Insurance has spent the past year experimenting with generative AI in the most obvious corners of the enterprise: summarising emails, extracting details from claims forms, drafting responses for call-centre agents, and speeding up document-heavy back-office work.
Those are useful gains. But a new whitepaper by Google Cloud and CoverGo argues that they may also be a distraction.
Also Read: OneByZero raises US$20M Series A to help enterprises move AI from pilots to production
The paper, authored by Maxim Afanasyev and Tomas Holub, frames the industry’s current AI moment as a potential “Kodak moment”: a point at which incumbents optimise an existing business model while the market quietly moves somewhere else. As management thinker Oren Harari put it, “The electric light did not come from continuous improvement of candles.” For insurers, the warning is that better workflow automation will not be enough if the fundamentals of claims, underwriting and distribution are being rewritten by autonomous software agents.
The paper points to studies from MIT and BCG suggesting that 95 per cent of organisations struggle to generate tangible returns from agentic AI. Its diagnosis is blunt: many institutions still treat AI as a series of isolated productivity projects, rather than a new operating layer for the business.
For Southeast Asia’s insurance sector, where mobile-first customers, digital wallets, embedded finance and platform ecosystems are already reshaping financial services, that distinction matters.
When every market participant has an AI agent
Agentic AI refers to systems that do more than answer prompts. These agents can plan tasks, call tools, retrieve data, make decisions within defined limits, and interact with other agents or systems. In insurance, that could mean an underwriting agent checking data sources, a claims agent assessing evidence, or a broker agent negotiating terms.
The bigger shift begins when insurers are no longer the only ones deploying these tools.
Customers may use personal AI agents to read policy wording and optimise claims submissions. Healthcare providers, repair workshops, brokers and even regulators may use agents to submit, verify, challenge or audit information in near real time. At that point, insurance stops being a slow exchange of forms between humans and becomes a machine-speed negotiation between multiple software actors.
The whitepaper gives a simple example: if a dental policy reimburses treatment linked to acute tooth pain, a customer’s AI agent may learn to frame a routine visit around that symptom. In underwriting, agents could identify weak spots in pricing models and generate applications that exploit them. In distribution, broker agents may flood the market with automated outreach, while customer agents filter or block them.
The result is not just more volume. It is a change in market behaviour.
Also Read: Health, wealth, and legacy planning converge as new wave of SEA insurtechs emerges
Economists have long discussed signalling: the idea that reliable information is costly to fake. AI weakens that assumption. If polished applications, claims narratives and supporting documents can be generated cheaply, insurers need to think beyond traditional risk scoring. They need mechanisms that encourage truthful behaviour, detect strategic manipulation, and decide when a human should intervene.
Why Southeast Asia cannot treat this as a distant problem
In mature Western markets, insurers often carry decades of legacy infrastructure and rigid internal rulebooks. Southeast Asia has its own constraints, including fragmented regulation, uneven data quality and protection gaps, but it also has a different adoption curve.
The region’s consumers are young, mobile-first and used to digital financial services moving quickly. In markets such as Singapore, Indonesia, Vietnam, Thailand and the Philippines, customers increasingly expect insurance to be embedded into payments, e-commerce, lending, travel, logistics and healthcare journeys. The next insurance interaction may not start on an insurer’s website. It may start inside a superapp, a hospital system, a broker platform or a small business finance tool.
That makes the agentic shift especially relevant. If every partner in the value chain deploys AI, insurers cannot rely on isolated internal pilots. A claims chatbot here and an underwriting assistant there will not be enough. The core systems must be able to coordinate agents, enforce authority limits, maintain audit trails and connect with external systems safely.
This is where the paper introduces the idea of an Agentic Core Operating System.
From AI tools to an agentic operating layer
The argument is not that insurers should replace humans with autonomous machines. It is that they need a technology foundation where human experts and AI agents work as coordinated teams.
In practical terms, this means moving away from monolithic, hardcoded systems towards modular architecture. One building block highlighted in the paper is the Model Context Protocol (MCP), often described as a “USB-C for AI”. The idea is to give agents a standardised way to access context from systems such as customer databases, policy administration platforms and claims records, without building a custom integration every time.
Other emerging layers include Agent-to-Agent protocols, which allow different agents to exchange information and negotiate tasks, and agentic payment protocols that could enable transactions under strict controls. Workflow engines based on directed acyclic graphs can help manage the recurring loop of plan, act, observe and reflect that agentic systems use.
Also Read: Why Southeast Asian enterprises need AI governance before scaling generative AI
For many Southeast Asian insurers, the appeal is clear. Legacy replacement projects are expensive and risky. A protocol-driven architecture could allow incumbents to build around old systems first, rather than rip everything out at once. But the governance challenge is just as important as the technology.
The new failure modes
A traditional AI tool can make a bad recommendation. A network of agents can create system-level failures.
The whitepaper identifies several risks. One is an agent loop, where two or more agents keep triggering each other without resolution. Another is agent disagreement, where different agents reach conflicting conclusions on the same case. A third is authority violation, where an agent acts beyond its approved mandate.
These risks matter in insurance because decisions affect money, compliance and trust. A wrongly denied claim can damage a customer relationship. A mispriced portfolio can create financial losses. A poorly governed automated process can draw regulatory scrutiny.
The proposed answer is a schema-driven agent model, where business rules, authority limits and escalation triggers are defined in configurable schemas rather than buried in code. Human-in-the-loop controls remain central: confidence scoring, mandatory escalation for sensitive cases, audit logs, and clear accountability for decisions.
For regulators in Southeast Asia, this could become a key question. The issue will not simply be whether an insurer uses AI, but whether it can explain how autonomous decisions are made, challenged and reversed.
The road beyond pilots
The paper’s modernisation roadmap is deliberately incremental. It begins with intelligent document processing to extract logic from legacy documentation, then separates solution design from execution so business and technology teams can validate workflows before coding. New agentic platforms can be built in parallel with legacy systems, master data can be loaded first, and policies can gradually move to the new core before old systems are retired.
That may sound unglamorous, but it addresses a real problem in insurance technology: transformation often fails when companies try to jump straight from legacy complexity to a grand new platform.
Also Read: What I’m learning about the second wave of insurance digital transformation in Indonesia
The deeper message for founders, investors and operators is that AI strategy in insurance can no longer be measured only by productivity gains. The question is whether the business is prepared for a market where customers, intermediaries and service providers also have intelligent software acting on their behalf.
Peter Drucker’s warning remains apt: “There is nothing so useless as doing effectively that which should not be done at all.”
For Southeast Asia’s insurers, the risk is not that they fail to automate enough tasks. It is that they automate yesterday’s business model while the market shifts to a new one. The next phase of insurtech will belong to those that build not just AI use cases, but the operating system for an agentic economy.
The post SEA’s insurers face a new question: what happens when customers have agents? appeared first on e27.