Customer Churn Prediction in Indian Commercial Insurance

Acquiring a new commercial insurance client in India costs significantly more than retaining an existing one. Industry estimates across non-life commercial lines suggest that new business acquisition costs, including broker commissions, underwriting time, inspection fees, and onboarding administration, run 5 to 7 times higher than the cost of renewing an existing policy. Yet most Indian non-life insurers invest a disproportionate share of their analytics and sales resources in new business, treating renewal as a largely administrative process rather than a revenue management opportunity.

Commercial insurance renewal rates in India vary substantially by segment and insurer type. Mid-market commercial accounts (fire and property sums insured between INR 5 crore and INR 100 crore) typically renew at rates of 65 to 78%, according to industry practitioner surveys and broker feedback gathered at the 2025 Indian Non-Life Insurance Summit. Large corporate accounts above INR 100 crore renew at higher rates given the switching friction involved, but they also generate competitive bidding processes that compress margins. SME commercial accounts below INR 5 crore have the weakest renewal rates, often falling below 60%, driven by price sensitivity, broker-switching, and the fact that many SME policyholders do not clearly attribute claims service quality to their insurer versus their broker.

Churn prediction models change this dynamic by identifying, well before renewal, which accounts are at elevated risk of non-renewal or switching. An account flagged as high churn risk six months before renewal gives the insurer, the broker, or both, time to intervene with proactive underwriting review, service uplift, or targeted retention initiatives. The same flag raised two weeks before renewal is operationally useless: there is not enough time to change the account relationship in a meaningful way.

The key distinction from retail insurance churn models is that commercial insurance involves intermediaries. In most cases, the broker, not the insurer, holds the client relationship. A commercial policyholder does not simply switch insurers; they are often moved by their broker to a different market. Churn prediction in commercial lines must therefore model broker behaviour as a first-order variable, not merely an account characteristic.

Churn prediction models are only as good as the signals they consume. The richest predictive signals for commercial insurance non-renewal in India combine policy transaction data, claims history, service interaction records, and market context. Most Indian non-life insurers have the raw data for at least some of these signals but have not systematically organised them into a format that supports model training.

Claims frequency and severity are the most obvious signals and the ones that cut in counterintuitive directions. A commercial account with zero claims over three years is at higher churn risk than commonly assumed: the policyholder may question whether they are receiving value, and a competing insurer can offer a lower premium with the same apparent coverage. An account with one moderate claim that was settled quickly and fairly is actually a strong retention indicator: the insurer has demonstrated its promise. An account with a disputed or delayed claim is the highest-risk category for non-renewal.

Premium growth rate is a direct churn driver. Indian commercial property insurance premiums are influenced by reinsurance treaty pricing, monsoon loss history, and individual risk features. When a renewal premium increases by more than 15 to 20% over the prior year without a corresponding change in risk profile, the account becomes an active candidate for market testing through the broker. Models trained on historical data from Bajaj Allianz and HDFC ERGO commercial portfolios (discussed at claims analytics forums in 2024) consistently show premium increase percentage as one of the top three churn predictors across commercial property and liability lines.

Mid-term endorsement requests are a signal that many insurers overlook. A commercial policyholder who requests multiple endorsements during the policy year (adding locations, increasing sums insured, changing coverage conditions) is an engaged account that is growing and finds value in the product. An account with zero endorsements across two consecutive years may be disengaged or may have shifted business activity that the insurer is not aware of. The absence of interaction can be as informative as its presence.

Broker relationship changes are among the most powerful predictive signals. When the broker managing an account changes (whether through broker consolidation, client-initiated broker switch, or broker staff turnover within the same firm), the probability of insurer switching at renewal increases substantially. Broker data is often not systematically captured in insurer CRM systems, creating a significant gap in the input data available to churn models. Insurers who invest in maintaining up-to-date broker-account linkage data materially improve their model accuracy.

Commercial insurance churn prediction models face a design challenge that retail models do not: the low volume and high heterogeneity of the prediction targets. A large Indian retail life insurer might have millions of policies to train on; a mid-sized commercial non-life insurer might have 20,000 to 80,000 active commercial accounts, with renewal events distributed unevenly across months. This data constraint shapes every modelling decision.

Gradient boosted tree models (XGBoost, LightGBM) are the most widely deployed approach for commercial insurance churn prediction in India, for reasons of interpretability as much as accuracy. Unlike deep learning models, gradient boosted trees produce feature importance scores that claims and underwriting managers can interrogate: "the model flags high churn risk for this account primarily because the renewal premium increased 22% and there was a disputed claim in month 7 of the current policy year." This interpretability is operationally essential: a retention call with a broker is far more effective when the relationship manager understands specifically what the model is responding to.

Feature engineering for commercial churn models requires careful thought about the temporal dimension. The most predictive features are often not static account characteristics but trajectories over time: premium growth trend over three years, service interaction frequency trend over the last six months, claims settlement time relative to policy average. Capturing these trajectories requires the model training pipeline to construct time-series features from transactional data rather than using a single policy-year snapshot, which adds engineering complexity but significantly improves prediction accuracy.

Class imbalance is a persistent challenge. If 25% of accounts churn at renewal, the model sees three times as many retention outcomes as churn outcomes in training data. Standard techniques (SMOTE oversampling, class-weight adjustment, threshold calibration) address this but require careful tuning for the specific class distribution in the insurer's portfolio. The evaluation metric matters significantly here: overall accuracy is a misleading metric for an imbalanced dataset. Precision and recall on the churn class, plus the area under the ROC curve, are the appropriate metrics for a retention model.

Survival analysis methods, particularly Cox proportional hazards models, offer an alternative to binary churn classification that better captures the time dimension of renewal risk. Rather than asking "will this account churn at next renewal?", survival models ask "what is the probability that this account renews for at least N more years given its current characteristics?" This is a more useful formulation for long-term account planning and for prioritising retention investments across a portfolio with different renewal dates.

Indian commercial insurance is overwhelmingly a broker-intermediated market. In commercial property, engineering, and liability lines, brokers control the placement relationship for the vast majority of accounts. This creates a fundamental complication for churn prediction: the insurer is often not the proximate cause of churn, and intervention may need to go through the broker rather than directly to the policyholder.

Broker-driven churn takes several forms that models need to distinguish. Competitive repricing churn occurs when a broker shops an account at renewal to find a lower premium, often as a routine service to the client regardless of any service issue. Relationship-driven churn occurs when the insurer has damaged its relationship with the broker, perhaps through disputed commissions, delayed claim settlements on broker-managed accounts, or poor underwriting responsiveness. Structural churn occurs when a broker consolidates its panel of insurers and the smaller insurer loses the relationship without any individual account fault.

Predicting competitive repricing churn requires external market data: what is the current rating environment for the relevant line of business, and is the insurer's renewal premium competitive with current market rates? Indian non-life insurance premium rates for commercial property are influenced by reinsurance pricing cycles, IRDAI's detariffed environment, and individual risk assessments. An insurer whose commercial property rates are currently above market due to loss experience has a structural churn risk that is not reducible by relationship management alone.

The broker performance scorecard, if maintained systematically by the insurer, provides inputs that materially improve churn model accuracy. Brokers with high account retention rates, consistent on-time renewal placement, and low claim dispute rates are positive signals for the accounts they manage. Brokers with deteriorating performance metrics, new broker appointments replacing experienced handlers, or recent disputes with the insurer's claims team are churn risk multipliers. New India Assurance and United India Insurance, as the two largest commercial non-life insurers by premium volume, maintain broker performance data as part of their empanelled broker management processes, which can be integrated into churn prediction models with appropriate data governance.

Identifying an at-risk account is only useful if the insurer can act on that identification in a commercially sensible way. The intervention strategy depends on the primary driver of churn risk, which the model's feature importances can indicate at the individual account level.

For accounts flagged primarily due to premium growth, the intervention is a proactive underwriting review before the renewal is formally processed. The underwriter or relationship manager contacts the broker 90 days before renewal, reviews the risk features that drove the premium increase, and determines whether any can be addressed (improved fire suppression, reduced stock values due to business changes, installation of security systems) to justify a lower premium. Bajaj Allianz General Insurance's commercial lines teams have implemented structured pre-renewal reviews for accounts flagged by their retention model as premium-sensitive, with reported success in retaining accounts that would otherwise have gone to market.

For accounts flagged due to claims service issues, the intervention is different: a service recovery call from a senior claims manager or the branch head, acknowledging the difficulty of the claims experience and committing to specific service improvements. This is most effective when the claim is still within the settlement process (the surveyor appointment has been made but the final payment has not been released), as a mid-course correction carries more credibility than a retrospective apology. When the claim has already been settled, the intervention must focus on demonstrating future commitment through concrete actions, such as a dedicated claims contact for the account's next loss event.

For accounts flagged due to broker relationship changes, the appropriate intervention is a meeting with the new broker contact, not with the policyholder directly. The insurer's sales team should invest in understanding the new broker's preferences, account management expectations, and any service standards their clients typically require, and communicate how the insurer intends to meet those standards. The commercial relationship with the broker is the asset to protect; the policyholder relationship flows from it.

Retention program ROI can be calculated at the portfolio level with reasonable precision. If the model identifies 500 high-risk accounts representing INR 35 crore in annual premium, and the intervention program costs INR 80 lakh in relationship manager time and underwriting review costs, a retention lift of even 10 percentage points (retaining 50 additional accounts at an average premium of INR 7 lakh) generates INR 3.5 crore in retained premium against INR 80 lakh in intervention cost. The payback case is strong relative to new business acquisition costs for the same premium quantum.

Churn prediction models use policyholder data to generate predictions that drive commercial actions, specifically, targeted retention interventions aimed at influencing renewal decisions. This use of data has specific implications under IRDAI's data governance framework and India's evolving data protection regulations.

IRDAI's Guidelines on Insurance Data Analytics, 2024 require insurers to maintain clear documentation of the purpose for which policyholder data is collected and used. Using claims history data to price risk is an established and accepted purpose. Using the same claims history data to predict churn probability and drive retention outreach is a different purpose that requires separate documentation and, depending on the insurer's privacy policy, potentially additional consent or disclosure to policyholders. Insurers building churn prediction capabilities should engage their compliance teams to confirm that the data usage is within the scope of consents obtained at policy inception.

India's Digital Personal Data Protection Act, 2023 (DPDPA) introduces additional considerations. Under the DPDPA, the use of personal data for purposes beyond the originally stated purpose requires a new consent or a legitimate basis under the Act. For commercial insurance policyholders that are corporate entities rather than individuals, the DPDPA applies differently: the Act primarily covers personal data of natural persons. However, individual contact data (relationship manager names, personal email addresses, mobile numbers) used within CRM systems for retention outreach may qualify as personal data subject to DPDPA requirements even where the policyholder entity is a company.

Data minimisation principles also apply. A churn prediction model does not need every field in a policyholder record to generate an accurate prediction. Building the model on the minimum necessary data fields, and excluding sensitive personal data that is not predictive, reduces both regulatory risk and the consequences of any data breach. Insurers should require their data science teams to document which specific data fields the churn model uses and why each is necessary, as this documentation supports regulatory audit responses.

Sharing churn prediction outputs with brokers (so that brokers can take retention action on behalf of the insurer) introduces additional data governance questions. The broker receives information derived from the policyholder's own claims and service history. Most broker agreements in India do not explicitly address sharing of analytics outputs, and insurers should review their broker data-sharing agreements before operationalising broker-facing retention dashboards.

The financial case for churn prediction investment is built on two comparisons: the cost of retaining an at-risk account versus the cost of replacing lost premium through new business acquisition, and the improvement in retention rates achievable through targeted intervention versus the baseline retention rate without intervention.

New business acquisition costs in Indian commercial insurance are driven primarily by broker commission rates, which are regulated by IRDAI but vary by line of business and negotiated agreement. For commercial property, broker commissions on new business placements typically run 10 to 15% of premium in addition to underwriting expenses, inspection costs, and administrative onboarding. For a new commercial account generating INR 10 lakh in annual premium, total acquisition cost can reach INR 1.5 to 2.5 lakh. Retention of the same account at renewal requires a much smaller investment in relationship management, underwriting review, and broker support.

Retention rate improvement from predictive model-driven interventions is typically measured at 3 to 8 percentage points above baseline in the first year of deployment, based on published findings from property and casualty churn prediction programs in comparable markets (the UK, Australia, and Singapore, whose commercial lines dynamics have meaningful parallels to India's). Indian insurer data on measured retention lift from churn prediction programs is sparse in the public domain, though practitioners at the 2025 Indian Non-Life Insurance Summit cited similar improvement ranges for commercial portfolio programs.

The compounding effect of improved retention over multiple years is substantial. An insurer that retains 5 percentage points more of its commercial portfolio each year does not merely preserve the premium from those accounts in year one; it also retains the renewal base for year two, three, and beyond, along with the cross-sell opportunities those accounts represent. Over a five-year horizon, a 5% improvement in annual retention rate for a portfolio of INR 500 crore in commercial premium generates cumulative retained premium that significantly exceeds the model development and intervention program costs.

Building a production-grade commercial insurance churn prediction capability in India typically progresses through four stages, each requiring different investments in data infrastructure, model development, and operational integration.

The first stage is data consolidation. Most Indian non-life insurers have policy data, claims data, and broker data in separate systems with limited linkage. Before any model can be trained, these datasets must be joined at the policy-account level with consistent identifiers, and historical data must be organised into a training set where the outcome (renewed or churned) is known for each account. This data engineering work often takes three to six months and is the stage most frequently underestimated in project planning.

The second stage is baseline model development and validation. A gradient boosted tree model trained on two to three years of historical renewal data, validated on a held-out period, can typically be built in four to six weeks by a team of two data scientists familiar with the insurance domain. The output at this stage is a model that produces churn probability scores for every account approaching renewal, with feature importance outputs that identify the primary risk drivers for each flagged account.

The third stage is operational integration: connecting the model output to the CRM system used by relationship managers and the broker portal used by distribution partners. This integration determines whether the model's insights actually reach the people who can act on them. The integration should display churn risk scores alongside actionable context, not just a probability number but the two or three specific factors driving the score for each account, with suggested next actions tailored to each factor type.

The fourth stage is continuous model improvement through outcome tracking. When a flagged account is retained, was the retention driven by the intervention or by factors outside the insurer's control? When a flagged account churned despite intervention, what was the intervention and why did it fail? Building this outcome tracking into the model governance process allows the insurer to progressively refine both the prediction model and the intervention playbook. Insurers who skip this stage see model performance degrade over one to two years as market conditions shift and the original training data becomes stale.