Fledgegate

Support teams may want some AI capabilities to remain assist-only while lower-risk, well-tested capabilities become customer-facing. Fledgegate proposes a separate graduation record for each defined skill, such as answering a policy question or preparing a bounded account action. The supplied research found feature-control, observability and response-correction products as structural analogues but no direct per-skill autonomy gate. It also confirms only one of three referenced interfaces; the other connections remain pre-build gates. A limited competitor search does not prove an empty market.

Human override frequency is not accuracy: staff can correct style, use shortcuts, disagree or make mistakes. Edit distance has little meaning without semantic labels. Model-stated confidence is not comparable across models or skills unless calibrated against representative, reviewer-adjudicated outcomes. Production traffic drifts, and observed outcomes can arrive late or be confounded by policy changes. Refund eligibility, plan changes and identity recovery carry different authority, reversibility and customer harm. A global score or cross-tenant benchmark would erase those differences and can leak confidential policy and behavior. Human agents must never be ranked from override data.

Skill definition, allowed action, policy version, risk tier, assist-mode draft, human action, override reason, adjudicated label, calibration sample, metric contract, threshold proposal, risk-owner approval, promoted state, customer interaction, outcome observation, appeal, drift candidate, preapproved stop condition, automatic fail-safe demotion, owner acknowledgment, remediation, retest and repromotion remain separate. Promotion always requires accountable approval. Automatic demotion is allowed only as a reversible fail-safe under a versioned preapproved policy, never from an opaque model judgment.

A support AI operations, quality or risk leader responsible for deciding which agent capabilities may face customers and under what evidence, limits and rollback policy.