The Ladder
01 — Shadow to auto-post
Autonomy is earned, never switched on.
We do not flip an autonomy switch. We move the system up a trust ladder: one supplier, one client, one consistent outcome at a time. Each promotion requires a measured history; demotion is automatic the moment a correction appears.
- The promotion metric is not the model's confidence. It is how often a human accepted the proposal without changing it.
- Demotion is immediate and automatic on any subsequent correction. There is a global kill switch.
- Aligned with Article 14 of the EU AI Act.
The Number
02 — Composed, not declared
Where the number comes from
Confidence is composed, not declared. A model reporting its own confidence is reporting an opinion.
A rubric, not intuition.
95%+ requires five or more concordant prior entries. With no history for that client, a hard ceiling of 85%.
A second opinion, with fixed arithmetic.
A reviewing agent independently re-derives the classification on every document. Agreement raises the number; disagreement discounts it.
History becomes statistics.
Where prior entries exist, the number is frequency, not opinion — an account used in 11 of 12 prior entries is 92%.
Provenance is mandatory.
An account with no source is capped at 50%. The weakest link decides — a document's confidence is the minimum across every element of the entry, never the average.
The Human
03 — Against automation bias
The human only stays in command if the human keeps looking.
The better the proposals get, the less the human checks them. Automation bias is the real failure mode of human-in-the-loop, and most systems ignore it. We design against it.
Disagreement penalises, it doesn't just warn
When proposer and reviewer diverge, confidence is deterministically demoted, pushing the document to human review.
Re-verification where complacency is highest
The reviewer runs again precisely on high-confidence entries, where the human's guard is lowest.
Forced justification on what's risky
Large or high-risk entries require a written reason to approve. It breaks the rubber stamp.
Measuring attention
Deliberately injecting incorrect documents to measure whether the reviewer is still looking.
If this is the kind of system you want in your finance function — start small, and make us earn it.