DemoPricing

What is Beneficial vs Harmful Drift in AI Decisions?

Drift is a lasting difference between what a written rule says and what is actually decided. It is beneficial when the departure leads to better outcomes than the rule would have, and harmful when it leads to worse. Telling the two apart needs outcomes, not just decisions.

Why the direction matters

Treating all drift as a breach throws away what experienced people have learned. Treating all of it as acceptable lets bad habits stand. The useful question is whether the departure changed results, which needs the outcome of each decision.

How Studio tests it

Studio runs a causal estimate (DoWhy) on the decisions where practice and policy differ: it adjusts for the other factors that predict the outcome, checks that the sample is large enough for the effect it sees, corrects for testing many rules at once (Benjamini-Hochberg, q ≤ 0.10), and runs refutation tests. When the result is not significant, it says the drift is inconclusive rather than guessing.

What happens next

Harmful drift with a robust causal effect can be drafted into a proposed rule change, for example an escalation for the cases that did worse. The proposal is reviewed and sealed by people before it takes effect; NOMOS never changes the rules in force by itself.

Example

Scenario

A credit policy caps the debt-burden ratio at 50%, and decision logs show officers approved 12 of 30 applications above it.

Without governance

The exceptions are either banned outright, or ignored. Nobody knows whether they did better or worse.

With NOMOS

With repayment outcomes joined to the decisions, Studio estimates the effect of the exception on default and reports whether it is significant and robust. If the exceptions did better, the bank can write the condition into the policy; if worse, it can draft an escalation for them.

Frequently asked questions

How much data is needed?+

Enough decisions on both sides of the rule for the effect being measured; Studio states when there are too few and calls the result inconclusive. With realised outcomes, it needs at least 20 decisions joined to their outcomes.

Related concepts

See it in practice

Upload a policy, confirm its rules, seal them: any system can then check an action against them.

Open Studio →