How Insurance Teams Are Using Agent Skills in 2026

Insurance Adopted AI Fast — But It's Shallow, Risky, and Unmeasured

Insurance agents have rushed into AI faster than almost any other profession — and their agencies have not kept up. Individual adoption jumped from 37% in 2025 to 65% in 2026, with users reporting an average of four hours saved per week, according to Liberty Mutual's 2026 agent study. That is a genuine productivity shift, and it happened largely under the radar of formal agency governance.

The gap between individual enthusiasm and institutional readiness is stark. Only 14% of agencies have formally implemented an AI tool, just 18% have a well-defined AI policy, and a separate industry survey of insurers found that 42% never measured AI outcomes at all. Meanwhile, trust in AI with client and business data sits at 22%. Agents are experimenting with unvetted tools, on their own time, with no shared rules about what data can leave the building.

Most current usage stays shallow — the "Assist" tier rather than anything resembling autonomy:

• Marketing content generation: 18% of agents
• Summarizing meeting notes: 17%
• Comparing policy details: 13%
• Customer data analysis: 8%

Notably, 26% of agents will not consider AI for taking a customer's initial contact information at all. The closer a task sits to the client relationship, the more resistance it meets — and that instinct is defensible when E&O carriers have not meaningfully updated their products for AI-mediated workflows.

This is the core tension heading into 2026: adoption is real, but depth, governance, and measurement are not. Enterprises elsewhere are learning the same lesson about agent skills — instruction packages that let an AI agent perform a specific task reliably. Published research has found that roughly 26% of them carry at least one vulnerability, which is exactly the kind of risk an agency with no AI policy cannot see, let alone manage. SkillzBench is building a free directory and MCP connector designed to close that gap: every skill pinned to a commit SHA, security-scanned, license-checked, and benchmarked in a sandbox with the same model, with and without the skill, so the Lift in percentage points is published separately from any rating. Agents that only see the skills you choose are considerably easier to govern than agents that see everything.

Where Agent Skills Actually Fit: The Insurance Workflows That Work

The most useful AI in insurance isn't the flashiest — it's the quiet, repetitive paperwork that eats an adjuster's afternoon. That's exactly where agent skills, the instruction packages that teach coding agents like Claude Code, Cursor, and Codex how to handle a specific task, are finding their footing.

Where the numbers point

The adoption data maps cleanly onto what skills do well. Among insurers using AI in claims, 35% use it for summarization and 22% for document processing — both are classic skill territory: feed the agent a claim file, a police report, or a stack of medical bills, and get back a structured summary a human can verify in minutes. A Nordic insurer found 70% of documents fed into its system are correctly extracted and interpreted automatically, freeing agents for customer relationships.

On the agency side, the workflow breakdown tells a similar story:

• Quoting support — 71% workflow adoption among AI-using agencies, the single highest lane
• Lead intake — 58% adoption, where conversational intake cut abandonment roughly in half for early adopters
• Routine data entry — 26% of agents saving 2+ hours weekly automate data entry and document maintenance
• Claims handling — 49%, concentrated in summarization and routing rather than decisions

The human-in-the-loop boundary

Every credible source draws the same line. As one industry analysis puts it, AI agents can prepare and route the claim file, but an adjuster retains payout and denial decisions. The framing among adopting agencies is "AI proposes, human agent disposes."

That boundary matters for how you deploy skills. Skills fit naturally into "Assist"-level work — summarizing, extracting, drafting, comparing — which is precisely where adoption is proven. Only 17% of insurers run agentic AI today, and only 10% of P&C insurers have successfully scaled AI at all. Frontier autonomy is not where the evidence is.

The practical takeaway: build skills for the paperwork, keep humans on coverage, pricing, and payouts. If you want to test that approach, SkillzBench is building a free directory of agent skills — security-scanned, version-pinned, and benchmarked with and without each skill — plus a free MCP connector so your agent only sees the skills you choose. Given that published research found roughly 26% of skills carry at least one vulnerability, verifying before you deploy isn't optional.

The Hidden Risk: ~26% of Agent Skills Carry a Vulnerability

Here's a number that should stop every insurance agent mid-workflow: published research on agent skills found that roughly 26% carry at least one vulnerability. That means about one in four of the instruction packages people are installing into their AI tools has a flaw — and if your agent touches policyholder information, that flaw becomes your problem.

This lands on an industry already anxious about exactly this. E&O liability is the most-cited blocker to AI adoption among non-adopting agencies, and the liability picture hasn't improved: as industry analysis puts it, "the answer in 2026 is still 'the agent'" — E&O carriers have not meaningfully updated their products for AI-mediated workflows. When an AI tool misuses client data or mishandles a policy detail, there's no carrier backstop absorbing the error. The agent bears the risk.

The governance gap makes it worse. Only 18% of agencies have a well-defined AI policy, up from 12% in 2025, meaning most AI use happens with no formal guardrails at all. Meanwhile, just 22% of agents trust AI with business data and client information — an instinct worth honoring. A skill installed casually into an agent that handles renewals, claims notes, or customer contact details is precisely the kind of unvetted exposure that instinct warns against.

The core issue is that skills are easy to install and hard to evaluate. A skill is just a set of instructions your agent will follow, and nothing stops a poorly written or malicious one from directing your agent to leak data, call the wrong tools, or behave unpredictably. Before adding any skill to an agent that touches client information, ask:

• Where did this skill come from, and who maintains it?
• Has it been security-scanned, not just eyeballed?
• Is the version pinned so upstream changes can't silently alter behavior?
• Does it actually improve results, or just add instructions?

Most teams have no way to answer those questions today, which is why SkillzBench is building a free, independent directory that continuously crawls and dedupes agent skills, security-scans and validates each one, pins every version to a commit SHA, and publishes a trust score from security, quality, license, maintenance, and community signals. Its free MCP connector also lets you control exactly which skills each agent can see — a per-agent loadout instead of a grab bag.

The math is unforgiving: with 65% of individual agents now using AI and 83% expecting usage to increase, skill exposure only grows. Vetting isn't optional overhead — it's the difference between an agent that saves you four hours a week and one that creates the E&O claim you spent years avoiding.

A Practical Rollout: Vet Skills, Equip Agents, Keep Humans in Charge

The fastest way to derail an insurance AI rollout is to skip straight to the ambitious stuff. Individual agents are already saving an average of four hours per week with AI, yet only 14% of agencies have formally implemented a tool — and just 18% have a well-defined AI policy, according to Liberty Mutual's 2026 adoption research. The gap between enthusiasm and infrastructure is where mistakes happen.

Start where the evidence already points. The proven "Assist" use cases — summarizing meeting notes (17% current use), generating marketing content (18%), and comparing policy details (13%) — deliver measurable time savings without touching coverage, pricing, or payouts. These are the workflows to pilot first.

Write the policy before you scale. ACT's Chris Cline advises agency leaders to set expectations and formal guidance for AI use before deployment. A non-negotiable rule: never enter personally identifiable information into free, public AI tools.

Then measure, don't assume. Across the industry, 42% of insurers never measured AI outcomes, and most who did couldn't demonstrate clear ROI, per industry AI statistics. A practical rollout looks like:

• Pilot proven Assist use cases with a small group and a defined baseline
• Publish an AI policy covering permitted uses and PII handling
• Vet every tool before it touches client workflows
• Track hours saved, error rates, and review time — not vibes
• Keep humans in charge of coverage, pricing, payouts, and denials

Vetting is the hard part, and it's where SkillzBench fits. SkillzBench is building a free, independent directory of agent skills where every version is pinned to a commit SHA, security-scanned, license-checked, and benchmarked in a sandbox with and without the skill to show Lift in percentage points — plus a free MCP connector and per-agent loadouts so an agent only sees the skills you chose. That gives insurance teams a way to test skills safely before they touch client workflows. SkillzBench is pre-launch; the waitlist is open.

Skills, proven.