Agents, Agents Everywhere: Governing the Next Era of AI in Insurance

By Matthew Twist, VP Sales EMEA, Earnix

Agentic AI is gaining considerable momentum, alongside growing concern among insurers and regulators about autonomy, accountability and control. Technology providers need to address those concerns directly, moving the conversation beyond the novelty of AI agents to the governance required to deploy them safely in regulated environments. For insurers, trust, human oversight, explainability and deep industry context will be defining success factors. That makes the choice of technology partner particularly important. AI deployed within critical insurance workflows needs to be built with the realities of insurance in mind from the outset.

That timing matters because enterprise adoption of agentic AI is accelerating rapidly. Industry reports predict significant growth in the number of AI agents deployed by large enterprises over the next several years, even as many organisations acknowledge that their governance frameworks are not yet ready for that scale.

Organisations are moving quickly from experimentation toward production, while governance frameworks are still catching up.

It is tempting to see this proliferation as straightforward progress. More agents can mean greater productivity, more automation and faster innovation. They can also create more opportunities for inconsistency, unintended consequences and loss of control.

There is another risk too: agent proliferation can recreate the same silos insurers have spent years trying to remove. An underwriting agent, pricing agent and customer agent may each perform their individual tasks well, but that does not mean their decisions are aligned. As insurers deploy more agents, the challenge becomes not only how to govern each one individually, but how to orchestrate them around shared business rules, risk appetite and objectives.

Much of the conversation around AI agents still focuses on whether they are capable enough. Increasingly, the more important question is whether they are governable. In highly regulated industries, capability without appropriate controls creates risk.

Insurance illustrates this particularly clearly. Pricing adjustments, underwriting decisions, fraud investigations and customer interactions can all carry financial, legal, regulatory and reputational consequences. Decisions need to be accurate, but they also need to be explainable, auditable and demonstrably consistent with an insurer’s rules, appetite and obligations. An AI agent that schedules a meeting incorrectly creates an inconvenience.

An AI agent that recommends or executes an underwriting or pricing decision without appropriate controls creates an entirely different category of risk. Insurers need to resist the assumption that an agent capable of sophisticated reasoning can automatically be trusted inside a mission-critical workflow. Intelligence is only one part of the equation. The agent also needs the right context, constraints, permissions and accountability.

The objective should not be maximum autonomy. It should be appropriate autonomy. Insurers need the flexibility to determine where an agent can recommend, where it can execute, where it must escalate and where a human must always make the final decision. Those boundaries will differ according to the decision, the level of risk and the business context.

That requires governance to be built into the architecture rather than added after deployment. Autonomous decisions need to be traceable and explainable, with clear mechanisms for challenging, stopping or overriding an action. Human oversight needs to be an active part of the decisioning process, particularly where financial, regulatory or customer outcomes are at stake.

Governance should not be seen as the brake on agentic AI. In a regulated industry, it is what makes greater autonomy possible. When permissions, decision rights, traceability, escalation and auditability are built into the decision flow from the outset, insurers can give AI greater scope to act without giving up control.

This is particularly important in insurance, where domain expertise cannot simply be inferred from a general-purpose large language model. A general-purpose model does not inherently have access to an insurer’s specific underwriting appetite, pricing strategy, regulatory obligations, portfolio objectives or operational processes.

Insurance-native context goes beyond terminology or industry data. It includes the business logic surrounding the decision: how underwriting appetite interacts with pricing, how a recommendation affects portfolio performance, which regulatory requirements apply, and where human authority begins and ends.

Agentic AI for insurance therefore needs more than a capable model. It needs an orchestration layer that connects AI with the insurer’s data, models, rules, workflows and governance framework, while controlling what each agent is permitted to see, recommend and do.

It also means resisting the idea that every problem needs an agent. Insurance decisions require different forms of intelligence. Predictive AI may identify a risk or behavioural signal, generative AI may interpret or explain it, and an agent may execute the next steps. The real challenge is orchestrating the right intelligence around the right decision, with the appropriate level of human involvement and control.

That distinction matters because the value of an insurance agent should not ultimately be measured by how many tasks it completes. It should be measured by whether it improves the decisions that determine risk, growth, profitability and customer outcomes.

Before giving an AI agent permission to act, insurers should be able to answer some fundamental questions:

• Can we understand why it made a decision and reconstruct that decision afterwards?
• Who determines what the agent is, and is not, permitted to do?
• Where is human review or approval required?
• Is there a complete audit trail of the data, models, rules and actions involved? 
• Can its behaviour and performance be monitored continuously in production?
• What happens when the agent encounters something outside its remit, or its confidence falls below an acceptable threshold? 
• Can an action be stopped or overridden?
• Is the agent operating with the insurance-specific context required for the decision, including underwriting appetite, pricing rules, regulatory obligations and portfolio considerations?
• Can we measure whether the agent’s decisions are actually improving the business outcome it was deployed to influence?
• Can we understand how one agent’s actions affect decisions elsewhere in the insurance lifecycle?

These are not questions to ask once during procurement and then file away. Governance needs to remain active throughout the agent’s operation: clear permissions, explainability, traceability, continuous monitoring, escalation to humans and the ability to stop or override an action. In insurance, those should be the entry requirements for autonomy.

They are also the requirements for scale. An agent that performs well in a controlled pilot may still fail when deployed across multiple products, markets, workflows or regulatory environments. Production-ready agentic AI needs to behave consistently, remain governable as conditions change and operate within the same enterprise decisioning framework as the rest of the business.

The enthusiasm surrounding agentic AI is understandable. The opportunity to improve how insurers analyse information, support decisions and respond to changing market conditions is significant. Realising that opportunity will depend on whether insurers can introduce greater autonomy without weakening accountability.

But accountability is only part of the equation. Insurers also need to avoid creating a new generation of disconnected AI tools. As agents become more capable, the competitive advantage will come from coordinating intelligence across pricing, underwriting, claims and customer engagement so that individual actions support broader portfolio and business objectives.

The next era of AI in insurance will not be defined by which insurer deploys the most agents. It will be defined by which insurers can turn greater autonomy into better decisions without losing accountability. That requires insurance context, intelligent orchestration and governance to move together. 

That is the challenge we are focused on at Earnix: bringing predictive, generative and agentic AI together within an insurance-native orchestration environment, so intelligence can move into production decisioning with the control the industry requires. As agentic AI moves from experimentation into core insurance workflows, the combination of intelligence, orchestration and control will become increasingly important. We will have more to share on how Earnix is approaching that challenge soon.