What if your computer didn't just help you type or access YouTube but acted like a smart assistant instead? That's what we broadly call agentic AI. Some might think AI is here to help us improve productivity and reduce costs, but the change is much deeper (Afanasyev, 2026). It defines what the insurance industry would be in five years and how your organization should adapt.
Asia's First-Mover Advantage and Early Missteps
Asia is the first-mover in AI adoption among financial institutions, compared to developed markets often hindered by tech debt and bureaucracy. The learnings here uncover that, for the industry, agentic AI is a change in the "rules of the game."
Asian financial institutions originally rushed into Agentic AI by treating it like a project focused on multiple use cases, prioritizing 'impact vs. feasibility'. This approach targets quick wins, assuming the organization is the only player implementing AI and hence making the journey risky and expensive. No surprise to see both academia, MIT (Challapally et al, 2025) and consultants, BCG (Apotheker et al, 2025), report that 95% of organizations struggle with Agentic AI. Such organizations miss that their customers and partners are also deploying AI Agents (Citrini and Shah, 2026).
When Productivity Gains Backfire: The Email Case Study
What happens when Agentic AI use cases are assessed without context? Take AI for emails, a productivity use case deployed three years ago, which used AI to help write emails or summarise the received emails. It seemed like a win but lacked context; certain employees used AI Agents to imitate work and secure visibility by generating volumes of emails. AI-generated content is hard (with some researchers claiming impossible) for a human or an algorithm to distinguish from genuinely important messages. The recipients would have to apply AI Agents to search for information. Both senders and receivers myopically reported AI bringing gains, although the insurers likely deployed a use case that resulted in miscommunications and productivity drains.
The New Reality: Everyone Has AI Agents
The right approach for an insurer deploying AI is to accept that others are doing the same, so focusing on internal productivity may not be enough. Think about claims. Your policyholders' and providers' AI Agents are scrutinizing policies and preparing documents to maximize approval. Recall your dental insurance policy, which offers reimbursements only if certain symptoms, such as tooth pain, are present. The AI Agent ensures that the symptom is stated, even when the policyholder visits dentists for routine checks.
In underwriting, similar to how Mythos from Anthropic identifies cybersecurity vulnerabilities (Bloomberg, 2026; Azhar and Williams, 2026), your customers' agents might be identifying gaps. Insurers whose underwriting processes do not handle AI-generated applications may end up underpricing risks. For wealth management, why would a customer having access to LLM models pay advisors who use generic LLMs for generating ideas?
Think about distribution, where an insurance broker deploys AI Agents to identify and message prospects. The competing brokers are doing it too, leading to overwhelmed prospects who might avoid purchases altogether.
The latter is highly illustrative and applies in reverse to a financial institution thinking about its agentic strategy. Previous breakthroughs, such as electricity and the internet, accelerated business but the ability to make intelligent decisions broadly remained intact. AI Agents generate volumes of persuasive, credible content but the overload reduces decision-making ability.
The "Business First" Approach: Three Pillars for Success
Asian financial institutions concluded that to succeed, they need to shift to the "Business First" or "Applied AI" approach, which combines three views—Business domain, Transformation, and Technology—under a single leader (Afanasyev and Milind, 2026). McKinsey used to name leaders with such skills as "Advanced Analytics Translators" and, since the early AI days, has been urging organizations to grow leaders who "ensure that organizations achieve real impact from their analytics initiatives" (Henke, 2018).
Pillar 1: Business Domain — Reimagining Market Forces
The Business domain reimagines how market forces will change, how the industry will provide value, and what an organization's role will be in the long run. This is the least-discussed pillar and precisely what puts organizations at risk.
The majority of Asian organizations are privately controlled, meaning limited dependency on reporting cycles and an ability to prioritize long-term needs. Rapid economic changes over the past two decades have led them to embrace practicality and challenge the status quo. For instance, in developed markets with high labor costs, many insurers seem to be prioritizing Agentic AI technology to improve efficiency and cut costs. Most Asian markets are, however, known to have low labor costs, and local institutions have rightly questioned us about deploying AI for just operational efficiency in the regional context. However, young people, the largest segment of Asia's population, adopt AI Agents for their financial needs the fastest. This has accelerated economic changes not yet noticeable in developed markets.
How Generative AI Disrupts Insurance Fundamentals
We have discussed how Generative AI disrupts signaling theory, which is foundational for modern insurance. In some cases, the policy buyers may know more about the likelihood that they will suffer a loss than the insurance company. The 'market for lemons' theory shows that only the ' worst ' buyers (expecting larger, more likely losses) may end up interested in a policy. Signaling theory explains how parties with superior information (actors, for example, policyholders) convey attributes to parties with less information (decision-makers, for example, policy underwriters) to overcome information asymmetry. For the industry to function, credible signals—think medical records for life insurance—must be costly to imitate. Generative AI has empowered any actor to generate strong signals, disrupting the fundamentals of risk-taking, with empirical studies already demonstrating an erosion of trust.
How can actors credibly signal quality when signals can be cheaply fabricated? The answer, we believe, is grounded in mechanism design, the economics discipline concerned with a decision-maker designing the rules so that rational actors reveal truthful attributes even when their interest might be in using Generative AI. The Applied AI leader has to possess deep insurance domain knowledge and also a practical level of academic perspective. Such a leader analyses disruptions to come up with a view on what the organization's role would be. The outcome could be that the value proposition should be in risk prevention instead of risk pricing; changing target customer segments; vertical integration bringing certain services in-house; or leveling AI to address moral hazards.
Pillar 2: Transformation — Rethinking Processes and Culture
Transformation should focus on agility, changes in processes, decision-making, and human roles. "A growing number of studies reveal that human–AI systems do not necessarily achieve better results than the best of humans or AI alone" (Vaccaro et al, 2024). Many organizations ground their AI performance evaluation on efficiency, accuracy, or productivity. However, "nowadays, productivity gain is no longer the single evaluation criterion. In many instances, computer systems are expected to enhance our creativity, reveal opportunities and open new vistas of uncharted frontiers" (Avital et al, 2009).
Legacy institutions have to unwind certain cultures religiously followed for years. Certain advancements have ensured that employees follow rule books and processes prescribed by management. As a result, many organizations operate as factories, with processes equivalent to production lines of the early 1900s. Employees and mid-management have limited or no discretion outside of what is prescribed. This practice might have been appropriate earlier, but the Agentic AI economy is unpredictable; rule books and processes become outdated in the blink of an eye.
Pillar 3: Technology — Building the Infrastructure for AI Agents
The Technology view is most discussed and depicts how non-human resources are augmented by AI Agents. The learning from Asian progress in Agentic AI is a need for Applied AI leaders to take a broad view on financial industry technology and ground hot topics, such as tech debt, data quality, MCP and A2A, Agentic AI workflow designs, etc.
Asia's Unique Technology Landscape
Asia is home to 5 billion people—about 10x of Western Europe and US populations combined—with interactions among various ethnic groups who have their distinct languages, cultures, and business styles. Asia accounts for about 50% of non-cash transactions globally (Capgemini, 2025), with most countries, including India and China, already cash-free and unimaginable GDP growth in developed markets. Asian institutions deal with unprecedented transaction volumes and diversity, so scale and agility have traditionally been prioritized over politics and offering sophisticated products. Moreover, many overly complicated software products adopted in developed markets for the past two decades found little relevance in Asia.
Before LLMs and Agentic AI, the Asian tech ecosystem had been focused on developing local software solutions to handle volumes. Native software providers have mastered agentic workflow designs using Directed Acyclic Graphs with state management, allowing agents to "Plan, Act, Observe and Reflect". MCP, the "USB-C for AI", has landed in Asia as a standardized interface, allowing agents to pull real-time context from sources like SQL databases, CRM systems, and core banking. For developed markets, MCP would solve the chronic pain point of tech debt by creating an abstraction layer over legacy core systems. Complementing this is the A2A protocol, which facilitates peer-to-peer coordination, and AP2, empowering agents to do financial transactions with or without human presence.
Architectural Challenges: Making AI Agents Work Together
Most legacy systems were architected for workflows that functioned adequately for rule-based operations. Meanwhile, Agentic AI demands open interfaces where workflows are not hardcoded and any given process can be decomposed, assigned to an agent, and reassembled. Organizations are adopting AI capabilities in separate departments and in relative isolation. The containment limits risk, allowing teams to demonstrate value; the actual issue is how these modules function as a coherent whole.
Agent loops, where Agent A triggers Agent B which re-triggers Agent A, require modern architectures to include circuit breakers that terminate chains exceeding defined depth or duration thresholds. Agent disagreements surface when two agents with complementary mandates reach different conclusions. Authority boundary violations occur when agents act beyond their defined scope.
Data Governance and Schema Management
A schema-driven model is recommended. Insurers should adopt a master schema configurator that allows users to govern agent behavior without code. Insurance is challenging in the volume of its unstructured data. Decades of paper-based documentation and inconsistent digitization have left organizations with data states that cannot be queried in real time. For Agentic AI to function, three areas are foundational. Vector databases enable semantic search across document corpora, allowing context retrieval. Structured extraction pipelines convert unstructured documents into queryable records. Master data management maintains reference datasets which can be used for real-time lookup, matching and adjudication.
Regulatory Safeguards: Ensuring Accountability and Control
As regulators, we are of the view that autonomous systems making customer-impacting decisions must be auditable, explainable and interruptible (IAIS, 2024). To ensure agents do not overstep authority, safeguards must be included. Confidence scoring requires that every AI output carry a numerical confidence estimate, with automatic escalation to human staff when results are low. Mandatory review workflows gate high-stakes decisions behind human approval that cannot be bypassed. Continuous feedback loops mean that each correction trains the models. Comprehensive audit trails provide thorough records.
A Five-Step Modernization Roadmap
A five-step modernization approach is required:
Step 1: Start with structured discovery. Most attempts fail because organizations jump to engineering straightaway. Insurers must produce a clear blueprint of their processes. This discovery phase should identify opportunities to separate capabilities into components — units that can become plug-and-play services with open APIs, allowing future agents and partner systems to come in flexibly.
Step 2: Separate design and validation from execution. Insurers should introduce a solutioning phase where a small team defines the future state by determining which capabilities will be rebuilt or replaced, how workflows should operate, etc. Only once the design has been validated should engineering teams begin executing. AI systems can assist by extracting undocumented patterns from interviews, translating requirements into architectures, and simulating potential models to stress-test assumptions.
Step 3: Build in parallel. The target platform should be developed independently while legacy systems continue running operations. A specific sequence is recommended — load non-member-facing master data first, accepting a short window (as brief as possible) during which updates may need to be reflected in both legacy and new systems. Once it's loaded, validate the setup by running test policies.
Step 4: Migrate data through controlled, incremental processes. A single large cutover concentrates risk and creates extended recovery windows if things go south. Hence, a controlled migration strategy is beneficial. Bring the new system online and then migrate data in staggered batches. Not all data needs to migrate; historical data is often better suited for archival storage or data lakes. Before migration, the new platform should be tested in lower environments. Once validated, the system can go live while legacy systems continue managing the existing portfolio. AI systems support this by inferring undocumented schemas, mapping fields, and extracting structured data from records.
Step 5: Retire legacy systems. The final, frequently overlooked step is decommissioning legacy systems. Leaving them partially operational introduces unnecessary cost and risk. Instead, insurers should archive historical data where necessary, ensure compliance for record retention, and fully decommission said infrastructure.
The Kodak Moment: Adapt or Become Obsolete
Organizations focusing on productivity use cases miss the big picture, but the insurance industry will likely not exist in its current form in five years. Think about Kodak 30 years ago: a leader in optimizing film production costs that missed how digitalization changed preferences. In this democratization, insurers need to shift to business models redesigned to cater to change and structure their AI journeys accordingly.
For the full white paper from which this article is adapted, click here.
