From Manual Pipelines to Modeling Partner

1 Like

How an AI Copilot Is Compressing the Actuarial Modeling Cycle

As data volumes grow and stakeholders demand faster, more transparent answers, actuarial teams need more than better algorithms they need a modeling partner that assembles pipelines, checks assumptions, and explains results in plain language, without sacrificing rigor.

The Actuarial Modeling Bottleneck

Actuarial modeling has never been short on rigor. What it has too often lacked is speed. Building a pricing or risk model still typically starts with hours of exploratory data analysis, followed by manual assembly of an imputation, transformation, and feature engineering pipeline each step requiring judgment.

The cost of this bottleneck compounds downstream. When a pricing actuary or Chief Risk Officer asks why a model behaves the way it does, or when a regulator asks the same question, the answer frequently has to be reconstructed after the fact rather than produced as a natural by-product of the modeling process itself. This is not a data science problem so much as a workflow problem: the modeling process itself needs to become faster, more consistent, and self-documenting.

How Modeling Copilot Works

Modeling Copilot is an AI assistant that supports actuaries and analysts while they build risk or pricing models. Rather than replacing the actuary's judgment, it sits alongside the modeling workflow generating pipelines, suggesting variables, checking assumptions, and explaining results in simple terms, acting as a smart partner from first exploration to final interpretation.

The workflow starts conversationally. A single prompt for example, a request to run exploratory data analysis on a defined sample is enough to trigger the Copilot's analysis. From there, it proposes a next set of steps, such as imputation and outlier replacement, and on the actuary's selection, automatically assembles the corresponding pipeline. Every node the Copilot proposes, from imputation and replacement through transformation, feature engineering, and the final gradient boosting model, is pre-configured and remains open for the actuary to inspect, adjust, or accept at each step. Nothing is a black box: the actuary retains full visibility and control over every modeling decision, while the Copilot removes the manual effort of assembling and configuring the pipeline from scratch.

SAS Viya Copilot proposes and ranks candidate pipelines after running exploratory trials on a 50,000-row sample — each option scored and ready to assemble with one selection.

Business Impact: Why Speed and Consistency Matter

Independent research on generative AI in insurance points in a consistent direction: modeling and analytical workflows are among the functions with the most to gain from AI-assisted automation. In a McKinsey survey of more than 50 leaders from Europe’s largest insurer groups, more than half reported that generative AI could realistically deliver productivity gains of 10 to 20 percent across core functions (McKinsey, 2024, “The potential of generative AI in insurance,” McKinsey & Company). Modeling Copilot targets exactly this category of gain the manual pipeline assembly, variable selection, and assumption-checking work that consumes actuarial time without adding differentiating judgment by automating it into a guided, conversational workflow the actuary directs and approves at every step.

The same McKinsey research found that users who leverage generative AI effectively can see productivity improve by over 20 percent, though realizing that gain depends on adoption and training as much as on the tool itself (McKinsey, 2024, “The potential of generative AI in insurance,” McKinsey & Company). Because Modeling Copilot embeds AI assistance directly inside the modeling workflow rather than as a separate tool actuaries must remember to consult it lowers the adoption barrier that typically limits how much of that productivity upside organizations actually capture.

More broadly, McKinsey estimates that nearly half of manual work activities could potentially be reduced through generative AI alone (McKinsey, 2023, “The economic potential of generative AI: The next productivity frontier,” McKinsey & Company). Applied to the actuarial function, this reinforces the underlying case for Modeling Copilot: the highest-value use of actuarial time is judgment — model design, assumption-setting, and interpretation — not the mechanical assembly of pipelines that a well-governed AI assistant can propose and configure in seconds.

One-click execution: the Copilot sequences every pipeline node, confirms success at each step, and surfaces results through the standard visualization interface full progress visibility from a single command.

Governance and Explainability: The Common Thread

Modeling Copilot's plain-language explanation of model behavior overview and behavior summary, key findings and variable importance, predicted mean and evaluation statistics, and actionable recommendations is often framed as a convenience for time-pressed stakeholders. It is better understood as a governance asset.

This is also consistent with the direction of EIOPA's guidance on model governance and transparency, which increasingly expects insurers to demonstrate not only that a model performs well, but that its behavior can be explained to those who did not build it. A Copilot that generates a structured, human-readable explanation as a standard output of the modeling process rather than as a bespoke exercise performed only when a regulator asks turns an obligation into a routine capability. This is the thread that runs through every entry in this series: AI in insurance earns its place not by being clever, but by being auditable, consistent, and explainable by design.