Customer Support Agent Designer

Full prompt

You are designing a customer support AI agent for {{company_name}}. The agent will handle {{primary_support_domain}} for {{target_customer_segment}}.

Define the agent's constraints:
- Core policies it must never violate (e.g., refund limits, data access rules, SLA commitments)
- Escalation triggers (e.g., "if sentiment < -0.7 OR request involves {{sensitive_topic}}, escalate to {{team_name}}")
- Approval gates (e.g., "refunds over ${{threshold}} require human sign-off")
- Failure modes to watch for (e.g., "agent offers discount without checking eligibility")

For each policy, specify:
1. The rule in plain language
2. How the agent detects when it applies
3. What the agent does if it applies
4. What happens if the agent gets it wrong

For escalations, define:
1. The trigger condition
2. The target team or queue
3. The context the agent must pass along
4. The SLA for human response

For evaluation, list:
1. Metrics to measure agent quality (e.g., "% of tickets resolved without escalation", "customer satisfaction on agent-handled tickets")
2. Red flags that signal a broken agent (e.g., "agent offers refunds to ineligible customers", "agent repeats same response 3+ times")
3. Weekly audit criteria

Output as a structured specification, not prose.

How to use

Run this prompt with your support team lead and your product/ops person in the room. Fill in the {{placeholders}} with your actual constraints, not your aspirational ones. The goal is to write down the rules the agent will actually follow—the ones you'd enforce manually if you had to. Once you have the spec, use it to: (1) brief your AI vendor or engineering team on non-negotiables, (2) build test cases that verify the agent respects each rule, (3) set up monitoring dashboards for the red flags you identified. Don't deploy the agent until you've run 100+ test tickets through the spec and caught at least one failure mode.

Customization tips

• Replace {{threshold}} with your actual refund limit, not your ideal limit. If your team would never approve a $2k refund without escalation, set the threshold at $500. The agent should be more conservative than your best human agent.
• Add a 'context passing' rule: when the agent escalates, it must include the customer's previous messages, the specific rule that triggered escalation, and what the agent already tried. Escalations without context waste your team's time.
• Build a weekly audit loop: pull 20 random agent-handled tickets, check them against your policies, and log any violations. After 2 weeks, you'll know which policies the agent actually understands and which ones it's faking.
• Define what 'resolved' means. If a customer says 'thanks, that helped' but the underlying issue isn't fixed, that's not resolved. Your evaluation metrics depend on this definition.
• Set a 'confidence threshold': if the agent is <70% confident it understands the request, it should escalate rather than guess. Measure how often this happens—if it's >30%, your policies are too complex or your agent needs better training data.