Agentic AI: Human in the middle
Agentic AI represents a new wave of intelligence for insurance, moving from automation to autonomy, and can help in tasks that range from dynamic underwriting to real-time claims orchestration.
By Anandi Iyer

Since the 1990s, insurers have leveraged AI for automation streamlining claims and underwriting processes. Over the past two decades, it has moved from relying on rules-based to machine learning-driven insights, culminating perhaps in the recent surge of generative AI to enhance customer interactions and decision making. However, agentic AI arguably represents the next wave of AI, a true transformation, not just an incremental improvement.
Flood risk as a use case
Celent senior analyst insurance Karun Arathil said floods present an ideal proving ground for agentic AI. They are geographically complex, require fast decision-making and expose insurers to both operational strain and reputational risk.
“In a flood event, an agentic AI system can immediately begin to analyse rainfall intensity, river levels and satellite imagery to identify at-risk policyholders. It can launch proactive communications, update exposure models and even begin assembling the data required for automated claims initiation – all without waiting for a formal notification of loss,” Mr Arathil said.
“This kind of coordination across underwriting, claims and customer service has traditionally been impossible to manage manually at scale. But with autonomous agents embedded across business units, insurers can operate with something far more powerful than automation: orchestration,” he said.
What makes agentic AI different
While conventional AI models typically assist with tasks like document classification or chatbot queries, agentic AI is goal-driven. It works toward business objectives, whether reducing loss ratios or speeding up claims resolution, by reasoning through complex scenarios and choosing actions dynamically.
He said agentic AI enables insurers to respond to catastrophes in real time by deploying specialised sub-agents – for underwriting, customer communication, and fraud detection – that work independently but share context and rules.
Unlike RPA, which is rigidly rules-based, agentic AI can assess situations and activate the right sub-agent. Compared to generative AI, which risks hallucinations, agentic AI incorporates logic, reasoning, and verification, making it more reliable for secure, enterprise-level decision-making.
Early signs of impact
The early promise of agentic AI is already showing through pilot deployments, said Mr Arathil.
He cited recent Celent research saying, “Insurers in the US expect the greatest benefits to emerge in claims processing (69%), underwriting (66%) and customer engagement (59%). Flood response scenarios are a natural extension of these capabilities.”
While fully autonomous AI may still be years away from widespread use, these early-stage systems – often termed ‘agents on rails’ – are constrained by rules, approvals and oversight. Yet even within these bounds, they are already demonstrating value. This staged approach marks a transition from ‘human in the loop’ to ‘human in the middle’, a paradigm where AI performs the heavy lifting, but humans remain supervisory anchors.
Flood resilience in Asia and the Middle East
Mr Arathil said that the application of agentic AI takes on urgency in Asia and other parts of the world where flood risk is rising sharply but operational maturity varies widely.
In regions with fast-growing insurance penetration and digital infrastructure still in development, agentic AI may help insurers leapfrog legacy systems altogether.
For example, a Gulf-based insurer could use agentic systems to track real time rainfall thresholds and link them to parametric triggers, enabling faster payouts in flood-prone urban centres. In Southeast Asia, insurers might deploy multilingual AI agents to assist rural customers with claims filing during seasonal monsoons, reducing delays and alleviating pressure on overstretched call centres.
The way forward in these regions is to start with internal-facing use cases, such as claims triage, portfolio monitoring, exposure modelling, before moving toward fully autonomous customer-facing systems. It’s not about replacing humans; it’s about building adaptive layers of intelligence that support them.
Tipping points
Widespread adoption will depend not only on technological capability but also on trust, governance, and return on investment. The biggest barriers remain: fragmented legacy systems that make cross-functional coordination difficult; inconsistent data quality, especially in emerging markets; and customer unease with the idea of AI making claims or pricing decisions.
There is also concern that, like any digital system, agentic AI could be subverted for fraud – with malicious actors attempting to manipulate inputs or exploit AI behaviour.
Yet, the long-term direction is clear. As the industry’s exposure to climate, cyber and systemic risks grow more complex, agentic AI offers something traditional systems cannot: adaptability at scale.
“For risks such as floods, this shift is not just about speed or efficiency. It’s about resilience. The insurers who succeed will be those who invest not just in automation, but in adaptive intelligence, systems that can think, act and evolve in the face of a changing world,” he said.