Key facts
| Agent stack | Chat completions, function calling and embeddings are live |
| Use cases | Order status support, returns and refunds, recommendations, catalogue enrichment |
| Policy tools | Expose returns and refund rules as tools with enforced limits |
| Payment data | Keep card data in your payment provider and out of prompts |
| Models | 30+ models; small aliases for classification, stronger ones for conversations |
| Deployment | Plugsky cloud, your VPC, on-prem and air-gapped options |
| Pricing model | Flat monthly self-serve plans with unlimited fair-use usage |
| Roadmap | Batch and assistants endpoints are coming soon |
TL;DR
- Resolve routine order questions end to end with read-only tools.
- Enforce refund limits in code, not in the prompt.
- Escalate exceptions, disputes and fraud signals to a human.
- Keep payment credentials out of every prompt and log.
- Measure deflection, escalation rate and customer satisfaction together.
How it works, step by step
- Pick one queue such as order status or returns to start with.
- Expose read-only order tools and a policy tool for returns and refund rules.
- Set hard refund and discount limits in code, with human approval above them.
- Ground product answers in catalogue data and citations rather than model memory.
- Keep card data out of prompts and send order references instead.
- Log tool calls and model versions, then measure deflection and escalation quality.
Try it yourself
Ecommerce workflows suited to agents
Support and catalogue work repeat endlessly, which suits bounded agents:
- Order support: answer status, delivery and address questions from order data.
- Returns and refunds: apply the policy, process routine cases and escalate exceptions.
- Product guidance: recommend based on compatibility and stock using catalogue retrieval.
- Catalogue enrichment: fill attributes and write descriptions for review before publishing.
The agent handles volume; people handle disputes, exceptions and brand decisions.
Architecture and hard limits
Put the rules in code where they cannot be argued with:
- Read-only tools: order lookup, tracking, stock and policy retrieval.
- Action tools with caps: refunds and discounts only within configured limits.
- Approval queue: anything above the limit or outside policy goes to a human.
- Retriever: embeddings over catalogue, policies and help content for grounded answers.
- Logger: records tool calls, inputs, model version and outcome for each conversation.
Payment card data never enters a prompt; use tokens and order references.
Peak readiness and customer trust
Peak trading exposes every weakness, so test at volume before the season. Rate-limit per channel, queue non-urgent work, and keep an escalation path that a human can join mid-conversation. Be transparent about automation: a clear handoff builds more trust than a confident wrong answer. Use region selection to match customer privacy expectations and keep audit logs for dispute resolution. Where a marketplace or enterprise customer requires isolation, deploy in a VPC or on-prem using the same compatible API.
Pilot and measurement
Prototype on plugsky-micro or plugsky-lite on one queue, then use the 14-day full-access trial to test a stronger model on complex conversations. Measure containment rate, escalation rate, refund accuracy, handling time and satisfaction, and review a sample of escalated cases weekly. Self-serve plans are flat monthly with fair-use usage, so seasonal peaks and catalogue sweeps stay predictable. Roll out queue by queue with versioned prompts, and re-test whenever policies or catalogue structures change.
Honest comparison
| Capability | Plugsky | Typical agent framework | Building in-house |
|---|---|---|---|
| Tool calling | Live function calling over order, stock and policy APIs | Framework-specific adapters | You build and maintain |
| Policy enforcement | Caps and approval gates enforced in application code | Varies | You design controls |
| Deployment | Cloud, VPC, on-prem and air-gapped | Usually cloud-only | You operate GPUs and serving |
| Pricing | Flat monthly self-serve, fair-use usage | Varies by framework and model | GPU plus operations cost |
| Model choice | 30+ models behind one API | Varies by provider | You host every model |
Frequently asked questions
Can the agent issue refunds?
Yes, within hard limits you set in code. Above the limit or outside policy, route the case to a human approver with the evidence attached.
Which endpoints do ecommerce agents use?
Chat completions, function calling, JSON mode and embeddings are live. Batch and assistants endpoints are coming soon.
Is payment data sent to the model?
No. Keep card data in your payment provider, send order references and tokens, and exclude credentials from logs and prompts.
Is there a free plan?
Yes. The free plan includes plugsky-micro and plugsky-lite with no credit card, enough to prototype one support queue.
How does pricing work?
Self-serve plans are flat monthly with unlimited fair-use usage, which suits seasonal volume. See the live pricing page for current plans.
How do we handle peak season?
Rate-limit per channel, queue non-urgent work, keep a human escalation path and load-test the agent before the peak.
How should we measure it?
Track containment, escalation rate, refund accuracy, handling time and satisfaction together, and review escalated cases weekly.