Trust Score Factor 5: The Art of Escalation Judgment
An AI agent that escalates everything is useless — it is a notification machine with extra steps. An agent that never escalates is dangerous — it will eventually encounter a situation it cannot handle and fail silently. The sweet spot is an agent that knows the difference between what it can handle and what it should not try. That judgment is what escalation scoring measures, and it accounts for 15% of the trust score.
The Two Failure Modes
Escalation judgment has two failure modes, and Shulam penalizes both:
Over-escalation (too eager)
The agent escalates tasks it could have handled independently. This creates unnecessary workload for human operators, defeats the purpose of automation, and signals that the agent lacks confidence in its own capabilities. Over-escalating agents typically have escalation rates above 25% — meaning one in four tasks gets pushed to a human. The network median for Act-level agents is 4.2%.
Under-escalation (too quiet)
The agent handles tasks it should have escalated. This is the more dangerous failure mode because the consequences are often invisible until they compound. An agent that processes an ambiguous compliance case instead of escalating it might produce a technically valid but contextually wrong decision. Under-escalation is detected through post-hoc review: human auditors sample completed tasks and identify cases where escalation would have been appropriate.
How Escalation Quality Is Calibrated
Measuring escalation judgment requires a ground truth: what should the agent have escalated? Shulam establishes this through a three-layer calibration process:
- Operator-defined rules.Each operator configures explicit escalation triggers: transaction amounts above a threshold, specific counterparty types, tasks involving PII, first-time scenarios. These are the baseline — if the rules say "escalate," the agent must escalate. Non-negotiable.
- Peer comparison.For ambiguous cases (no explicit rule applies), Shulam compares the agent's decision against what other agents with similar capability scopes did in analogous situations. If 80% of comparable agents escalated a similar case, an agent that did not escalate receives a negative signal. This is not majority-rules voting — it is a calibration input weighted at 30% of the ambiguous-case score.
- Human audit sampling. Every week, a random sample of 5-10% of completed tasks (non-escalated) are reviewed by human auditors. Auditors mark whether escalation would have been appropriate. These audit results feed back into the scoring model and refine the peer comparison baselines. The audit catch rate — how often auditors find a missed escalation — is the single strongest predictor of escalation factor scores.
What Good Escalation Looks Like
The best-performing agents on the escalation factor share three characteristics:
- They escalate with context.A bare escalation ("I cannot handle this") is worse than no escalation at all. High-scoring agents attach a summary of what they know, what they tried, why they are uncertain, and a recommended action. The human reviewer can make a decision in 30 seconds instead of 5 minutes.
- They escalate early. An agent that recognizes uncertainty at step 2 of a 10-step process and escalates immediately scores higher than one that plows through 8 more steps and escalates at step 10. Early escalation limits blast radius.
- They learn from escalation outcomes. When a human resolves an escalated task, the resolution becomes a training signal. Agents that incorporate these signals — handling similar cases independently next time — show improving escalation rates over time. The scoring model rewards this trajectory: a declining escalation rate (from 15% to 5% over 60 days) boosts the factor score even if the current rate is still above median.
The Escalation Rate Sweet Spot
There is no universal "correct" escalation rate — it depends on the agent's domain, risk profile, and authority level. But network data reveals clear patterns:
An Act-level agent with a 20% escalation rate is almost certainly over-escalating. An Authority-level agent with a 0% escalation rate over 30 days either has a very narrow scope or is under-escalating. Both patterns trigger a review signal in the scoring model.
Why This Factor Matters More Than You Think
Escalation judgment is the factor that most directly measures an agent's self-awareness. Task accuracy measures whether the agent can do the work. Compliance measures whether it follows the rules. Escalation measures whether it knows the limits of its own competence. That metacognitive ability — knowing what you do not know — is what separates agents that operators trust from agents that operators babysit.
When enterprises evaluate AI governance platforms, they consistently rank "knowing when to involve a human" as a top-3 requirement. The escalation factor is Shulam's quantitative answer to that requirement: a number between 0 and 100 that tells you exactly how well your agent makes that judgment call.
Explore how escalation interacts with the other six factors in the Trust Score documentation.
See How Your Agents Escalate
Model your agent's escalation rate against network benchmarks and see the impact on trust score.
Try the Calculator