This piece is by Alessandro De Leonardis, CIO, Armundia Group
Should we automate more in insurance brokerage?
Any institution must first reckon with a constraint that precedes any model: data quality. Information comes from different providers, must be used in real time, and rarely arrives clean. The challenge is particularly acute in Family Offices, which aggregate data from multiple banks to produce a consolidated analysis and therefore have to manage the collection, normalisation, and continuous monitoring of various, dynamic flows. Added to this is the need to optimise portfolios using Machine Learning algorithms capable of accounting for the relationships between holdings spread across different institutions. It’s on this foundation – one only as strong as its data – that everything else rests.
In the automation of brokerage, three key themes emerge—and they explain why computational power alone is not enough.
1. Propensity to purchase
The first problem is estimating the probability that a client will purchase a given insurance product. Technically, this is a supervised classification problem, trained on historical conversion data: the model learns from past behaviour to predict future behaviour. The work does not stop at the algorithm but begins upstream, in the research and definition of the optimal dataset, and ends in an accurate probabilistic calibration—indispensable if the output is then to be usable within an objective function.
But there’s one requirement that runs through the entire process and is non-negotiable: explainability. Being able to explain why the model considers a client likely to buy is not a technical refinement but a condition of regulatory compliance. Already at this first level, then, transparency is not an accessory to the model: it’s a constitutive part.
2. Optimising the proposal
There’s also the challenge of constructing the optimal insurance proposal within budget and regulatory constraints. Here the problem is of a different nature: a constrained optimisation that selects products and coverage limits while operating within rigid boundaries. The client’s budget is a hard constraint (it can’t be exceeded) and it sits alongside structural constraints such as dependencies between products and eligibility rules.
But the heart of the matter is the multi-criteria nature of the objective function, which must hold together two demands that are not always aligned: the suitability required by the IDD regulation and commercial propensity. The trade-off between the client’s objective need and the probability that they will buy isn’t a neutral parameter to be set once: it’s a matter of responsibility, and must stay configurable. Entrusting that balance to the algorithm alone would mean concealing a value-laden decision behind the appearance of automation.
3. Estimating the real insurance need
The third problem is the most insidious, and the most instructive. It involves measuring a client’s actual need on the basis of claims data, with an ML approach. The training set, in this case, is built on observed claims: it measures facts, not commercial choices. The framework is properly actuarial and requires modelling the frequency and the severity of claims separately.
It is precisely here, however, that a structural limitation surfaces that no amount of computational power can sidestep: claims exist only for those who were already insured. Those who were not remain simply invisible to the model. To estimate the risk of the uncovered population, one is therefore forced to resort to geographical and environmental variables as proxies, with all the approximation that entails. To this day, the partial-coverage bias remains an open research problem. It is the clearest demonstration of a general principle: the model cannot see what was never recorded.
Conclusion
From different angles, we reach the same conclusion. Propensity makes transparency a regulatory requirement. Optimisation calls for human arbitration over the trade-off between the client’s interest and the commercial objective. The estimation of need reveals a bias that only deliberate oversight can recognise and govern.
Automating insurance brokerage is possible—to some extent, it’s inevitable. But each of these problems shows that automation reaches a boundary beyond which human judgment is needed: to explain, to arbitrate, to see what the data does not contain. The brokers of the future are not born from replacing one with the other, but from their collaboration.
Be the first to comment