Audit-Proof Workflow
About this app
Question in plain language → AI parser → deterministic engine → AI explanation → sign-off → audit log. Replay returns the byte-identical result; let the AI compute directly and the numbers scatter, the log turns red.
The idea
In regulated fields such as banking supervision, every figure has to be traceable and repeatable. Language models, however, do not return the same output on every call. An audit-proof design therefore uses AI at the edges: it turns a question into parameters and explains the result in words. The calculation itself runs in a deterministic, versioned core – same input, same output. A human reviews and signs, and every step is logged. The app demonstrates this pattern with a simplified example of the capital requirement for credit risk.
What you see
A question such as “What happens to our capital if the SME portfolio grows by 20 %?” travels as a packet through five stations: request, AI parser, deterministic engine, AI explainer and human sign-off. The engine uses the simplified Basel/CRR IRB formula without maturity adjustment: K = LGD · [Φ((Φ⁻¹(PD) + √ρ · Φ⁻¹(0.999)) / √(1−ρ)) − PD] and RWA = 12.5 · K · EAD. The result panel shows the parameters, correlation ρ, capital ratio K, RWA and signature; the audit log records every run with an input hash.
Try it
Pick one of the three questions, run it and press “Replay” a few times: in deterministic mode every repetition is identical and the log line turns green. Then switch to “AI computes itself”. The packet skips the engine, the language model estimates the figure, and every replay differs – the line turns red. In the app this spread is simulated as a deviation of a few percent. Consider what a signature means under a number nobody can reproduce.
Frequently asked questions
- How can AI be used in an audit-proof way?
- By limiting AI to language tasks – understanding questions and explaining results – and leaving the actual calculation to a deterministic, versioned program. A human reviews and signs off, and every step is logged with its inputs and version. Any result can then be recomputed exactly later.
- Why shouldn’t a language model compute regulatory figures itself?
- Language models can produce different outputs from run to run and are not guaranteed to calculate exactly. Audits and supervisors need a figure that comes out the same every time for the same inputs. A fixed calculation core with a documented formula provides that.
- What does the IRB formula calculate?
- The IRB formula from the Basel framework and the CRR gives the capital requirement K for a credit exposure from probability of default PD, loss given default LGD and correlation ρ. Multiplied by 12.5 and the exposure at default EAD, it yields risk-weighted assets (RWA). The app uses a simplified version without maturity adjustment.
- What does the replay button in the app show?
- Replay runs the same input again and compares the result with the first run. In deterministic mode both match; in “AI computes itself” mode they differ. The audit log marks matches in green and deviations in red.
Subject: AI transformation in business
More from AI transformation in business
A VisuApp by heyprof: interactive, free in your browser, no sign-up. What is a VisuApp? · All apps