Accounting › Close & Compliance
Finance Intelligence
Nothing here decides anything. Detectors produce scored observations with their evidence attached; a person opens the case and records what turned out to be true.
In Accounting → Close & Compliance · 10 screens
WorkspaceRisk QueueWhat is open, what it is worth, and how quickly it is being worked — sorted by exposure rather than by score, because that sends people to the most expensive thing.OpenViewRisk SignalsA scored observation with its EVIDENCE attached. "Score 0.86" persuades nobody; the facts that produced it are what a person acts on.OpenRecordsRisk CasesA person’s queue, with an outcome. The outcome is what closes the loop: confirmed fraud and dismissed noise are both training data, and a system that does not record which cannot improve.OpenRecordsCase ActionsWhat was done about it in the world, as opposed to in the case.OpenRecordsRisk DetectorsEach answers one question, is separately switchable and separately measured. A detector nobody can turn off is one that gets ignored wholesale the first time it is noisy.OpenViewDetector PerformancePrecision — of the signals it raised, how many were real. The only number that decides whether a detector should keep running, and the reason outcomes are recorded at all.OpenViewAssistant SuggestionsSuggestion rather than detection, with the reasoning attached. "Because the last nine invoices from this supplier were coded 6410" is a reason a person can check; a confidence score is not.OpenRecordsCoding PatternsA learned lookup rather than a model. For GL coding this beats a model on both accuracy and explicability, and a person can read it.OpenRecordsNarrativesThe generated text and the edited one, both kept — the difference between them is the most useful feedback there is, and a narrative rewritten entirely was not worth generating.OpenViewAssistant AccuracyConfidence is only useful if it is calibrated: a suggestion offered at 0.9 and accepted half the time is worse than one offered at 0.5 and accepted half the time, because the first one lies.Open
Columns, fields and actions
| Screen | Kind | What it is | Columns | Fields | Actions |
|---|---|---|---|---|---|
| Risk Queue | Workspace | What is open, what it is worth, and how quickly it is being worked — sorted by exposure rather than by score, because that sends people to the most expensive thing. | 0 | 0 | 0 |
| Risk Signals | Read-only list | A scored observation with its EVIDENCE attached. "Score 0.86" persuades nobody; the facts that produced it are what a person acts on. | 15 | 0 | 2 |
| Risk Cases | List | A person’s queue, with an outcome. The outcome is what closes the loop: confirmed fraud and dismissed noise are both training data, and a system that does not record which cannot improve. | 15 | 15 | 1 |
| Case Actions | List | What was done about it in the world, as opposed to in the case. | 8 | 6 | 0 |
| Risk Detectors | List | Each answers one question, is separately switchable and separately measured. A detector nobody can turn off is one that gets ignored wholesale the first time it is noisy. | 15 | 16 | 2 |
| Detector Performance | Read-only list | Precision — of the signals it raised, how many were real. The only number that decides whether a detector should keep running, and the reason outcomes are recorded at all. | 14 | 0 | 0 |
| Assistant Suggestions | Read-only list | Suggestion rather than detection, with the reasoning attached. "Because the last nine invoices from this supplier were coded 6410" is a reason a person can check; a confidence score is not. | 10 | 0 | 1 |
| Coding Patterns | List | A learned lookup rather than a model. For GL coding this beats a model on both accuracy and explicability, and a person can read it. | 11 | 7 | 0 |
| Narratives | List | The generated text and the edited one, both kept — the difference between them is the most useful feedback there is, and a narrative rewritten entirely was not worth generating. | 9 | 7 | 0 |
| Assistant Accuracy | Read-only list | Confidence is only useful if it is calibrated: a suggestion offered at 0.9 and accepted half the time is worse than one offered at 0.5 and accepted half the time, because the first one lies. | 10 | 0 | 0 |