Key facts
| Endpoint | POST /v1/chat/completions with tools for account lookups and bounded actions |
| Models | 30+ models; cheap aliases for triage and stronger ones for complex cases |
| Identity | Verify before exposing account tools; never pass unverified identifiers into prompts |
| Streaming | Live for responsive replies in chat channels |
| Pricing model | Flat monthly self-serve plans with unlimited fair-use usage |
| Free tier | Free plan with 2 free AI models, no card required |
| Governance | Scoped keys, audit logs and sampled QA review |
| Roadmap | Moderation and audio endpoints are coming soon |
TL;DR
- Triage first: intent, sentiment and entitlement decide the path.
- Identity and permissions are code-level checks, never prompt instructions.
- Define action tiers so refunds and cancellations cannot happen implicitly.
- Escalate with a structured summary, not a raw transcript.
- Review a sample of conversations weekly; metrics alone miss tone failures.
How it works, step by step
- Map the top intents and mark which are answerable, which need account data and which need a human.
- Implement identity verification before any account-scoped tool becomes available.
- Add read tools for orders, subscriptions and case history, and write tools for bounded actions.
- Define an action policy: answer, suggest, execute with approval, or never.
- Run the agent loop with streaming replies and a full trace of tools and decisions.
- Escalate with a compact summary when confidence, entitlement or sentiment fails, and sample conversations for QA.
Original data
Try it yourself
Architecture of a support agent
Six parts, each with clear ownership:
- Channel adapter: chat widget, email or messaging platform delivering normalized events.
- Identity service: verifies the customer before account tools are exposed.
- Triage step: intent, urgency and sentiment classification on a cheap alias.
- Tool layer: read tools for context and write tools with explicit policies.
- Agent loop: policy-grounded answers and bounded actions with streaming replies.
- Handoff: structured summary, customer state and recommended next step for a human.
Action policy: the safety core
Support agents touch money and entitlements, so define tiers explicitly:
- Answer: policy and product questions answered from retrieved or provided content.
- Suggest: the agent proposes an action but a human or the customer confirms.
- Execute with approval: allowed actions inside limits — for example refunds below a threshold — with approval logic in code.
- Never: account closure, credit changes, legal commitments and anything irreversible.
The model selects intent; your policy engine decides permission. Never encode entitlement rules only in the prompt, because prompt injection and model drift both break them.
Implementation and evaluation
Operational practices that keep quality visible:
- Stream replies to reduce perceived wait, but persist the final answer for QA.
- Log model, tools, retrieved sources, approvals and handoff reason per conversation.
- Sample conversations for human QA against a rubric: accuracy, tone, policy compliance.
- Track containment only on intents you intend to contain, and resolution accuracy through follow-up signals.
Core metrics: containment, resolution accuracy, escalation precision, hallucinated commitments (target zero), and time to resolution. Review clusters of failures monthly to decide whether the fix is content, policy or model routing.
Limitations
Support agents amplify whatever your policies get wrong:
- Confident wrong answers damage trust faster than slow ones; grounding and refusal paths matter more than fluency.
- Empathy and de-escalation are hard to evaluate automatically and easy to get wrong; keep humans in the loop for distressed customers.
- Moderation endpoints are coming soon, so profanity, abuse and safety filtering need your own rules today.
- Identity verification adds friction; design one strong path rather than several weak ones.
- PII handling must comply with your own policies and the regions you serve — scoped keys and region pinning help, but data minimization starts in your pipeline.
Honest comparison
| Capability | Support agent on Plugsky | Macro-and-menu bot | Human-only support |
|---|---|---|---|
| Understanding | Intent, sentiment and context | Keyword matching | Full judgment |
| Account context | Read tools behind identity checks | None | Manual lookup |
| Actions | Tiered policy with approvals | None or fixed scripts | Policy-bound humans |
| Escalation | Structured summary with state | Transfer without context | Not applicable |
| Cost shape | Flat monthly self-serve, unlimited fair use | Per-seat bot platform | Headcount |
Frequently asked questions
What should a support agent be allowed to do?
Answer questions, fetch context and take bounded actions such as resending a receipt. Refunds, cancellations and credit changes should be tiered with approval rules enforced in code.
How do I prevent wrong account access?
Verify identity before exposing account tools, keep identifiers out of the model context unless required, and log every account access for audit.
Which model handles support chats?
Triage runs well on plugsky-micro or plugsky-lite, while complex cases and multi-step troubleshooting benefit from a stronger alias. Route by intent.
How do I evaluate support quality?
Track containment on intended intents, resolution accuracy through follow-ups, escalation precision and hallucinated commitments. Add weekly human QA sampling for tone.
When should the bot hand off?
On repeated failure, explicit request, distressed sentiment, or any intent outside its policy. Hand off with a summary and customer state.
Is there moderation for abusive input?
Not yet — moderation endpoints are coming soon. Apply your own filters and escalation rules for abusive interactions in the meantime.
Can I pilot support automation for free?
Yes. plugsky-micro and plugsky-lite are free with no card, and the 14-day full-access trial lets you test stronger models on real ticket samples.