The most important governance choice for any finance AI is how much autonomy it gets. A practical ladder has four rungs:
| Level | What AI does | Example |
|---|---|---|
| 1. Inform | Analyzes and highlights | Flags unusual journal entries |
| 2. Recommend | Proposes an action; people decide and execute | Suggests GL codes; drafts variance commentary |
| 3. Act with approval | Prepares the action; a person approves it | Prepares a payment run for sign-off |
| 4. Act alone (within limits) | Executes under limits; people monitor | Auto-applies cash under a matching-confidence threshold |
How to choose the level
- Consequence: what does an error cost, and can it be reversed?
- Volume: high volume makes human review expensive, which pushes toward level 4 with sampling.
- Evidence: how good is the measured accuracy on your data?
- Regulation and reputation: anything affecting customers, credit, tax or disclosures needs more human control.
Where finance leaders are today
Esker's 2026 survey found 32% of finance leaders use AI to recommend actions that teams execute manually, and 29% let AI act with human approval. CFOs are much more cautious than their teams on strategic decisions: 78% of CFOs say AI should not execute decisions around setting revenue targets, compared with 11% of other finance leaders.
Move up the ladder with evidence
Start one level lower than you think, measure accuracy and overrides, and promote a use case one level at a time when the evidence supports it. Write the level, limits and owner into a policy, just like signing authority.