What are confidence thresholds?
Every time an AI agent proposes an action — drafting content, categorizing feedback, sending a response — it assigns a confidence score between 0 and 100. The confidence threshold is the minimum score at which an agent in autonomous mode can execute an action without human approval. Proposals that fall below the threshold are routed to the agent inbox for review.
Where to set the threshold
- Navigate to Settings > Agents.
- Select the agent you want to configure.
- Find the Confidence threshold slider or input field in the agent's settings panel.
- Set the threshold to your desired value and save.
Understanding the scale
- 90–100 (Very high) — Only the most certain actions proceed automatically. Nearly everything goes to the inbox. Best for agents handling sensitive or customer-facing content where errors have real consequences.
- 70–89 (High) — A conservative but practical setting. Routine, straightforward actions execute automatically while anything ambiguous gets human review. Good starting point for most teams.
- 50–69 (Moderate) — The agent handles the majority of tasks on its own. Lower-confidence actions still get reviewed. Suitable for well-established workflows with consistent data.
- Below 50 (Low) — The agent has wide autonomy. Use this only for low-risk tasks where occasional mistakes are acceptable and easy to correct.
Setting thresholds per workflow
In addition to the global agent threshold, you can set a threshold for individual workflows and automations. This lets you be strict for public-facing content (high threshold) while allowing more autonomy for internal categorization tasks (lower threshold).
To set a workflow-specific threshold, open the workflow or automation settings and look for the Confidence threshold option. If no workflow-specific threshold is set, the agent's global threshold applies.
How to calibrate your threshold
- Start high (80+). Review everything the agent proposes for the first week or two.
- Track your approval rate. If you are approving 90% or more of proposals without changes, the threshold is likely too high.
- Lower gradually. Drop the threshold by 10 points and observe. Continue lowering until you reach a balance where the agent handles routine work and you only review edge cases.
- Monitor rejections. A spike in rejections after lowering the threshold means you went too far. Raise it back up.
Threshold and agent mode interaction
The confidence threshold only applies when an agent is in autonomous mode. In supervised mode, every proposal goes to the inbox regardless of its confidence score. In assistant mode, the agent only acts when you explicitly ask, so the threshold is not relevant.
Best practices
- Different thresholds for different risk levels. A feedback-categorization agent can safely run at 60. A public-announcement agent should run at 85 or higher.
- Review your rejected items. Patterns in rejections reveal where the agent needs better guidance, not just a higher threshold.
- Revisit periodically. As you refine agent guidance and the agent improves, you can safely lower thresholds over time.