated teams hire NikTech AI
Regulated organizations don't need more AI demos. They need someone who can get an agent through security review, into the real workflow, and past the auditors.
I spent fifteen years building the platforms agents depend on before I built the agents. That includes a group insurance carrier's cloud, hardened to pass compliance audits, and infrastructure where every change met federal standards. Identity, permissions, security review and audit questions are where most AI pilots stall. For me, they are familiar ground. You work directly with me, from the first call to the handoff.
Answers from 120 GB of R&D knowledge
Led the agent build on Azure AI Foundry, with retrieval over a 374,000-node ontology and existing models exposed as tools over MCP.
Kickoff to MVP in 16 weeks
Led data and AI engineering for a 29-person joint team, entirely inside the customer's own cloud.
Every change to federal standards
Onboarded applications through Terraform and GitOps and automated secret rotation with HashiCorp Vault.
Built the cloud, then passed the audits
Multi-region Azure and AWS, least-privilege identity, and hardening that passed compliance audits.
Delivered through engagements with Microsoft, Cisco, AMC and The Standard · MS Systems Engineering, Cal State Fullerton · Azure AI Engineer (AI-102), Terraform Associate, KCNA and 3 more certifications · Capability statement →
Questions buyers ask
FAQ
How is an AI agent different from an AI copilot for insurance?
A copilot answers when someone asks it something. An agent watches a queue, such as new FNOL packets or broker submissions, and starts the work itself: it reads the documents, extracts the facts, checks them against policy or appetite, and prepares a complete file. A named adjuster or underwriter then approves before anything consequential happens.
How do you keep an AI agent compliant with the NAIC model bulletin on AI?
The bulletin asks insurers for documented governance, testing and oversight of the AI systems they use, including systems from vendors. Every NikTech agent enforces named-person approval gates, keeps a hash-chained audit log of every model call, document and decision, and is measured with a test set built from your own cases before and after launch. That record is the documentation.
Does policyholder data leave our cloud?
No. The agent runs inside your own Azure, AWS or GCP tenant, under your identity model, and calls models through your cloud's own endpoints. Nothing is copied to a NikTech system.
How long does a pilot take?
About five weeks. Week 1 picks one queue, one metric and one approver. Weeks 2 to 5 build the agent with its approval gates and audit log and run it beside your team, so its results can be compared with theirs. Production hardening follows only if the numbers justify it.
What stops the agent from taking a wrong action?
The agent cannot act on its own. It prepares work, and a specific person approves it. The platform enforces each approval gate. Answers must cite a source document or they are not given, and accuracy tests run continuously so drift shows up before it reaches a customer.
Do you work with public sector and defense programs?
Yes. NikTech AI has FedRAMP High delivery experience and is available as a subcontractor to prime contractors. Nick Harinath is a US citizen and clearance-eligible, and can begin the clearance process when a program requires it.
Growing with our pilots
We're building the bench
NikTech AI is looking for senior AI, data-integration and cloud-security engineers, and for insurance operations leaders who want to advise on regulated agent builds.
Introduce yourselfContact
Start with one queue
Remote nationwide, and on-site across Southern California: Orange County, Greater Los Angeles, San Diego and the Inland Empire. Available for fixed-scope pilots, contract and subcontract engagements, including federal and defense work.
- Email[email protected]
- Websiteniktechai.com
- Based inCorona, California · US citizen, clearance-eligible