Plain language in. Governed actuarial work out.
Give actuarial teams a natural-language interface without turning production into a chat. Skills define the procedure, a harness executes and records it, and the approved engine performs the calculations.
Control layers demonstrated in the public reserving harness.
A plausible answer is not a repeatable process
An AI assistant can explain a calculation and still skip a diagnostic, lose a warning between messages or produce a pack that cannot be reproduced.
Actuarial production needs a defined route: run, diagnose, respond to evidence, assemble the review artefact and preserve exactly what happened.
Natural language over an approved production path
The model is an interface and coordinator. The controlled process remains explicit and inspectable.
Define the procedure
Encode the standard review route, commands, checks and stopping conditions in reusable skills.
Resolve the request
Translate the actuary's instruction into the approved run, diagnose and pack sequence.
Run the engine
Call the named reserving or projection engine rather than asking the language model to invent numbers.
React to evidence
Route warnings into prescribed diagnostics and sensitivities instead of narrating past them.
Capture the manifest
Store inputs, commands, outputs, thresholds, warnings and reroutes for the review trail.
Assemble the pack
Produce a reviewable actuarial artefact with figures, uncertainty, caveats and reviewer context.
A review pack, not a chat answer
The reserving example demonstrates the operating pattern on a Mack workflow while keeping the engine replaceable.
Diagnostic rerouting
An outlier link-ratio warning triggers the prescribed sensitivity run and carries into the pack.
Review memory
Warnings, sensitivities and caveats remain connected to the run that produced them.
Model adapter
The workflow layer can sit over another approved engine or an insurer's existing production process.
Consistent language
The same evidence produces the same controlled explanation and pack structure each cycle.
The workflow does not own actuarial judgement
The harness controls execution around a defined production process. It does not approve assumptions, select judgmental factors autonomously or make the final reserving decision.
What this does not claimWhat a production implementation includes
Engagement outputWhere control sits in an agent-enabled workflow
Questions & boundariesIs the actuarial agent autonomous?
Not in the decision-making sense. It interprets a request and routes it through approved procedures, tools and diagnostics; assumptions, overrides and sign-off remain explicit human responsibilities.
What does model-agnostic mean here?
The workflow and control layer is separated from the calculation engine. An approved internal engine can replace the public Mack demonstration through a defined adapter without rewriting the governance pattern.
What happens when a diagnostic fails?
The workflow follows a prescribed route: run the required sensitivity or exception procedure, retain the warning in the manifest and prevent a clean sign-off until a reviewer disposes of it.
What makes a good first pilot?
Choose one recurring production cycle with a known engine, stable inputs, repeatable diagnostics and a review pack people already understand. The pilot should improve control and traceability before expanding scope.
Related control patterns
Move the model. Keep the evidence.
Rebuild spreadsheet and legacy-platform models in Python while proving, step by step, that the original behaviour survived. AI accelerates construction; a deterministic harness decides whether it matches.
Explore this solutionSolution · Documentation & regulatory reviewDefine good. Then make the work pass.
Turn actuarial and regulatory expertise into evidence rules, procedures, checks and stopping conditions that AI must follow. The report becomes a review interface, not an unsupported draft.
Explore this solutionWhere does your team repeat controlled work around an existing actuarial engine?
Start with one production cycle, its diagnostics and the review artefact the team already knows.