🤖 SupportPilot

Autonomous AI customer service across every email

SupportPilot receives customer emails, identifies the customer and order, answers questions, executes permitted actions, requests approval for refunds, cancellations, and address changes, replies inside the original thread, and escalates to a human when needed.

It is an autonomous customer-service platform for businesses that want every support email handled quickly without surrendering control over sensitive actions.

Giving a model access to an inbox is easy. Letting it safely represent a company is not. Real customer service requires identity resolution, durable threads, accurate business context, explicit permissions, approval boundaries, delivery handling, human escalation, and a record of every source and action.

SupportPilot packages those pieces into an installable TypeScript SDK, an operational inbox, and a hosted control plane.

🧩 Set up an autonomous support agent

A business connects its email transport, company knowledge, available tools, and explicit action policy:

export default defineCustomerService({
  transport: agentMail({
    inbox: "[email protected]",
  }),

  knowledge: [
    website("https://moltencoffee.shop"),
    document("./support-policy.md"),
  ],

  tools: [
    shopify.orders.read(),
    shopify.addresses.update(),
    shopify.refunds.prepare(),
  ],

  policy: {
    automatic: ["faq", "order_status"],
    approvalRequired: ["refund", "cancel_order", "change_address"],
    prohibited: ["invent_product_claim", "override_refund_policy"],
  },
});

The platform provisions the inbox workflow, normalizes incoming messages, maintains each customer conversation, gathers business context, runs the agent inside declared policy, and produces one of four clear outcomes:

Respond — answer safely using verified company knowledge
Act after approval — prepare a sensitive action and wait for a human decision
Escalate — hand the case to a person with a complete summary
Clarify — ask the customer for information required to proceed

🧠 Customer-service architecture

Customer email
     │
     ▼
AgentMail / Resend Transport Adapter
     │
     ▼
Webhook Verification, Dedupe & Message Normalization
     │
     ├──────────────► Durable Thread + Customer Identity Store
     │                                   │
     ▼                                   ▼
GPT-5.6 Support Planner ◄────── Knowledge & Business Context
     │
     ├────────► Policy Runtime ────────► Allow / Approve / Deny
     │                                      │
     ├────────► Business Tools              ▼
     │          Shopify • CRM • Billing   Human Approval Queue
     │                                      │
     └──────────────────────┬───────────────┘
                            ▼
                   Reply / Action Executor
                            │
                            ▼
                  In-thread Customer Reply
                            │
                            ▼
             Trace, Replay, Metrics & Human Handoff

🔐 Policy-controlled autonomy

The platform does not use an unexplained model confidence percentage to decide whether a refund is safe. A business expresses its operational boundaries directly.

Routine questions and order lookups can be resolved autonomously. Sensitive actions such as refunds, cancellations, or address changes are prepared by the agent and paused for approval. Prohibited actions are denied, and unsupported questions are escalated.

Every customer case records:

The matched policy rule
The sources used in the answer
The customer and order context supplied to the model
Every proposed and completed tool call
The approver identity and timestamp
The exact outbound reply
Delivery, response, and escalation events

That makes autonomous customer service testable, reviewable, and safe to improve.

🔄 Complete customer-service lifecycle

1. Receive

Verify inbound AgentMail or Resend webhooks
Deduplicate retried events
Normalize provider-specific payloads
Extract clean replies, attachments, and participants
Reconstruct the correct customer conversation

2. Understand

Resolve the customer's identity
Retrieve previous conversation memory
Load the relevant support policy and storefront knowledge
Fetch current order, billing, or account context
Classify intent, urgency, and the requested outcome

3. Decide

GPT-5.6 plans the answer and any required business tools
The policy runtime evaluates every requested action
The agent automatically responds, pauses for approval, asks a clarifying question, or escalates

4. Act

Run only explicitly permitted tools
Keep customer and commerce actions idempotent
Reply inside the original email thread
Handle provider delivery events and failures
Resume the case after an approval decision

5. Observe and improve

Display a structured trace for every customer interaction
Replay a case using a new policy or agent version
Compare proposed replies before deploying changes
Track automatic resolution, escalations, and approval outcomes
Export a complete handoff summary for a human agent

🧪 Customer-service simulator

Businesses should not test autonomous support changes on real customers.

The included simulator runs realistic multi-message conversations against the complete knowledge, tool, and policy configuration:

scenario: late-order-then-refund
messages:
  - "Where is order MC-1042?"
  - "It will arrive too late. Please refund it."
expected:
  first_message: automatic_response
  second_message: approval_required
prohibited:
  - refund_without_approval
  - unsupported_delivery_promise

Each scenario produces a pass/fail trace. Teams can run the suite locally or in CI before changing a prompt, model, policy, integration, or support workflow.

☕ Working demonstration: Molten Coffee

Molten Coffee is the first demonstration business.

Scenario 1 — Routine product question

A customer asks when their coffee was roasted. The agent cites the storefront's roasted-to-order information and answers automatically without inventing a fulfillment date.

Scenario 2 — Order lookup

A customer emails, "Where is my order?" The platform resolves the customer, retrieves the matching Shopify order, and replies in the same AgentMail thread with the current order status.

Scenario 3 — Refund requiring approval

The customer requests a refund. The agent prepares the permitted refund operation and a customer-ready reply, then pauses in the approval queue. After the merchant approves it, the tool runs exactly once and the customer receives the final response.

Scenario 4 — Address change requiring approval

A customer asks to change the delivery address after ordering. The agent verifies the order is still eligible, prepares the address update, and requests approval before making the change.

Scenario 5 — Unsupported product-safety question

A customer asks a question that is not answered by the product catalog. The agent refuses to invent a product or health claim and escalates the conversation with the missing information highlighted.

🔌 AgentMail and Resend provider adapters

AgentMail is the first provider because it supplies agent-owned inboxes, threads, replies, attachments, realtime events, and custom domains. The customer-service platform wraps it behind a transport contract:

interface CommunicationsTransport {
  verify(request: Request): Promise<VerifiedEvent>;
  normalize(event: VerifiedEvent): Promise<InboundMessage>;
  reply(
    threadId: string,
    message: OutboundMessage
  ): Promise<Delivery>;
  subscribe(handler: EventHandler): Promise<Subscription>;
}

A Resend adapter implements the same contract while the platform owns customer identity, threading, deduplication, ordering, and reply state. Future adapters can bring SMS, live chat, or WhatsApp conversations through the same customer-service policy and trace runtime.

🧱 Technical implementation

GPT-5.6: Customer-intent understanding, grounded answer planning, tool planning, clarification questions, and human-handoff summaries
Codex: TypeScript SDK, transport contracts, webhook ingestion, idempotency, policy runtime, Shopify tools, trace viewer, scenario simulator, tests, and deployable sandbox
AgentMail: Agent-owned inboxes, two-way email threads, attachments, webhooks, and realtime events
Resend adapter: An alternative email transport behind the same communications contract
Next.js + TypeScript: Customer-service inbox, approval queue, agent configuration, and trace explorer
PostgreSQL: Threads, customer identities, policy versions, approvals, and execution records
Shopify: Live order context and controlled support actions for the Molten Coffee demonstration

🎬 Three-minute demo plan

Run npx supportpilot init and show the generated support configuration.
Send a real customer email to the Molten Coffee support inbox.
Watch the agent identify the customer, retrieve the order, and reply in-thread.
Ask for a refund and show the policy runtime pause the action.
Approve the refund from the customer-service control plane.
Show the completed action, final customer reply, and replayable execution trace.
Run the same conversation in the simulator to demonstrate how businesses safely test policy changes.

🏆 OpenAI Build Week

Track: Developer Tools

SupportPilot is not a vague chatbot or an inbox wrapper. It is an autonomous customer-service product that understands every email, resolves routine work, performs approved business actions, and brings a human in exactly when needed.

Built With

agentmail
codex
next.js
openai-gpt-5.6
postgresql
resend
shopify
typescript