Ask why a new insurance product launched late and the answer is rarely that the actuaries were slow. The base rates, relativities and underwriting rules are often agreed long before anyone can get a quote. The real wait sits between "the rating guide is signed off" and "the model is live, tested and plugged into distribution".
That gap is where AI is making the biggest difference.
Where the time goes
Rating guides arrive as PDFs, spreadsheets or regulatory filings. Someone has to turn them into something a system can run. Every table gets keyed in, every rule gets coded, the calculations get put in order, and then the results have to match the actuary's examples exactly.
Usually that someone is a developer who isn't a pricing expert, working from a document written by a pricing expert who isn't a developer. Each unclear line becomes a meeting. Each rate change restarts the cycle. The upshot is that plenty of insurers and MGAs change rates less often than they would like, simply because every change is a project.
What AI is good at here
Large language models read documents well, pull structure out of them and convert from one format to another. That covers most of the slow work:
- Extraction. Rate tables, factor definitions and rules lifted out of PDFs, spreadsheets and filings into structured data.
- Drafting. A written algorithm like "base rate by region, times an age factor and a breed factor, plus tax" turned into a working model.
- Checking. The draft validated against a schema and priced against the guide's worked examples, with any mismatch explained in plain words.
- Iterating. A rate change applied to a copy of the model, the tests re-run and the differences summarised.
A first working draft can now come together far faster than a hand build. It is still a draft.
Why AI alone isn't the answer
Pricing is regulated, and small errors are costly. A model that is nearly right is wrong. Misread one footnote and you can misprice a whole book. And a model nobody can read is a model nobody can sign off.
What works is a clear split of duties:
- AI does the extraction, the first draft and the first round of checks.
- A pricing specialist reviews every table and rule against the source and settles anything the document leaves open.
- Tests built from the actuary's worked examples prove the model prices correctly.
- A person approves and publishes. Not the AI.
Swallow is built around that split. Over the Swallow MCP server, an AI client can read the modelling rules, start from an example model, validate its draft and run the tests. It cannot create, edit or publish a project. The finished model goes into a draft in the dashboard, and from there it moves from draft to approved to published, with your team's roles deciding who can approve.
The format decides how fast you can review
For this to work, the model has to be readable by both the AI that writes it and the person who checks it. Generated program code is quick to produce and slow to review. A structured model, with named inputs, rate tables as rows and formulas written out, can be drafted by an agent and checked by an underwriter without either of them reading code.
Every Swallow model is a single JSON file in that shape, with its tests saved on each version. The agent drafts it, a person reviews it and the engine prices from it. Our pet insurance example walks through one end to end, and we go deeper on the format in why pricing models should be human readable.
What changes when implementation stops being the bottleneck
- More products get built. New lines, new territories and small trial propositions become worth doing once each one isn't a long build.
- Rates move when they should. Responding to claims inflation or a competitor move becomes routine, with every change tested before release. Our rate change process sets out the steps.
- Actuaries price. Time goes on rates and strategy, not chasing implementation errors.
AI doesn't replace the pricing actuary. It takes away the slowest, least valuable part of the job, and makes regular, well-tested rate updates realistic.
Start with what you have
A rating guide, a spreadsheet rater or a filing is enough to begin. Connect your AI client over MCP and build from an example, or see how we approach replacing a spreadsheet rater. If you'd rather talk it through, get in touch.
