End-to-End Agentic Claims Workflows in Indian Insurance: FNOL to Settlement With Human-in-the-Loop Gates

Indian non-life insurers in 2026 are no longer deploying agentic AI as a single autonomous claims handler. The architecture that has emerged from eighteen months of production experience is a chain of five specialised agents, each scoped tightly to one stage of the claims lifecycle, with explicit hand-off gates and human-in-the-loop checkpoints between them. The pattern is now visible at Bajaj Allianz, HDFC Ergo, ICICI Lombard, Tata AIG, and Go Digit across motor and SME commercial lines, with health and large commercial commercial property following in pilot.

The five agents in this chain are:

• the FNOL intake agent, which receives the first notification of loss and produces a registered claim record
• the triage agent, which assesses claim type, complexity, and routing
• the surveyor-dispatch agent, which assigns and schedules survey resources
• the settlement-recommendation agent, which produces a reasoned settlement proposal for human approval
• the audit agent, which reviews the closed claim for adherence to authority limits, policy wording, and regulatory requirements

The design separation matters. A monolithic agent that attempts to handle FNOL through settlement in a single reasoning chain produces failure modes that are difficult to debug, expensive to govern, and unacceptable to IRDAI's audit expectations under the Information and Cyber Security Guidelines 2023. A chain of narrow agents, each with a defined input contract, output contract, and tool scope, produces failure modes that are isolatable, reviewable, and recoverable.

The economic case is meaningful. Indian general insurers handle approximately 4.5 crore claims annually across motor, health, and commercial lines, with claims operating expenses running at 8 to 12 percent of incurred claims. Even a 25 percent reduction in operational cycle time on the routine portion of this volume, which is what the chain is delivering in production, represents annual savings in the range of INR 2,500 to 4,000 crore across the industry. The savings come not from headcount reduction (Indian claims operations remain headcount-constrained) but from cycle-time compression and from freeing experienced claims staff to focus on the 30 to 40 percent of claims where human judgement is the binding constraint.

The regulatory motivation is equally direct. The IRDAI claim settlement timeline circulars require specific cycle-time performance, with metrics tracked at the insurer level and reported in the IRDAI Annual Report. The Protection of Policyholders Interest Regulations 2024 add documentation and communication standards that have raised the operational bar for every claim regardless of size. Insurers that have built the agent chain are using it to meet these standards at scale; insurers that have not are absorbing the operational cost in their expense ratios or risking regulatory observations on cycle-time non-compliance.

The FNOL intake agent is the entry point. It receives a loss notification through any of the insurer's channels (call centre transcript, broker portal submission, customer self-service portal, WhatsApp via an IRDAI-approved Business API integration, email) and produces a structured claim record in the policy administration and claims management systems.

The agent's tool scope is restricted to read-only access on the policy administration system, write access on the claims registration endpoint, and outbound communication through the insurer's standard channels. It cannot approve coverage, set reserves, or commit the insurer to any liability position. Its hand-off output is a structured payload containing the policy number, the policyholder identifier, the loss date and location, the reported peril, the preliminary loss estimate (where provided), the claim type code, the channel of intimation, and a confidence score for each extracted field.

The tool-calling architecture matters at this stage because the FNOL flow involves at least four external systems. The agent calls the policy administration API to confirm the policy exists and is in force on the loss date. It calls the underwriting system to retrieve the schedule, covered perils, and applicable endorsements. It calls the claims management system to register the claim with a unique reference. It calls the communication system to acknowledge the loss to the insured with the assigned claim reference. Each call is logged with the full request payload, the response payload, and the model reasoning step that triggered it.

The gate between intake and triage is structural. A claim cannot proceed to the triage agent until the intake agent has produced a complete payload with confidence scores above thresholds. Low-confidence FNOLs (typically defined as any FNOL where one or more required fields was extracted at less than 90 percent confidence) are routed to a human claims executive for confirmation before triage proceeds. Tata AIG's deployment reports that approximately 62 percent of motor FNOLs pass through intake without human intervention, while approximately 38 percent require a confirmation pass for one or more fields.

The FNOL intake agent is also the gateway for DPDP Act 2023 consent capture. Every loss notification involves processing of personal data, and the agent must capture the policyholder's consent for the claim handling purpose at the point of intake. The consent record, with timestamp and the language used, is logged alongside the FNOL payload and retained as part of the immutable claim record. Insurers that have built the consent capture into the FNOL flow have closed a compliance gap that traditional telephone-based intake often left to manual handling.

The triage agent receives the intake payload and decides how the claim should be processed. Its decisions include whether the claim is straight-through-processable (typical for small motor own-damage claims below specific thresholds), whether it requires a physical or virtual survey, what level of surveyor expertise is needed, which adjudication queue the claim belongs to, and whether the claim shows any fraud indicators that should be flagged to the Special Investigation Unit (SIU).

The triage agent's tool scope is broader than the intake agent. It can query the claims database for similar prior claims, the historical loss database for benchmark patterns, the fraud indicator system for known-risk markers (shared addresses, surveyors with elevated fraud associations, claim frequency patterns), and the surveyor panel database to assess availability. It produces a structured routing decision, a reasoned justification, and a complexity tier.

The complexity tier drives downstream resource allocation. Tier 1 claims (typically motor own-damage below INR 1 lakh, simple personal accident, small health cashless) proceed through straight-through processing with minimal further human intervention. Tier 2 claims (most SME commercial claims, motor own-damage between INR 1 lakh and INR 25 lakh, group health claims with non-standard procedures) require human review of the triage decision before proceeding. Tier 3 claims (commercial property losses above INR 25 lakh, liability claims, complex marine claims, any claim with fraud indicators) bypass the agent chain entirely after triage and are handled by a senior human claims officer with the agent's analysis as input.

The gate at this stage is more nuanced than at intake. A Tier 1 claim proceeds to surveyor dispatch (or straight-through processing, where applicable) without further human approval. A Tier 2 claim requires a human approver to confirm the triage decision before dispatch. A Tier 3 claim is removed from the agent chain entirely. The thresholds are policy decisions set by the insurer's claims committee, not by the agent. Bajaj Allianz reports that approximately 45 percent of incoming claims are Tier 1, 40 percent are Tier 2, and 15 percent are Tier 3, with the Tier 1 proportion expanding as the agents accumulate confidence data on additional claim types.

The triage agent is also the natural point for fraud screening at the intake-to-process boundary. Rather than running fraud screening as a separate post-registration batch, modern agentic chains embed fraud signals into the triage decision itself, allowing flagged claims to be routed directly to the SIU pipeline rather than progressing through standard claims handling and then being recalled. This compresses the fraud-detection-to-investigation cycle from 5 to 9 days under traditional flows to under 24 hours in production deployments.

The surveyor-dispatch agent handles the operational coordination that has historically consumed the most time in commercial claims handling. For any claim above the straight-through threshold, a surveyor must be assigned, the survey must be scheduled with the insured, the surveyor must conduct the survey, and the survey report must be received and reviewed before settlement can proceed. Each of these steps has historically involved phone calls, email back-and-forth, and manual coordination that adds 3 to 10 working days to the claims cycle.

The dispatch agent's tool scope includes the surveyor panel database, the surveyor availability and load system, the insured's contact details, and the appointment scheduling system. It produces an assignment recommendation based on geography, specialisation, availability, current workload, and prior performance on similar claims. The recommendation is logged and proceeds to dispatch where the surveyor's load is below threshold; otherwise, it escalates to a human dispatcher.

The agent's reasoning is constrained by the IRDAI (Insurance Surveyors and Loss Assessors) Regulations 2015 and the associated panel governance rules. Surveyors above defined claim thresholds must hold the appropriate licence category. Independence requirements prohibit allocating a surveyor with a conflict of interest. The agent's recommendation logic explicitly checks each of these rules and flags any allocation that would violate them.

Once a surveyor is assigned, the agent handles the survey scheduling. It contacts the insured through the insurer's standard channel, proposes survey dates and times based on surveyor availability, captures the insured's preference, confirms the appointment, and sends preparation instructions. Throughout the survey window, the agent monitors progress, sends reminders to the surveyor at defined intervals if the report is delayed, and escalates non-responses after configured periods.

The gate between dispatch and settlement-recommendation is the survey report itself. The agent does not interpret the survey report substantively; it confirms that the report has been received, that it contains the required sections (cause of loss, extent of damage, assessment of admissibility, recommended settlement amount, supporting documentation references), and that the surveyor's qualifications match the claim category. A complete report with all required sections present proceeds to settlement-recommendation. An incomplete or anomalous report (for example, a settlement recommendation outside the surveyor's normal range for similar claims) is flagged for human review before proceeding.

HDFC Ergo reports that the surveyor-dispatch agent has compressed the average time from triage decision to survey scheduling from 48 hours to under 4 hours on Tier 2 claims, and the average time from survey completion to settlement-recommendation initiation from 72 hours to under 8 hours. The compression comes from elimination of phone-tag delays, automated reminder cycles, and structured intake of the survey report.

The settlement-recommendation agent is the most consequential agent in the chain. Its output is a reasoned proposal for settlement amount, settlement basis, applicable deductions, and any required clarifications. The proposal is not binding. Every settlement above the straight-through threshold requires explicit human approval before payment is authorised.

The agent's tool scope includes the survey report (as a structured document), the policy wording and endorsements (with semantic tagging for coverage clauses, exclusions, and conditions), the historical claims database for benchmark patterns, the reinsurance treaty terms (where applicable for treaty cession decisions), and the reserve guidance from the actuarial team. It produces a settlement recommendation with a full reasoning trace.

The reasoning trace is the engineering output that matters most for human review. A human claims officer receiving a settlement recommendation needs to see not only the recommended amount but also the path the agent took to that recommendation: which policy clauses it applied, which exclusions it evaluated and rejected, which deductibles it applied, which comparable claims it referenced, and which uncertainties remain. A recommendation without a reviewable reasoning trace is operationally equivalent to no recommendation at all, because the human officer must redo the analysis to verify the answer.

The agent operates under tight constraints on what it can and cannot produce. It can apply policy clauses and endorsements that are unambiguously triggered by the survey findings. It can apply standard deductibles, excess clauses, and average-clause calculations where the survey provides the necessary inputs. It can flag coverage questions that require legal interpretation. It cannot resolve ambiguous coverage in either direction; ambiguity is escalated to the human officer.

The gate between recommendation and payment is the hardest in the chain. Every settlement, regardless of amount, requires explicit human approval. Approvals are tiered by authority matrix:

• claims below INR 5 lakh are approved by a junior claims officer with the agent's recommendation as input
• claims between INR 5 lakh and INR 50 lakh are approved by a senior claims officer with mandatory review of the reasoning trace
• claims between INR 50 lakh and INR 5 crore are approved by the claims committee with formal documentation
• claims above INR 5 crore are approved at board or board-committee level with reinsurance concurrence where applicable

The agent's recommendation is one input to these approvals, not a substitute for them. ICICI Lombard's deployment of the settlement-recommendation agent on SME commercial claims reports that approximately 78 percent of human approvers accept the agent's recommendation as proposed, while 18 percent approve a modified amount, and 4 percent reject the recommendation entirely and rebuild the settlement from scratch. The high acceptance rate validates the architecture, while the modification and rejection rates confirm that the human gate is doing the work it is supposed to do.

The audit agent operates after the claim is settled and the file is closed. Its role is to review the closed claim for adherence to authority limits, policy wording, regulatory requirements, and internal claims handling standards. It does not have the authority to reopen a claim or modify a settlement; its output is a structured audit finding that feeds into the insurer's quality assurance and continuous improvement processes.

The audit agent's tool scope includes the complete claim file, the policy wording, the authority matrix, the regulatory checklist (including IRDAI claim settlement timeline requirements, mandatory documentation under the Protection of Policyholders Interest Regulations 2024, and any line-specific requirements), and the claims operating standards. It produces an audit report categorising the claim into one of four findings: compliant, compliant with observations, partially non-compliant, or non-compliant.

The value of the audit agent is in its ability to review every closed claim, not a sample. Traditional claims audit in Indian insurers operates on a sampling basis, typically reviewing 2 to 5 percent of closed claims with intensive manual review. Sample-based audit catches systemic issues but misses individual non-compliances that may have material customer or regulatory impact. A continuous audit agent reviewing every closed claim at a fraction of the manual cost expands the audit coverage to 100 percent while preserving the depth of review where it matters.

The audit findings feed three downstream uses. First, they identify individual claims that require remediation, which are escalated to the claims operations head for corrective action. Second, they generate aggregate statistics that the quality assurance function uses to identify patterns (specific surveyors with elevated audit findings, specific claim types with recurring documentation gaps, specific products with consistent coverage interpretation drift). Third, they feed the regulatory reporting required under the IRDAI claim settlement timeline circulars and the operational risk reporting mandated by the IRDAI (Operational Risk Management Framework) circular 2023.

The audit agent is also the natural point for DPDP Act 2023 compliance verification on closed claims. Every claim involves processing of personal data, and the DPDP Act 2023 imposes specific requirements on purpose limitation, retention, and data principal rights. An audit agent that verifies DPDP compliance on every closed claim catches gaps in real time rather than at the next external audit cycle.

The audit agent does not close the loop on the claim itself; the claim is already closed when the audit runs. It does close the loop on the operating model: findings feed back into agent prompts, surveyor performance assessments, claims officer training, and product policy reviews. Insurers that have integrated the audit agent into this feedback loop report measurable improvements in agent recommendation acceptance rates and in audit finding distributions over successive quarters.

The architectural separation across the five agents is enforced through a structured tool-calling layer. Each agent has access to a defined set of tools (read APIs, write APIs, lookup services, communication channels) and no others. The tool registry maintains the mapping of agent role to permitted tool, with every call logged at the registry layer and audit trails immutable.

The permissioning boundary is the single most consequential design decision in the chain. A claims-recommendation agent that has read access to the payment authorisation API is one prompt injection away from a payment release. A triage agent that has write access to the claims database can corrupt downstream agents' inputs. The principle of least privilege is not a stylistic preference; it is the fundamental defence against the failure modes that agentic systems introduce.

Production deployments use a layered permissioning architecture. The agent has a defined identity. Each identity has a set of granted scopes. Each scope maps to a set of API endpoints with specific permission levels. Every API call is authenticated, authorised, logged, and rate-limited. A call outside the agent's scope is blocked at the gateway, logged at the security operations centre, and reported to the agent governance committee.

The tool layer also enforces pre-execution validation. Before any write operation (claim registration, status update, communication dispatch), a validator checks the operation against business rules. A claim status change that skips required transitions is blocked. A communication to an insured outside the approved template library is blocked. A reserve update outside the configured range triggers a human approval workflow. The validator is rules-based for clear rules and model-based for fuzzier checks, with the rules engine acting as the primary defence and the model as a secondary filter for edge cases.

The IRDAI Information and Cyber Security Guidelines 2023 require risk assessments and access controls for systems handling policyholder data. The permissioning architecture described above is the operational implementation of those requirements for agentic systems. Insurers that have shortcut this layer have done so at their own regulatory risk; the audit and incident response evidence required when something goes wrong is impossible to reconstruct without the structured permissioning and logging architecture in place from the start.

A related architectural pattern is the kill switch. Every production agent in the chain has a single-command kill switch that disables all of its outbound tool calls within seconds. The kill switch is wired into the security operations centre and to the claims operations leadership, allowing immediate suspension of the agent in response to anomalous behaviour, incident response triggers, or governance-led pauses for model upgrades. Production deployments test the kill switch monthly as part of the operational resilience routine required under the IRDAI operational risk management framework.

The IRDAI's posture on multi-agent claims systems has evolved through 2025 and into 2026. The IRDAI Regulatory Sandbox, refreshed under the 2024 framework, has been the preferred testing route for novel agentic deployments. At least eleven AI-related claims applications entered the sandbox during 2025, with outcomes feeding into IRDAI's 2026 draft guidance on AI-assisted claims decisions.

Four governance principles have emerged consistently from this regulatory dialogue: explainability, accountability, fairness, and data protection. Explainability requires that every agent decision affecting a policyholder be reconstructable in plain language. Accountability places the insurer's board and senior management on the hook for agent outcomes, regardless of vendor involvement. Fairness requires monitoring for discriminatory outcomes on protected attributes (geography, occupation, claim history patterns). Data protection aligns with DPDP Act 2023 requirements on consent, purpose limitation, and data principal rights.

IRDAI has not mandated a specific architecture for agentic claims handling, but the architectural patterns that satisfy the four principles are converging. The chain-of-agents pattern with human-in-the-loop gates, immutable audit trails, and least-privilege tool permissioning has become the de facto reference architecture. Insurers piloting variants outside this architecture have faced longer sandbox cycles and more questions in the IRDAI Annual Inspection under the IRDAI (Inspection) Regulations.

The outlook for the next 18 to 24 months is a gradual expansion of the agent chain's autonomy on well-bounded claim types. Tier 1 motor own-damage claims are likely to move further into straight-through processing, with the agent chain handling the full FNOL-to-settlement flow under tight authority limits. Tier 2 SME commercial claims will continue to require human approval at settlement, but the proportion accepted as-recommended is likely to rise as the agent chain accumulates training data on insurer-specific patterns. Tier 3 large commercial claims will remain primarily human-led, with the agent chain providing analytic support rather than recommendations.

The Request Access path for brokers and insurers exploring agentic claims tooling typically starts with a single line of business, a single complexity tier, and a defined evaluation window. Successful pilots share three properties: clear baseline measurement of current cycle time and quality metrics, defined scope expansion criteria tied to measured improvement, and an executive sponsor with authority to clear governance friction. The pilots that fail tend to be over-scoped at the start, under-resourced on the engineering side, or run without explicit alignment with the legal, compliance, and operations functions that will own the agent chain in production. The pilots that succeed treat the agent chain as a cross-functional operating asset rather than as a technology project, with the claims operations head, the chief technology officer, the chief risk officer, and the compliance head sharing accountability from the first day of the pilot.