Where LLMs actually help in a pricing team

Most of the “AI for actuaries” talk is either hype or dismissal. From building these tools for real pricing work, my view is simpler: LLMs help with the language-heavy, ambiguous parts of the job, and they should stay out of the arithmetic.

Where they help

• Messy data mapping. Working out that SI_Bldg_AUD, Building Sum Ins. and bldg_si are the same field. The model proposes the mapping and a person approves it.
• Documentation Q&A. Retrieval over methodology notes and past review packs, with citations so the answer can be checked.
• Drafting. First drafts of review commentary, model documentation and code comments, based on outputs that code produced deterministically.
• Code acceleration. Scaffolding data pipelines, tests and plotting code, which is where most actuarial time actually goes.

Where they don’t

• Calculations. Loss ratios, rate changes and reserves are computed in code and passed to the model as facts, never computed by it.
• Judgement calls. Trend selection, large-loss thresholds and credibility weights are professional judgement, and the accountability stays with the actuary.

Guardrails that matter

1. Tools over prose. Give the model typed functions to call, so every step is logged and reproducible.
2. Citations or it didn’t happen. Any factual answer must point to a source.
3. A small evaluation set. Twenty real questions with known answers catch most regressions.
4. Human sign-off before anything reaches a rate filing or a board pack.

Used this way, LLMs don’t replace actuarial judgement. They free up more time for it.