A Member-Facing Agentic Assistant for Club Management, Delivered as an API

Industry: Hospitality and Membership Services. Offering: Agent and Automation Development, Conversational AI.

About Customer

Customer is a United States club operator running an established club-management platform with a relational data tier, a member identity provider, and member-facing web and mobile applications. Members use those applications for bookings, class reservations, billing, and account management.

Customer wanted members to ask for what they need in their own words and finish the task in the same conversation, with the existing platform kept as the system of record.

Business Objectives

1. Let members ask in natural language. Membership status, renewal dates, visit history, booking availability, class schedules, outstanding balances, facility hours, and club policy, answered conversationally in the apps members already use.

2. Complete the action in the same exchange. Check whether a court is free on Saturday morning and book it, move a session, cancel a class, or pay dues without leaving the conversation.

3. Support chat and voice from one implementation. Customer wanted voice and text to use the same reasoning path, with no second build.

4. Extend the existing platform instead of replacing it. The club-management platform holds the systems of record. No member-facing app was replaced in Phase 1.

5. Enforce per-member data isolation structurally. A member can only see or change their own information, enforced by the architecture instead of by instructions to the model.

6. Guarantee transactional correctness. An assistant that books and takes payment has to behave correctly when a request is retried or the network fails. Customer wanted idempotency and explicit confirmation on every write.

Solution Delivered

1. AspireNXT built the assistant as a single Agent API. The existing web and mobile apps send a member request (text or a short audio clip) and render the structured response themselves.

2. Amazon Bedrock AgentCore provides the agent runtime, with managed runtime, memory, gateway, identity, and observability components.

3. Data access runs through governed Model Context Protocol tools. The agent reads and writes club-management data through a defined set of MCP tools against the existing platform. It never issues direct, unbounded queries. Each tool has a fixed contract and a fixed scope.

4. Transactional actions are gated by confirmation and idempotency. For bookings and payments the agent checks availability or amount, presents the details, takes explicit member confirmation, then executes an idempotent write through the booking tool or the existing payment gateway. A retried request cannot produce a second booking or a second charge.

5. Voice runs through the same pipeline as chat. Spoken requests are transcribed with streaming speech-to-text and then follow the same intent, retrieval, and action path. Optional text-to-speech returns a spoken reply, configurable per surface.

6. Every interaction is observable. Prompt, retrieval, tool calls, and transaction outcome are captured to traces and Amazon CloudWatch with sensitive data redacted, so any member interaction can be reconstructed.

7. Some tasks stay with existing systems. Refunds, chargeback disputes, and proration adjustments stay in the existing systems, and the assistant hands them off.

8. Delivered into the customer's own AWS account as infrastructure as code.

Anthropic Claude on Amazon Bedrock

Claude interprets each request, chooses the tools, asks for anything missing, and completes the transaction.

1. Intent and entity extraction from free-form input. Claude turns a request like "move my Saturday session to Sunday morning" into a member, a booking, a target slot, and an action, from typed text or transcribed speech.

2. Agentic tool selection over the MCP layer. Claude decides which governed tools a request needs and in what order: check membership, look up availability, read the balance, call the booking tool. The tools are fixed and scoped. Claude picks the order for each request.

3. Multi-turn parameter gathering. When a request is missing a detail, such as a court booking with no time or a class with no date, Claude asks for it and keeps the earlier context in AgentCore memory.

4. Confirmation before every write. For bookings and payments Claude states back what it is about to do, in the member's own words, and waits for explicit confirmation. Claude writes the confirmation, and the tool layer enforces idempotency.

5. Grounded answers on club policy. Questions about guest policy, facility hours, or class rules are answered from an Amazon Bedrock Knowledge Base of approved club content, so answers reflect current club policy.

6. Routed to the model the task needs. Turning a short phrase into an intent is lighter work than running a multi-turn booking across several tools. Each request goes to the Claude model suited to that work, which keeps replies fast.

7. Isolated and bounded. Claude runs inside an authenticated member context, so tool calls cannot reach another member's data. Amazon Bedrock Guardrails refuse out-of-scope requests and redact sensitive data from responses and logs.

Services Used

• Anthropic Claude on Amazon Bedrock

• Amazon Bedrock AgentCore

• Amazon Bedrock Knowledge Bases and Bedrock Guardrails

• Model Context Protocol (MCP) tool layer

• Streaming speech-to-text and text-to-speech

• Amazon CloudWatch

• Langfuse for tracing

What Was Delivered

1. Members self-serving in natural language across membership, bookings, classes, billing, and facility questions, in the apps they already use.

2. Bookings and payments completed in the conversation, each behind an availability or amount check and explicit member confirmation.

3. One agent for chat and voice.

4. No platform replacement. The existing club-management platform, identity provider, and member applications stayed in place.

5. Structural data isolation, with per-member scoping enforced at the tool layer on every request.

6. Transactional safety, with idempotency and confirmation on every write.

7. Full traceability, with every request, retrieval, tool call, and transaction outcome captured and sensitive data redacted.

Considering a member-facing agent over an existing platform? Talk to the AspireNXT AI team.