AI in Reinsurance Treaty Optimisation for Indian Primary Insurers

Every Indian general insurer faces the same fundamental reinsurance question at each annual treaty renewal: how much risk to retain, in what structure, with which reinsurers, at what price. The answer determines the insurer's net underwriting result, its capital requirement under the solvency regime, its earnings volatility, and its competitive position in pricing primary risks. For a mid-sized Indian commercial lines insurer with a INR 1,500 to 3,000 crore gross written premium book, the difference between an optimised and a suboptimal treaty structure can amount to INR 50 to 200 crore in annual reinsurance cost, or equivalent variation in retained loss exposure.

Historically, the treaty renewal decision in India has been driven by actuarial convention and broking relationships rather than rigorous optimisation. An insurer's reinsurance manager, working with a reinsurance broker (commonly Gallagher Re, Guy Carpenter, Aon Reinsurance Solutions, or one of the Indian specialist brokers), would review the prior year's treaty performance, assess the market's appetite for Indian risk following any large catastrophe events, and negotiate terms primarily on the basis of experience rating: the actual loss record of the expiring treaty compared to the expected losses priced into the expiring rate.

This approach, while practical, leaves substantial value on the table. It does not systematically optimise the retention level to account for the insurer's current capital position and risk appetite. It does not model the interaction between different treaty layers or between quota share and excess of loss structures. It does not test whether the current treaty structure remains appropriate given changes in the portfolio's geographic concentration, line-of-business mix, or emerging catastrophe exposure. AI-based tools designed for reinsurance programme optimisation address each of these gaps by replacing judgment-based approximation with explicit modelling of the optimisation objective.

The framing of the reinsurance treaty as an optimisation problem is not new to global reinsurance markets, where catastrophe modellers at Guy Carpenter's GC Facilitation team and at Willis Re's analytics group have applied stochastic optimisation to reinsurance structure design for two decades. What is new is the accessibility of these methods to Indian primary insurers, driven by the increasing availability of cloud-based catastrophe modelling platforms, improved portfolio data quality following IRDAI's data submission requirements under IRDAI (Reinsurance) Regulations, 2018, and the arrival of AI-native reinsurance optimisation tools tailored to emerging market portfolios.

Experience rating, the dominant approach in Indian reinsurance pricing, uses actual historical losses to project future expected losses. An excess of loss treaty priced on experience rating looks at the insurer's loss record in the relevant layer over the past 5 to 10 years, applies development factors for unreported and unsettled claims, and derives an expected loss cost. The approach is sound when the insurer has a stable portfolio and long loss history, but it has known weaknesses that AI-based approaches can partially address.

The first weakness is sparse data in the tail. Excess of loss treaties, especially at higher attachment points, are triggered infrequently. An insurer with a 1-in-20 year loss event in its territory may have no such events in its 10-year history and therefore zero experience in the relevant layer. Experience rating, by definition, has nothing to say about this. Predictive models trained on industry-wide loss data, engineering-derived hazard intensities, and physical catastrophe model outputs can project loss frequencies and severities for low-probability events that experience rating cannot assess.

The second weakness is portfolio change. An insurer that has grown its commercial property portfolio from INR 500 crore to INR 2,000 crore in sum insured concentration in a coastal metropolitan area over the past five years has a materially different catastrophe exposure than its historical loss record suggests. A predictive model that ingests the current portfolio structure, applies a stochastic event set calibrated to Indian flood and cyclone hazard, and simulates losses under the current portfolio is capturing the current risk. Experience rating applied to the prior five years is capturing a smaller, differently composed portfolio.

The third weakness is correlation. A property portfolio concentrated in the manufacturing belt of Gujarat has correlated exposure to the same seismic events that also generate liability claims from product recall, business interruption from supplier disruption, and marine losses from port damage. Experience rating on each line separately does not capture the correlated impact of a single catastrophe event on multiple lines simultaneously. Predictive models built at the portfolio level can assess aggregate exposed value and correlate losses across lines for the same event, giving the reinsurance actuary a more complete picture of the aggregate probable maximum loss.

AI models improve on the traditional predictive approach by automating feature engineering from the insurer's portfolio data, identifying non-obvious correlations between underwriting characteristics and loss outcomes, and updating loss projections dynamically as the portfolio changes during the year. Indian firms including Quantbot Technologies and the analytics teams at large insurers are using gradient boosting and deep learning models for this purpose, calibrated against the Insurance Information Bureau of India (IIB) industry loss database and supplemented by proprietary claims data.

Indian commercial insurers use three primary treaty structures, each of which is optimised differently by AI tools.

Quota share treaties cede a fixed percentage of all premiums and losses to reinsurers in exchange for a reinsurance commission. For a primary insurer, a quota share treaty reduces both premium income and loss exposure proportionally. The optimisation question is: what percentage to cede, and at what commission rate? AI tools model the optimal cession percentage as a function of the insurer's capital position, target solvency ratio, growth plans, and the available commission rate in the market. A higher cession reduces retained premium income but also reduces capital consumption, which may enable faster portfolio growth. The optimal cession rate trades off these effects, and the solution changes as the insurer's capital position evolves through the year.

Surplus treaties are layered above a defined retention. The insurer retains risks up to a defined sum insured, and cedes the excess proportionally to the surplus line. For commercial property, a typical Indian arrangement might involve a retention of INR 2 crore per risk with a surplus treaty of 9 lines (giving combined capacity of INR 20 crore per risk). The optimisation question for surplus treaties is the retention level: too low, and the insurer cedes profitable premium unnecessarily; too high, and the insurer retains concentration risk it cannot comfortably absorb. AI tools model the optimal retention as a function of individual risk characteristics, portfolio concentration, and the marginal cost of the surplus capacity.

Excess of loss treaties provide protection once aggregate or per-occurrence losses exceed a defined attachment point. The optimisation questions for XL treaties are the attachment point, the limit purchased, and whether to buy multiple layers. Too low an attachment means the insurer is buying expensive catastrophe cover for attritional losses it can absorb from operating income. Too high an attachment leaves the insurer exposed to events it cannot absorb. AI optimisation models the attachment point as a function of the insurer's earnings capacity, capital buffer, and the stochastic loss distribution at different layers.

Interaction effects between treaty structures

The most sophisticated AI-based optimisation work models the interaction between all three treaty types simultaneously, rather than optimising each in isolation. An insurer with a quota share that covers the first 40% of each risk, a surplus treaty above the retained line, and an excess of loss treaty protecting the net retained account is running three overlapping protections. The optimal parameterisation of each depends on the others. AI simulation tools model thousands of treaty structure combinations across the stochastic event set and identify the Pareto frontier of structures that minimise reinsurance cost for a given level of retained loss volatility, or that maximise portfolio growth capacity for a given solvency constraint.

Any discussion of reinsurance optimisation for Indian primary insurers must begin with GIC Re, because IRDAI (Reinsurance) Regulations, 2018 establish GIC Re as the mandatory first refusal reinsurer. Indian insurers are required to offer GIC Re a mandatory cession on all reinsurance, with GIC Re entitled to retain a defined percentage before the insurer places the balance with other reinsurers. The mandatory cession percentage has evolved over IRDAI's regulatory cycles, and as of the 2025-26 financial year, GIC Re retains a right of first refusal at terms that Indian insurers must accept before accessing international capacity.

The mandatory cession constraint has direct implications for treaty optimisation. The AI optimisation model cannot freely choose how to distribute risk across reinsurers: GIC Re receives its mandatory cession first, and the optimisation operates over the remaining portion. This reduces the degrees of freedom in the optimisation but does not eliminate it. The insurer can still optimise its retention level, its surplus and XL treaty structures, the distribution of the balance between GIC Re's additional facultative capacity and international treaty reinsurers, and the choice between quota share and surplus arrangements for different segments of the portfolio.

GIC Re's own pricing and treaty terms are not set by market competition in the same way as international reinsurers. GIC Re applies its own rating methodology, which has historically been less responsive to individual insurer data quality improvements than international reinsurers who compete actively for well-performing Indian cedants. This creates an AI optimisation challenge: the benefit of improved data and modelling is realised primarily in negotiations with international reinsurers who adjust terms based on the quality of the insurer's analysis, rather than with GIC Re whose terms are more administratively determined.

The practical implication is that AI-based reinsurance optimisation in India has its greatest impact on the international market-facing portion of the programme. Insurers who bring stochastic loss modelling, AI-validated PML estimates, and portfolio analytics to their Singapore or London negotiations can demonstrate that their risk is better understood and better managed than the average Indian cedant, supporting better terms on the international layers. The GIC Re layer remains subject to the mandatory framework, but even there, better data has informational value in managing the relationship and in planning treaty programme design.

Catastrophe modelling is the technical foundation of reinsurance treaty optimisation for property-dominant portfolios, and its integration with AI optimisation tools represents the most technically demanding element of modern reinsurance programme design. Indian natural peril risks differ from global average assumptions in ways that matter significantly for treaty optimisation.

Flood is the dominant catastrophe peril for Indian commercial property. India experiences severe flood events in multiple river basins annually, with the Brahmaputra, Kosi, Mahanadi, Krishna, and Godavari systems generating significant insured losses in most years. International catastrophe model providers including AIR Worldwide (now Verisk), RMS (now Moody's RMS), and EQECAT have India flood models, but industry practitioners consistently note that these models underestimate Indian flood loss frequency and severity for certain basins, particularly in the Northeast and Bihar-Jharkhand corridor, where drainage infrastructure and flood defence conditions differ materially from model assumptions. GIC Re's own catastrophe modelling team has produced proprietary flood loss estimates for Indian basins that are used in treaty discussions alongside commercial model output.

Cyclone risk affects the eastern and western coastlines differently. The Bay of Bengal generates more frequent and intense cyclones than the Arabian Sea, creating different natural peril profiles for insurers concentrated in Odisha, Andhra Pradesh, and West Bengal versus those with Gujarat and Maharashtra concentration. AI-enhanced catastrophe models improve on standard commercial models by incorporating local building vulnerability data, historical insurance loss data from the IIB, and satellite-derived land use and construction data to calibrate event loss estimates to actual Indian portfolio characteristics rather than global average assumptions.

Earthquake is significant for Himalayan-foothills portfolios (Uttarakhand, Himachal Pradesh, Northeast India) and for legacy high-risk zones in Gujarat (remembering that the 2001 Bhuj earthquake caused insured losses estimated at INR 2,500 crore). AI-assisted seismic risk assessment integrates ShakeMap data, soil amplification factors from the Indian Meteorological Department, and individual building vulnerability assessments to produce portfolio-level probable maximum loss estimates at defined return periods.

The integration of catastrophe model output into the treaty optimisation workflow requires the AI optimisation tool to ingest event loss tables (ELTs) from the catastrophe models and combine them with the insurer's portfolio data. The combined stochastic event set, typically 10,000 to 100,000 simulated years, drives the loss distribution from which attachment points, expected loss costs, and aggregate probable maximum losses are derived. AI tools automate the computation across this event set, enabling the optimisation to run against the full distribution rather than against a handful of deterministic scenarios.

The annual reinsurance treaty renewal is fundamentally a negotiation between the primary insurer (cedant) and reinsurers about how much of the cedant's risk each reinsurer is willing to assume and at what price. The cedant's analytical capability directly affects the quality of that negotiation. An insurer that arrives at renewal with detailed portfolio analytics, stochastic loss modelling, and AI-validated risk assessments is in a fundamentally different negotiating position than an insurer relying on static bordereaux and experience summaries.

International reinsurers assess Indian cedants partly on the quality of the data they can provide. At the January and April renewal meetings in Singapore, London, and Dubai (the primary hubs for Indian treaty negotiations), cedants who provide detailed risk location data with latitude-longitude coordinates, construction type breakdowns, occupancy classification, and individual risk PML estimates receive more competitive terms from analytical reinsurers than cedants providing only aggregate premium and loss summaries. AI tools that can process raw policy data from the insurer's administration system and produce a structured reinsurance submission package in standard formats (RDS, RDOS, OED) substantially reduce the cost of producing high-quality submission materials.

AI-based pricing model validation tools allow the insurer to arrive at renewal with an independent assessment of fair treaty pricing, challenging the reinsurer's rate indication when the model analysis suggests the risk is better than the reinsurer's pricing implies. This is a shift from the traditional asymmetry in which the reinsurer's analytical team substantially outresourced the cedant's. Indian insurers including HDFC Ergo and New India Assurance have invested in reinsurance analytics teams using commercial and proprietary modelling platforms to build this internal capability.

The timing of the renewal negotiation is also affected by AI tools. Continuous portfolio monitoring using AI systems that flag emerging concentrations or shifts in risk profile allows the insurer to bring real-time data to mid-year discussions with reinsurers about treaty amendments. Rather than waiting for the annual renewal to address a growing geographic concentration or a new line of business, the insurer can open discussions with reinsurers 6 to 9 months before the treaty year end, when the reinsurer has more flexibility and the cedant has more negotiating options.

The effectiveness of AI-based reinsurance treaty optimisation is directly bounded by the quality of the underlying portfolio data. Indian commercial insurance portfolios present specific data quality challenges that affect the reliability of AI optimisation outputs.

Location data quality is the most frequently cited constraint. Many historical commercial property policies in India were written with address-only risk locations, without latitude-longitude geocoding. Catastrophe models require geocoded locations to apply hazard footprints. AI geocoding tools trained on Indian address formats can convert address strings to coordinates with 85 to 92% accuracy for structured addresses in urban areas, declining to 60 to 75% for semi-urban and rural addresses with less standardised formats. The ungeocoded or miscoded fraction of a portfolio, which may represent 15 to 30% of locations in a legacy commercial book, introduces error into the catastrophe model output that cascades into the treaty optimisation.

Sum insured accuracy is the second major constraint, directly related to the underinsurance problem discussed elsewhere. A portfolio where 35% of industrial risks have sum insured values that are 30 to 60% below current reinstatement cost produces a catastrophe model output that understates the insurer's true exposure. AI tools can cross-validate declared sum insured values against benchmark reinstatement cost databases (maintained by providers including CBRE, JLL, and the government's Central Public Works Department cost indices), flag properties where declared values appear inconsistent with market benchmarks, and suggest adjusted exposures for treaty purposes. But this cross-validation cannot fully substitute for accurate sum insured declarations, and insurers using AI-adjusted exposures in treaty negotiations must disclose the adjustment methodology to reinsurers.

Claims data quality affects the experience rating component of reinsurance pricing. Indian insurance claims data in legacy systems often lacks the granularity required for sophisticated experience analysis: individual claim records may lack peril coding, cause of loss description, or individual risk identifiers that would allow attribution to specific policies. AI-powered data cleaning and enrichment tools can recover some of this information by matching claims to policies, inferring peril from loss description text using NLP, and applying pattern recognition to identify potential duplicate or split claims. ICICI Lombard and Bajaj Allianz have both invested in claims data remediation projects using these tools in preparation for treaty renewals.

The IIB maintains an industry loss database that supplements individual insurer experience for pricing purposes, particularly for lines with sparse individual insurer loss history. IRDAI's data submission requirements under IRDAI (Obligation of Insurers to Rural or Social Sectors) Regulations and related data collection circulars have progressively improved IIB data quality. AI tools that access IIB data as a supplementary source for benchmarking individual insurer loss ratios and for calibrating model parameters improve the reliability of the treaty optimisation without requiring the individual insurer to have a long individual loss history in every line and geography.

Realising the benefits of AI-based reinsurance treaty optimisation requires an organisational model that does not currently exist at most Indian primary insurers. The traditional reinsurance function consists of a reinsurance manager with treaty administration expertise, supported by the insurer's appointed actuary for reserving and pricing inputs. AI-based optimisation adds a data science capability that most Indian reinsurance departments lack.

The leading Indian insurers are addressing this through three models. The first is internal team building: hiring data scientists with catastrophe modelling and stochastic optimisation skills into the actuarial or reinsurance function, and training them on the insurance domain. This is effective but slow, given the competition for qualified candidates between Indian insurers, global reinsurers, insurtechs, and technology companies.

The second model is reinsurance broker partnership: working with the analytics arms of reinsurance brokers, whose modelling teams (including Guy Carpenter's GC Analytics, Aon's Reinsurance Solutions Analytics, and Gallagher Re's analytics practice) have deep catastrophe modelling and optimisation capabilities. Broker-led optimisation models have the advantage of access to market pricing intelligence that individual insurers cannot replicate, but create a dependency on the broker relationship and may not give the insurer fully proprietary analytical capability.

The third model is specialist vendor engagement: procuring AI-based reinsurance optimisation software from providers including Sequel (part of Verisk), AIR Worldwide's ReMetrica platform, or emerging Indian-market tools from vendors including Sarvada Intelligence and Quantbot Technologies. These platforms provide the modelling infrastructure while the insurer's own team provides the portfolio data and domain judgment. The vendor model gives the insurer proprietary capability built on the vendor's platform, with the vendor providing model maintenance and updates as the underlying catastrophe science improves.

The governance dimension of AI-based reinsurance optimisation is material. The insurer's board is required by IRDAI (Reinsurance) Regulations, 2018 to approve the annual reinsurance programme. When that programme is informed by AI optimisation outputs, the board must have sufficient understanding of the methodology to discharge its oversight responsibility. Board-level education on catastrophe modelling and AI optimisation is now a standard agenda item at the more analytically sophisticated Indian insurers, delivered through the actuary, the CRO, or external consultants. The goal is not to turn board members into modellers but to ensure they can ask the right questions about model assumptions, data quality, and the basis for retention decisions.