Azure AI Foundry Agent Confidence Scores: Enterprise Governance Built Into the Platform
Microsoft has shipped Agent Confidence Scores in Azure AI Foundry — and this is the enterprise governance feature that AI architects have been asking for since agent deployments started going into production. Every agent output now gets a 0-1 reliability rating. Below 95 percent, the action auto-routes to a human reviewer before it executes. This is not a wrapper you build yourself — it is native to the Foundry runtime and requires configuration, not code.
What the Confidence Score Actually Is
The confidence score is a composite signal, not just the model's self-reported certainty. Microsoft uses two inputs:
- Model self-confidence: The model's internal probability estimate for the output it produced — derived from the token probability distribution, not a separate evaluation call
- Independent evaluator score: A separate lightweight model that reviews the primary model's output against the task specification and rates alignment — this is the more reliable of the two signals
The composite score is weighted and normalised to a 0-1 scale. A score of 0.95 means the system has high confidence that the output is correct, well-aligned with the task, and safe to execute. Below 0.95, the system flags the output for human review before any action is taken.
Why Microsoft Chose 95 Percent as the Default Threshold
The 95 percent default is calibrated for general enterprise workflows — the kind where a mistake is costly but reversible. At that threshold, approximately 5 percent of agent outputs route to human review. Microsoft's internal testing found that threshold captures the majority of genuinely risky outputs without overwhelming human reviewers with false positives.
The threshold is not fixed. Microsoft provides guidance on when to raise it:
- 98 percent threshold for actions that are difficult to reverse — sending external communications, writing to production databases, initiating financial transactions
- 99 percent threshold for truly irreversible actions — deleting records without backup, decommissioning infrastructure, regulatory submissions
How to Wire Confidence Scores Into Your Existing Foundry Agent
This is a configuration change, not a code rewrite. In your Foundry agent deployment manifest, add the confidence score policy block with your threshold and the routing destination for flagged outputs — typically a Teams channel or a ticketing system queue that your operations team monitors.
The human review interface in Foundry shows the reviewer the full agent context: the original task, the proposed action, the confidence score, and the reasoning chain. Reviewers approve or reject. Approved outputs execute; rejected outputs return to the agent with reviewer feedback for a retry.
Three Honest Catches
- Calibration is not perfect: The confidence score is a probability estimate, not a guarantee. A score of 0.96 does not mean the output is correct — it means the system believes it is. Novel task types the model was not trained on may have poorly calibrated confidence scores, appearing higher confidence than the actual correctness rate warrants.
- High confidence does not equal correct: Overconfident wrong answers are a known failure mode in large language models. The confidence score reduces risk but does not eliminate it. For highest-stakes actions, threshold tuning alone is insufficient — you also need domain-specific evaluation criteria baked into the evaluator prompt.
- Human reviewer queue SLA: Routing to human review only works if humans actually review in a timely manner. A 95 percent threshold that generates 20 reviews per hour requires staffing the review queue. This is an operational commitment, not just a technical configuration.
3 Moves to Ship This Week
- Enable confidence scores on your highest-stakes Foundry agent: Start with the agent that takes the most consequential actions. Set the threshold at 98 percent initially and monitor what percentage of outputs route to review — this gives you calibration data before tuning.
- Define your irreversible action taxonomy: Make a list of every action your agents take and classify each as reversible, difficult-to-reverse, or irreversible. Map different threshold values to each category. This is the policy document your confidence score configuration should reflect.
- Build the reviewer experience before you enable routing: The technical integration is straightforward. The harder part is building the review interface and workflow that makes human reviewers efficient. Define what information they need, where they receive alerts, and how they record their decisions before the first flagged output arrives.
Key Takeaways
- Azure AI Foundry now ships a native 0-1 confidence score for every agent output — below 95 percent auto-routes to human review before execution
- The score combines model self-confidence with an independent evaluator — the composite is more reliable than either signal alone
- The 95 percent default covers general enterprise workflows; raise to 98-99 percent for irreversible actions
- This is a configuration change to your existing Foundry agent, not a code rewrite — but it requires operational commitment to staff the human review queue
- Confidence scores reduce risk but do not eliminate it — overconfident wrong answers remain a known failure mode that requires domain-specific evaluator tuning for the highest-stakes workloads
