Back to registryConditional
agt-003
Summarizes long customer support threads into a structured handoff brief for agents.
- Owner
- Dana Whitfield
- Customer Support
- Model
- GPT-5.5
- Data Access
- Customer/Regulated
- Autonomy
- Read-only
- Value Tier
- High
- Hours Saved / mo
- 120
- Connected Tools
- Zendesk, Slack
Launch Readiness
79/ 100
Rubric breakdown
Seven criteria, each scored 0–4 and weighted. Data Safety and Tool Permission Risk carry hard safety gates.
- Business Value4/4weight ×1.0
- Use Case Clarity4/4weight ×1.0
- Data Safety4/4weight ×2.0
- Tool Permission Risk4/4weight ×2.0
- Evaluation Quality2/4weight ×1.5
- Human Review Coverage1/4weight ×1.5
- Auditability3/4weight ×1.0
Why this verdict
- Evaluation Quality is below target (2/4).
- Human Review Coverage is critically low (1/4).
Evaluation
89%
Accuracy
higher is better
5%
Hallucination
lower is better
12%
Escalation
routed to a human
82%
Satisfaction
higher is better
1.8s
Latency
per task
$0.03
Cost / task
estimated
Failure modes
- May omit nuance in emotionally charged threads.
- Limited evaluation coverage on multi-language tickets.
- Human review currently spot-checks only a sample of summaries.
Audit log
| Timestamp (UTC) | Actor | Event |
|---|---|---|
| Feb 5, 2026, 8:50 AM | Dana Whitfield | Agent registered with read-only Zendesk scope. |
| Feb 26, 2026, 2:10 PM | Privacy Review | Customer-data handling reviewed; PII redaction confirmed in summaries. |
| Apr 2, 2026, 9:35 AM | Eval Pipeline | Evaluation flagged limited coverage on non-English threads. |