AI Personas

A persona is the operating profile for one AI agent. It combines a system prompt, tone rules, a catalog of tools it may call, and the knowledge base scope it may retrieve from. Beamdesk evaluates every candidate reply against that profile before the answer is sent or queued for review.

Treat personas as production roles, not just writing styles. A support persona may answer setup questions, a billing persona may inspect invoices, and a technical persona may collect logs before escalating to engineering.

The system prompt describes the role, allowed topics, refusal policy, and escalation threshold. Tone settings keep language consistent across channels. The tool catalog grants explicit access to HTTP actions, integrations, or internal lookup functions. Knowledge scope limits retrieval to selected collections, article tags, or source groups.

Persona: Support
System prompt:
  You resolve setup, pricing, and account questions for Beamdesk customers.
  Use approved knowledge before answering.
  Escalate legal, security review, refund exception, and outage questions.

Tone:
  Direct, concise, calm. No invented timelines.

Knowledge scope:
  collections: ["public-docs", "billing-faq", "onboarding"]
  excluded_tags: ["internal-only", "draft"]

Tools:
  get_customer
  list_orders
  create_handoff

Multi-persona workflows let Beamdesk route the same queue to specialized agents without forcing customers through separate entry points. The classifier picks the first owner, and automations can move the ticket again if a later message changes the intent.

Use /settings/ai/playground to compare personas before they reach production. Paste real customer messages, lock the knowledge snapshot, and run two persona versions against the same prompt set.

A/B tests should measure resolution quality, escalation accuracy, latency, and token cost. Keep only one variable different between versions: prompt, tone, tool access, or knowledge scope.

Experiment: billing-refund-tone
Traffic: 20% of billing-tagged tickets
Variant A: Billing persona v3
Variant B: Billing persona v4
Primary metric: human override rate
Guardrails:
  max_auto_refund_amount: 0
  require_handoff_for_policy_exception: true

Persona analytics show whether a specialist is actually improving support operations. Review resolution rate, confidence distribution, cost per resolved ticket, deflection by channel, handoff rate, and the percentage of answers edited by humans.

Low confidence with high escalation quality usually means the persona is conservative. High confidence with many human edits means the persona needs tighter knowledge scope, stricter refusal language, or a narrower tool catalog.

Personas call custom HTTP tools only when the tool is listed in the persona catalog and the current ticket satisfies the tool policy. Tool results are attached to the ticket timeline so reviewers can see which customer data influenced the answer.

{
  "name": "lookup_subscription",
  "method": "GET",
  "url": "https://api.example.com/customers/{{customer.email}}/subscription",
  "headers": {
    "Authorization": "Bearer {{secrets.billing_api_key}}"
  },
  "allowed_personas": ["Billing"],
  "timeout_ms": 5000
}

Escalation paths decide what happens when a persona should not answer. Beamdesk can hand off to another persona, assign a human team, notify Slack, create a Linear or Jira issue, or leave a reviewed draft for approval.

Use persona-to-persona escalation for intent changes, such as a support ticket becoming a billing question. Use human escalation for risk, uncertainty, compliance, angry customers, or tool failures.