We built a free AI risk calculator that runs in minutes, using Fermi estimation with honest confidence intervals

We have been arguing internally for months about how to give people a fast estimate of their AI risk exposure without pretending the number is precise. Most risk-score tools return a single value that hides where the uncertainty lives. We wanted to build something that is structured, shows its work, and admits what it does not know. You answer a short form covering deployment type, jurisdiction, company size, automation level, and data sensitivity. This takes about three minutes, after which an agent (GPT-5.5 under the hood) runs for several more, streaming progress while it computes the estimate. The output is an expected annual loss with a 90% confidence interval, broken into five categories: technical, operational, legal and compliance, ethical and reputational, and governance. Every category surfaces its drivers, assumptions, and mitigations, and you also get a downloadable PDF. The method is Fermi estimation. For each risk we estimate incident frequency and the financial impact when an incident happens, with impact split into fines, legal costs, remediation, and indirect losses like brand damage. Base rates come from industry precedent and get adjusted for your context, so jurisdiction matters considerably. EU AI Act fines, for instance, scale to 7% of global turnover for prohibited practices. I want feedback from this sub specifically because risk quantification is hard, and honest people will disagree about the priors. Here are a few things I expect to be wrong or contested. 1. Base rates for AI-specific incidents are noisy, and we are extrapolating from a thin precedent that will look better in two years. 2. The single-year horizon hides compounding effects, which is a deliberate choice for a screening tool, but a real limitation worth flagging. 3. Governance risk is the hardest to monetise, and we took a swing at it; tell me where the estimate is off. 4. The 90% intervals come out wide, and people hate that, but we think narrow ranges are dishonest, and the trade-off is worth arguing about. The tool does not require a login, and no email is needed to see the result, though the PDF download asks for one. I would especially value three kinds of feedback. * Run it against a system you know well, and tell me whether the number passes your sniff test. * Tell me which assumption you would change first. * Tell me which of the five categories we got most wrong. \[Disclosure: I work at Modulos, where we make AI governance software, and this calculator is a free lightweight version of what the full platform does.\]