PræYou ask the question. The agent builds a glassbox.
Nightingale is an AI agent for statistical modelling. Give it a question about your business. It collects the evidence, draws the graph, and returns the answer as a distribution, with every source attached. Below is a worked example, from question to answer.
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?
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 follows the example. It shows how a year is built from the evidence, how ten thousand such years are sampled, and how the count becomes the answer. 86 of 100 years stay quiet. A qualifying year costs EUR 1.0 to 2.0 million. The worst year in a hundred costs about EUR 3.1 million. The same study runs on your own claims history.
THE PROBLEM
Many business decisions depend on a question that should have a numerical answer. What will this cost? How often does it happen? How bad can it get? Deep analysis is scarce. Expert hours are expensive, so most risks get a shallow look. Two experts read the same evidence and reach different views. And the work is rebuilt from zero for the next risk.
THE SHIFT
More analysis used to mean more analysts. That link is broken. An agent can now do the analytical work. Your experts keep the judgement: they set the question, check the working, and make the decision.
HOW IT WORKS
One run moves through three passes. Each pass is written down, and you can open any of it.
PASS 01
Collect the evidence
The agent reads the question and the material around it: official statistics, industry research, claims data. Where your own systems hold part of the answer, the agent reads those too. Every figure is tied to a named source. Where two sources disagree, the disagreement stays on the record.
PASS 02
Draw the graph
The agent draws a graph of the question: what drives the outcome, how the parts relate, and what is not known. Every number in the graph comes from the evidence. And each number must be confirmed by an independent source before it is used.
PASS 03
Price the options
The graph is sampled thousands of times. Each sample is one possible year. The run counts how often each outcome occurs, and prices each option on the same samples: accept, decline, restructure.
Every estimate is stored with its sources. When the outcome arrives, the estimate is scored against it.
IT LEARNS
A tested graph makes the next run better. Priors sharpen. Assumptions that failed once are not made twice. And the agent extends its own methods to classes of problem it has not met before.
The record does not start empty. The same study runs on your own claims history, tested against years that have already happened. The model of the day can be copied. The record behind it cannot.
WHY NOT GENERIC AI
A general assistant can read the same evidence and write a fluent answer. That part is now cheap. Even when both outputs are a number that looks like a probability, the two differ in type, not in degree.
The evidence in this work is messy, and the events happen once. You cannot rerun the year. The method is Bayesian by necessity: state the assumptions, weigh the evidence, and say what the answer rests on.
A prediction is not a model. Ptolemy predicted the planets well. Newton modeled them. Only the model survives the question “what if Mars were heavier”. Your decisions are questions of that kind: accept, decline, restructure.
The graph earns the number. The agent states its assumptions, and every estimate must be confirmed by an independent source before it is used. The answer rests on a graph that survived, not on a pattern fitted to the past.
The working stays open. The agent writes down the plan, the evidence log, the graph, and the samples as it works. You choose the model provider, closed-source or self-hosted. If the data cannot leave your infrastructure, the full workload stays there. You can check any step before you rely on the number. Where the evidence cannot support a number, the run refuses and names what is missing.
Partner with us early.
We are taking on a small number of design partners: teams facing a decision that existing tools cannot price. Tell us what you are about to spend, launch, or abandon — write to [email protected]. We will show you what the numbers say, and what the answer is worth before you pay for it. We are also speaking with investors at the same address.