A customer asks for competitor price matching. The policy covers own-website price adjustments but says nothing about competitors. The agent approves it for one customer and denies it for the next. Same request type, opposite outcomes. 23 complaints about unfair treatment in one month.
The agent is making unauthorized decisions in a policy gap. It lacks the authority to either approve or deny — both are policy-level decisions that require human judgment.
The Rule
When no policy section covers the customer’s request: escalate. Do not approve (creating unauthorized precedent). Do not deny (making unauthorized prohibition). Route to someone with policy-level authority.
After implementing this rule: unfair treatment complaints dropped from 23 to 0. Wait-time complaints: 2. A massive net improvement.
Policy Gap vs Ambiguous Policy
Policy gap: No policy section addresses the request at all (competitor matching, digital refunds, warranty extensions). The agent has zero guidance.
Ambiguous policy: A policy section exists but gives conflicting or unclear guidance. The agent has some guidance but cannot interpret it reliably.
Both require escalation. The agent should not interpret gaps as denial (“not mentioned = not allowed”) or as approval (“not mentioned = okay”).
The Data
1,000 requests tracked over a month:
| Category | Requests | Handling |
|---|---|---|
| Policy covered | 850 | 94% correct |
| Policy gap | 100 | 55 approved, 45 denied (unauthorized) |
| Ambiguous | 50 | Inconsistent interpretation |
100% of unauthorized decisions came from the 15% of requests without policy coverage. After adding an escalation rule: unauthorized decisions dropped to 3%.
Reducing Escalation Volume Over Time
The iterative approach satisfies both “maximize automation” and “never decide on gaps”:
- Escalate all gaps (constraint 2 met)
- Track recurring types — the same 5 gaps account for 80% of escalations
- Create official policy for common gaps with the policy team
- Update agent guidelines — now the agent can handle these authoritatively
- Monitor for new gaps as products and services evolve
This continuously reduces escalation volume while every decision remains policy-backed.
What Does Not Replace Escalation
Training on past decisions. Mimicking how humans resolved previous escalations doesn’t create policy. The agent still lacks authority — it just makes unauthorized decisions that look consistent.
Auto-approving based on precedent. Three human approvals don’t constitute policy. If the policy team hasn’t formalized the rule, agent auto-approval is still unauthorized.
Deny-by-default. Denial without policy basis can violate customer rights the policy was meant to protect. Both approval and denial in a gap are unauthorized.
Confidence scoring. The agent may be “confident” about a gap decision it has no authority to make. Confidence doesn’t create authority.
One-liner: When no policy covers a request, escalate — the agent lacks authority to either approve or deny, and the 52/48 random split on identical requests generates 23 unfair-treatment complaints that drop to zero after adding the escalation rule.