Open finance skill
Categorization history health
Is transaction history consistent enough to support reliable automation?
Measure vendor-level categorization consistency, overrides, ambiguity, and time drift before proposing automation rules.
Inputs
- ✓Categorized transactions
- ✓Vendor normalization
- ✓Override history
Outputs
- ✓Stable candidates
- ✓Conflicting history
- ✓Manual-policy queue
Checks performed
- ✓History coverage
- ✓Category consistency
- ✓Time drift
Guided synthetic example
See the check from input to decision
A vendor appears frequently, but its historical categories are inconsistent.
This is the operating model we use at Level. The percentages are estimated shares of the work in a typical review. Actual effort varies with data quality, complexity, and the issue found.
Start with the source records
About 15% of work| Vendor | Transactions | Most common category | Consistency |
|---|---|---|---|
| Example Hardware | 62 | Materials | 61% |
| Sample Telecom | 38 | Telephone | 97% |
Run the deterministic check
About 20% of workMeasure normalized-vendor history, category concentration, overrides, and time drift before proposing automation.
Surface the flagged result
About 10% of workNeeds review
Example Hardware is only 61% consistent and should not receive a broad automatic category rule.
Use recorded finance judgment to analyze possible reasons
About 25% of workAt Level, we record the review logic, known explanations, and questions our finance professionals use in this situation. AI applies that documented human judgment to organize the most plausible reasons, without pretending it knows which reason is true.
Possible reasons to investigate
The vendor may legitimately span materials, tools, repairs, and office purchases, or the historical books may contain miscoding.
Make the final human judgment
About 20% of workA Level finance professional validates the source evidence, challenges the AI-assisted analysis, and decides which explanation is supported.
Make the operating decision
About 10% of workKeep the vendor manual or add a narrow secondary condition after finance policy review.
Use this system
Start with the example, then inspect the structure
- Download the invented input and expected output.
- Replace one field at a time with a safe test value.
- Compare the result with the expected structure.
- Have a finance professional review every exception before acting.
MCP preview
run_synthetic_example {"slug":"categorization-history-health"}Human review is part of the system
This check organizes evidence and surfaces exceptions. A person with finance domain expertise still needs to interpret the result, validate the source records, and discuss the operating decision.
Open source does not mean open client data. We publish reusable finance tools, checks, templates, and playbooks. Client records, workpapers, credentials, communications, and private configurations remain confidential.