In my last article, I unpacked four defining characteristics of agentic artificial intelligence (AI) — autonomy, probabilistic reasoning, proactivity, and human‑in‑the‑loop design and grounded them in insurance scenarios. That definition work matters, but most leaders quickly ask a more practical question: How autonomous should my AI actually be for this decision?
To answer that, it helps to move from a yes/no debate about autonomy to a spectrum. This article introduces a five‑level autonomy framework tailored to insurance, then tackles the key question: Where does “agentic” actually begin?
A Revised Autonomy Framework for Insurance
Multiple frameworks now classify AI systems by autonomy level, inspired by the Society of Automotive Engineers levels for autonomous vehicles. Having reviewed these through an insurance lens, the industry may find a five‑level model useful with one key twist: separating agentic reasoning from operational agency.
Autonomy Levels in Insurance Context
| Level | Type | Description | Insurance example | Agentic? |
| Level one | Assisted | – Human prompts; AI suggests – Trigger for execution is 100% human | – Chatbot suggests answers to FAQs – Copilot auto-completes underwriting notes as typed | No: tool‑like, reactive |
| Level two | Guided | – AI plans multistep actions and proposes a course of action – Human must approve each consequential step before execution | – System assembles a complete underwriting package with data enrichment, risk scoring, and recommended terms – Underwriter must approve before binding | Mixed: agentic reasoning, constrained execution |
| Level three | Semi-autonomous | – AI executes multistep tasks and self-corrects within boundaries – Escalates when outside parameters | – Straight‑through claims processing: first-notice-of-loss intake, coverage verification, damage assessment, and settlement for low‑complexity claims, escalating complex losses | Yes: operational agency within guardrails |
| Level four | High autonomy | – AI operates independently within a domain – Human monitors on‑the‑loop and intervenes by exception | – Continuous portfolio monitoring detects geographic accumulation risk, rebalances exposure, and adjusts pricing – Weekly underwriting review | Yes: broad, domain‑level agency |
| Level five | Full autonomy | – AI sets subgoals and operates without real‑time oversight | – Largely theoretical in insurance – Regulatory requirements and fiduciary obligations create structural barriers for critical decisions | Theoretical in regulated insurance |
Where “Agentic” Actually Begins
In the broader AI field, genuine agentic behavior generally begins at level three in systems that can plan and execute actions within constraints, without per-step human approval. Level one plainly fails this test; it is sophisticated tooling that responds when asked.
Level two is a gray area. These systems demonstrate AI agent‑like reasoning: They can interpret context, decompose goals into steps, select tools, and assemble a plan. However, nothing actually happens until a human approves each consequential step. That’s why level two in an insurance context is referred to as “agentic reasoning, constrained execution.” It’s akin to a new underwriter who can analyze risk well but lacks binding authority: The analysis is agentic, but the agency to act is still withheld.
For the purposes of this framework, level three is treated as the practical starting point for full agency, while acknowledging that many insurance systems at level two are cognitively agentic but procedurally constrained by design. That distinction matters when evaluating vendor claims: Not every product labeled “agentic” has meaningful operational autonomy.
How Insurers Can Use the Autonomy Spectrum
This autonomy spectrum is not an academic exercise. It gives insurance leaders a way to:
- Classify existing and proposed AI initiatives by autonomy level
- See where current systems truly sit versus where marketing labels place them
- Decide where on the spectrum it is safe and valuable to move for specific decision types
For example, a carrier might decide on the following applications:
- Personal auto glass claims can safely move from level two to level three or even to low-end level four within tight guardrails.
- Midmarket commercial property may stay at level two for the foreseeable future, with strong agentic reasoning but human‑in‑the‑loop approvals for binding.
- Complex casualty placement or suitability determinations for complex investment and insurance products remain squarely at levels one or two for now, with AI agents supporting but not replacing human decision-makers.
In the final article in this series, I’ll shift from classification to action—how to match autonomy to consequence, design oversight patterns by decision type, and the questions leaders should ask vendors and internal teams before labeling any system “agentic.”
