This text was written in collaboration with César Ortega, whose insights and discussions helped form the concepts introduced right here.
the best knowledge product begins with sitting down with enterprise companions to grasp day-to-day workflows, handoffs, and bottlenecks. On this article, we focus on a problem that doesn’t require a sophisticated answer, only a easy optimization downside. It’s an excellent instance of how primary instruments can nonetheless remedy high-value issues. Particularly, we concentrate on optimizing the project of on-line insurance coverage insurance policies to trusted companions (unbiased insurance coverage companies: iia) at a worldwide insurance coverage firm.
Impartial insurance coverage companies are privately owned intermediaries that promote insurance coverage insurance policies from a number of insurers. In contrast to massive insurance coverage firms, they don’t design merchandise, set costs, underwrite danger, or pay claims; as an alternative, they examine choices throughout carriers, and place protection that most closely fits the consumer’s wants, usually incomes commissions for doing so. Right here, the thought is to work collectively to ship the perfect worth for each the company and the consumer.
Decreasing complexity
Optimization in the actual world is a spectrum. At one finish are actual strategies that may show optimality, however they typically might be computationally heavy at scale and might wrestle as the issue grows in dimension and operational element. On the different finish are heuristics, starting from easy rule-based baselines which can be simple to elucidate however arduous to take care of as complexity grows (typically residing in massive excel sheets), to extra superior metaheuristics that scale effectively computationally however might be more durable to justify, audit, or debug.
In observe, the simplest strategy typically sits within the center: pragmatic “good-enough” formulations, constructed with fastidiously chosen constraints that mirror each enterprise guidelines and actual operational limits as human workload and repair high quality.
The aim shouldn’t be theoretical perfection, however an answer that’s deliverable, comparable towards baselines, and simple to iterate. With a modular construction and a staged modeling technique, we are able to begin easy, measure influence with KPIs: tangible (time to project, optimum company choice, and so on.) and intangible (keep away from unfair concentrations of insurance policies in a couple of companies, and so on.), and evolve the system by way of small, secure enhancements moderately than ready months for a textbook-optimal mannequin.
That’s why we selected a light-weight optimization formulation. It captures the constraints that matter (capability, geographic eligibility, equity, and bucket combine) and delivers a deterministic, auditable reply quick sufficient for real-time latency necessities. If wanted, we are able to later prolong the strategy with decomposition strategies, stronger solvers, or heuristics with out altering the system’s core contract.
The baseline
Traditionally, these digital policy-to-agency of assignments have been accomplished manually, guided by non-standard standards and particular person judgment. Whereas this strategy typically works, this typically resembled a round-robin strategy: insurance policies had been distributed sequentially amongst accessible companies (iia’s), with little consideration for variations in capability, experience, or anticipated efficiency.
Whereas easy and seemingly truthful, it typically results in delays, missed alternatives, and uncertainty about which company (iia) is the perfect match. The method additionally didn’t scale effectively, creating additional project delays, and the outcomes didn’t persistently align with strategic objectives comparable to profitability, high quality, reproducibility, and transparency.
For that reason, we current how we solved an vital downside utilizing a light-weight integer programming strategy that matches incoming on-line insurance coverage insurance policies to companies in actual time. The tactic maximizes a productiveness rating (reflecting how effectively an company has carried out prior to now) whereas balancing company capability, equity, and geographic admissibility constraints primarily based on ZIP codes. We define the mathematical formulation, the live-update logic, and the PuLP implementation.
What downside are we fixing?
When a brand new on-line coverage is bought for a consumer, somebody nonetheless has to determine which company ought to deal with it. We depend on companies as a result of they add worth past the same old, comparable to advocating at declare time, servicing adjustments and renewals, cross-selling, and extra. Importantly, companies additionally originate demand: they bring about new shoppers (and consequently new insurance policies) into the funnel by way of their relationships and native presence, which compounds progress for the insurance coverage firm.
From a buyer perspective, this issues as a result of the company is commonly the major level of contact: the standard and velocity of company (iia) service can form the general expertise, particularly throughout high-stress moments like claims or pressing protection adjustments.
Since companies differ in licensing, geography, product strengths, gross sales attain, and day-to-day capability, the “greatest” company can differ from second to second. An actual-time project optimization system routes every new coverage to eligible, accessible companies which can be more than likely to ship worth to each the enterprise and the consumer, are handled pretty underneath clear guidelines, and are greatest positioned to drive future progress.
Good Previous-Long-established optimization
To create a transparent project course of, it’s important to think about broader enterprise objectives: comparable to ensuring the best company handles the best kind of coverage to maximise key efficiency indicators (KPIs) like coverage quantity and high quality. It’s additionally vital that companies perceive how these choices are made.
So, the carried out optimization algorithm ought to intelligently allocates insurance policies to companies primarily based on KPIs, together with the quantity and high quality of insurance policies they deal with. As a substitute of counting on subjective or inconsistent human judgment, the algorithm makes use of real-time, data-driven choices to optimize the coverage project course of effectively and pretty.
The optimization mannequin allocates insurance policies to companies primarily based on measurable efficiency indicators moderately than subjective judgment. To make choices reproducible, we translate company efficiency right into a numeric worth the optimizer can use. That is accomplished by way of productiveness weights, the place the important thing enter is the swap ratio: a metric that captures how a lot worth an company brings per unit of coverage it receives (for instance loss ratio, tenure, premium, cross-selling, and so on.).
In observe, the swap ratio permits the mannequin to distinguish companies that persistently ship robust outcomes from people who underperform. Larger-value insurance policies can then be directed towards companies which have demonstrated the power to deal with them successfully, whereas nonetheless respecting capability limits, geographic eligibility, equity necessities, and bucket-mix constraints.
Fairly than counting on static guidelines, the system recalculates choices as constraints, guaranteeing that assignments stay aligned with present operational capability and enterprise priorities.
The system operates in two modes:
- Batch mode: Optimizes primarily based on historic allowances, offering a complete evaluate of previous knowledge to enhance future allocations.
- On-line mode: Re-optimizes with every new incoming coverage, together with these new insurance policies within the optimization course of, then updates the stock and refines the batch optimization accordingly.
In essence, the batch mode handles historic knowledge to ascertain baseline guidelines and patterns, whereas the web mode ensures real-time adaptability by dynamically adjusting to new insurance policies and circumstances. This strategy helps preserve optimum efficiency in a consistently altering atmosphere.
The Resolution: Optimization Algorithm
Given a set of companies A and an incoming circulation of insurance policies P, we wish to determine what number of insurance policies to assign to every company and every coverage class (Gold, Silver, Bronze) in order that we maximize whole productiveness whereas adhering to sure constraints (company capability, ZIP code eligibility, , whole depend, penalties, and so on.).
Objetive perform:
- x is the determination variable within the optimization downside and represents the variety of insurance policies assigned to company a and class c, we solely handle optimistic integer values solely.
- A: set of companies (dimension |A| = m); a∈A.
- C: set of classes {Gold, Silver, Bronze} (|C| = p = 3); c ∈ C.
- The productiveness weights w is one quantity per company that estimates the advantage of sending yet one more coverage to that company. That is calculated with the time the company have over the swap ratio.
Guidelines we should respect (constraints):
Logical constraints:
Logical constraints are those required for the mannequin to be mathematically well-defined no matter enterprise context (e.g., variables are integers and totals steadiness).
- Integrality & Non-negativity: you possibly can’t ship detrimental or fractional insurance policies.
2. International conservation: the overall variety of insurance policies assigned throughout all companies and buckets should equal the overall stock accessible for project on this run (the sum of all company capacities).
Enterprise constraints:
Enterprise constraints encode area coverage decisions or operational guidelines (e.g., per‑company capability, ZIP admissibility, bucket combine, on-line flooring) that might change if the enterprise guidelines change.
- Per-agency capability: an company can’t obtain extra insurance policies than it may possibly at the moment deal with (Ua), which corresponds to the sum of the rows within the coverage project matrix.
2. ZIP admissibility: companies are solely licensed or approved to service insurance policies in particular geographic areas.
If a ZIP is inadmissible for company a, lock its row whole
By implementing ZIP eligibility within the optimization, we guarantee each project is operationally possible, defending service high quality, as a result of companies are strongest within the areas the place they’ve native presence and experience.
3. Bucket bounds: enterprise management that maintain the month-to-month allocation balanced throughout coverage tiers.
With out them, the optimizer would possibly push virtually all the things into essentially the most worthwhile tier, which may create danger focus and operational pressure. By setting minimums and maximums per bucket, you implement a wholesome combine that displays danger urge for food, service capability, and strategic targets.
What’s Not within the batch
Batch mode is a full re‑optimization on a hard and fast stock. It finds the perfect baseline allocation with out reacting to a single new coverage occasion. For that purpose, we exclude the next “stay” constraints which can be solely wanted when a brand new coverage arrives:
- Per‑company flooring from the earlier allocation. Flooring are a web-based safeguard that forestalls any company from shedding insurance policies when a brand new one arrives. In batch we’re computing the baseline itself, so there’s no “earlier” baseline to guard.
- ZIP lock is a stay‑mode security rule: when a single new coverage arrives, if that coverage’s ZIP is not allowed for company A, we freeze company A at cell stage (Gold/Silver/Bronze) at its earlier cell values so the brand new coverage can’t be assigned there and we don’t transfer any current insurance policies away.
- No headroom (“+1”) trick. Headroom is utilized in on-line mode to maintain feasibility when including precisely one new coverage. Batch mode doesn’t add a single coverage; it allocates your complete stock directly.
- Bucket bounds nonetheless apply on-line: every new coverage should maintain Gold/Silver/Bronze totals inside their min/max. These restrictions are up to date on a month-to-month foundation or as enterprise necessities change.
Why this works
By separating the method into batch (world steadiness) and on-line (native adjustment), the system achieves each stability and responsiveness. Batch optimization offers a constant, auditable reference level, whereas stay decisioning handles real-time arrivals with out disrupting the general construction. This mixture permits quick operational choices whereas preserving equity, capability management, and alignment with strategic targets.
E2E Implementation
The top-to-end course of includes greater than encoding guidelines in an optimization mannequin. In our AWS setup, Airflow orchestrates scheduled knowledge pipelines that refresh intermediate tables on every day, weekly, and month-to-month cadences. These jobs pull upstream knowledge, construct curated datasets and stay stock tables, and retailer them in S3. The Optimization service reads the newest inputs from S3 and, when wanted, calls a SageMaker endpoint to attain candidates and choose the perfect company underneath the capability, equity, and ZIP-code constraints described earlier. Exterior functions ship requests by way of an HTTPS endpoint on API Gateway, which routes them through middleware chargeable for authentication, validation, and request transformation earlier than invoking the Optimization service (and SageMaker, if required). The response (containing the chosen company and determination metadata) is returned to the Contact Middle and in the end the tip consumer. Lastly, outcomes and logs are written again to S3, feeding Airflow-driven monitoring and retraining, and Jenkins redeploys up to date parts to shut the loop.
Toy instance
To exemplify the mechanics of the unique manufacturing implementation in a simplified and self-contained method we create an artificial, runnable toy instance demonstrating the core logic behind policy-to-agency project utilizing linear integer programming with the PuLP library in Python.
The instance units up a small state of affairs with 4 companies and three coverage classes (“Gold,” “Silver,” and “Bronze”). Productiveness scores and capability limits are assigned for every company, together with constraints comparable to ZIP code eligibility and minimal/most coverage combine per class. The aim is to maximise the overall productiveness rating whereas respecting these constraints.
Whereas the instance is artificial and makes use of randomly generated weights and capacities, it successfully illustrates the basic optimization logic and workflow, together with variable development, constraint enforcement, and answer interpretation. This strategy might be immediately scaled and tailored to real-world knowledge and enterprise constraints as demonstrated within the full implementation.
In Desk 1, we illustrate a easy iteration. Batch mode first computes a baseline month-to-month plan that allocates the preliminary stock. On-line mode then simulates incoming insurance policies one by one towards a goal month-to-month whole; every arrival triggers a re-optimization that preserves current allocations and assigns solely the incremental coverage to an eligible company (e.g., respecting ZIP admissibility). On this instance, the brand new coverage is a high-value (Gold) coverage and its ZIP is admissible for A1, so the increment goes to A1. If the ZIP had been inadmissible for A1, the coverage can be routed to the perfect admissible company as an alternative. This course of repeats till the month-to-month bucket goal is reached.
Code
The code is accessible on this repository: Link to the repository
To run the experiments, arrange a Python ≥3.11 atmosphere with the required libraries (e.g., pulp, and so on.). It is strongly recommended to make use of a digital atmosphere (through venv or conda) to maintain dependencies remoted.
Conclusion
In comparison with a round-robin baseline that assigns insurance policies with no intelligence, our strategy makes use of a productiveness matrix derived from an swap ratio to route insurance policies the place they’re anticipated to create essentially the most worth. The optimization balances tangible metrics (the measurable worth and capability every company can ship) with intangible concerns (equity, stability, and the belief companies place in a predictable allocation course of). In brief, it replaces a blind rotation with a clear, auditable determination rule that displays each efficiency and operational constraints.
By making coverage assignments extra clear and predictable, we’ve constructed belief and collaboration. Businesses (iia’s) now perceive how choices are being made, which has elevated their confidence within the course of.
This instance exhibits how even a comparatively small optimization downside can generate significant enhancements. By beginning with a easy, well-defined formulation, we create a stable basis that delivers speedy worth whereas enabling future evolution. The identical framework might be prolonged by way of incremental iterations, incorporating richer indicators, and extra superior determination logic. In observe, the best influence typically comes not from constructing a posh system upfront, however from beginning easy and bettering constantly because the enterprise learns and the information matures.
References
[1]PuLP documentation, “PuLP 3.3.0 documentation.” COIN-OR. https://coin-or.github.io/pulp/main/includeme.html
