PræYou ask the question. The agent builds a glass box.
Nightingale is our first AI agent, built to automate the risk analysis behind insurance underwriting. Give it a risk and it gathers the evidence, models what drives the loss, and returns a distribution with a source behind every number. The worked example below follows one question from start to finish. It shows how a run works, and its figures are for illustration only.
IN DEVELOPMENT · TALKING TO DESIGN PARTNERS
The question
How likely is a German mid-market company to lose more than EUR 1 million to cyber incidents in 2027?
In this illustrative run, the estimated probability is 10.3%. Read the same evidence a thousand different ways, and the answer stays between 8.9% and 12.0%.
The film walks through that run. It builds a single year from the evidence, samples ten thousand of them, and counts the years that cross EUR 1 million. Most years, 86 in 100, stay quiet. A typical year with a qualifying incident costs EUR 1.0 to 2.0 million, and the worst year in a hundred costs about EUR 3.1 million.
THE PROBLEM
Every underwriting decision comes down to a few numbers: how often a risk will produce claims, how large those claims will be, and how bad a single year can get. Working them out properly takes underwriting and actuarial hours that few submissions can justify, so most get a quick look. Two underwriters reading the same file will often land on different answers, and none of the work carries over to the next submission.
THE SHIFT
Until recently, doing more analysis meant hiring more analysts. Nightingale does that work itself, on every submission, while your underwriters keep the decision. They frame the question, review the work, and make the call.
HOW IT WORKS
Each run goes through three passes, and every pass leaves a written record you can open.
PASS 01
Collect the evidence
The agent starts from the submission and pulls in what surrounds it: official statistics, industry loss studies, market claims data. If your own systems hold policy or claims history, it reads that as well, and it can run entirely inside your infrastructure when the data cannot leave. Every figure is tied to a named source, and when two sources disagree, both stay on the record.
PASS 02
Draw the graph
From that evidence the agent draws a graph of the risk, showing what drives claim frequency and severity, how those drivers interact, and where the gaps are. No number enters the graph until a second, independent source confirms it.
PASS 03
Price the options
The agent then samples the graph thousands of times, with each sample standing for one possible policy year. Counting across those years gives the probability of each loss level, and the same samples price every option on the table: quote as submitted, quote with a higher retention or lower limit, or decline.
Every estimate is stored with its sources before the outcome is known, so an underwriter, reviewer, or auditor can retrace any step. When the evidence is too thin to support a number, the agent says so and lists what is missing instead of guessing. As claims develop on the policies you write, the estimates are checked against them across the whole book, so you can see whether a 10% risk really behaves like one.
IT LEARNS
Those checks feed the next run. Priors tighten, assumptions that failed are dropped, and the agent extends its methods to lines of business it has not modeled before.
You do not have to wait years for the first results. The agent can backtest on your own claims history, checking its estimates against years whose claims have already developed.
Partner with us early.
We are taking on a few design partners among carriers, MGAs, and reinsurers. Send us a line of business or a submission you find hard to price at contact@praevidere.org. We will run the analysis and show you the result.