Automated Annuity Modeling | Elliott Actuarial

AI coding agents can revolutionize how we develop and validate actuarial models. The below proof of concept shows model configuration for a new product in an hour’s time, including single-cell Excel replication and a summary validation report.

Agents can impact model development across the board. This example focuses on configuration and validation.

	Today	Agent Enabled
Documentation	Time-consuming and rarely validated against model	Generated automatically; continuously compared against code
Code changes & configuration	Manual, error-prone process	Agents turn an Excel example into a code change and update validation
Validation	Excel replication may take weeks to complete	Agents replicate model functionality in Excel in minutes
Automation	Available at the whim of your model provider	Agents batch runs, build exhibits, and begin interpreting results

At a glance — FIA GLWB model build

Coding agent receives product spec
Agent has full access to the necessary context: model code, model spec, and custom skills covering model build and validation
Agent digests product materials, configures model features, inserts example assumptions, and creates the Excel replication and validation report in minutes

An example validation report can be found here. The report is long but a quick skim shows the depth of coverage.

This report would be the starting point for an actuary’s user-acceptance review. This example simulated the process of replicating a competitor product and used a product filing/brochure instead of a spec.

I have not edited the report. The results are generally strong but there are a few shaky conclusions and some quirks in the language. The agent tends to be over-detailed. Even with these shortcomings I believe this would save an actuary many hours of time in their build and validation process.

The setup

To use coding agents effectively they must have full access to the underlying model: the code, the assumptions, and the documentation. This setup is unlikely to work with most model-as-a-service providers, where the model itself isn’t easily consumed as context by the agent.

Coding agents are able to digest the entire code base and documentation at once and may also display a high degree of actuarial knowledge (see AnnuityBench for an example).

Detailed example

This example shows the process of replicating a competitor product where you have access to a brochure and filing. A product spec would work in place of the filing/brochure without issue.

Product

Stacking rollup (4% + 125% of credited rate)
Income growing at 75% of credited rate after election
Hedging done on an economic basis

Pricing model

Custom FIA model in development over the past few months

Tools used

Claude Code
Opus orchestrator agent that delegates to Opus/Sonnet subagents

Context

Full model code and specs available to the agent

Skills

Custom agent skills govern the research, configuration, and validation process

Process:

Agent receives product filing and brochure and configures the model
Model is automatically run for a single cell to produce full audit traces and single-cell replication
Reviewer skill automatically kicked off to compare configuration to provided materials and assess the quality of the build
After completion an actuary can use this report to quickly review and confirm the model implementation

Output:

Output	Description
Product configuration	Single cell with GLWB features modeled Actuary can now configure cells and assumptions
Audit traces	Excel replication of model functionality for a single cell I'm excited about this replication because formulas are automatically adjusted based on product design to provide an easy review
Review binder	Key excerpts from audit files to use in the report and user-acceptance review
Validation report	First pass review completed by coding agent to speed review process

Cost & time:

The full research/build/validation pipeline was run in Claude Code using Opus 5. The entire process took approximately one hour and cost about $60.

If you're interested in learning more about these capabilities, please reach out.