← Home

How iziRisk uses AI with your data

Simulations never use AI: the Monte Carlo engines run in your browser or on our servers. Only the features marked ✦ send content to an AI model, and only when you press them.

Who receives it

  • Anthropic (Claude) — the ✦ model builders, suggestions, result analyses and reports, and the AI-assisted import of loss registers.
  • OpenAI — turns search text into vectors for the knowledge base: your Ask-the-Experts questions, the task and risk names behind the distribution, risk and correlation suggestions, and approved expert contributions.

The iziRisk knowledge base (RAG)

Several ✦ features are grounded in : before the AI answers, iziRisk retrieves the most relevant passages from it and gives them to the model with your request (retrieval-augmented generation, RAG). That is why the answers reflect established QRA practice instead of a generic assistant’s guess.

  • Which features use it: Ask the Experts, and the distribution, risk and correlation suggestions in iziRisk Projects.
  • How the search works: the search text is converted into a vector by OpenAI; the matching itself runs in iziRisk’s own database, and the passages found travel to Anthropic together with your request.
  • What goes into it: courses, reference material and expert contributions approved by an iziRisk administrator. Your models, registers, results and questions are never added to it.
  • Unanswered questions: when the knowledge base has nothing close to an Ask-the-Experts question, the question is saved so the team knows which content to add.

What is sent

  • Model builders and suggestions: the description you type and the structure of the model (task, risk and variable names, ranges, distributions).
  • Analyses and reports: the model’s summary results — percentiles, probabilities, drivers and the names that label them — not your files.
  • AI-assisted import of a register: the column headers, a summary of each column with a few sample values, the list of category values, and the first 8 rows. The rest of the register is read and modelled in your browser.

What is not sent

  • The files you upload, as files. Payment details (they go to Stripe only). Your login codes. Any data from features without the ✦ mark.

Training and retention

Under Anthropic’s and OpenAI’s commercial API terms, content sent through their APIs is not used to train their models by default. Each provider keeps API data only for the limited period its terms set out.

Your controls

  • Use the hub without the ✦ features: every model, simulation and chart works without them.
  • Before an AI-assisted import, remove columns with names, ID numbers or free-text descriptions you do not want to share — the first 8 rows are sent as they are.
  • Enterprise organisations can switch AI off for all their members, and can have portfolio analyses sent with organisation, programme and project names replaced by codes.

Questions, or need the full subprocessor list for a security review?

Ask a question