Architecture

Reference architecture for the Consumer AI-Agent Channel.

The orchestration hub is the control plane. AI vendors are the channel endpoints. Enterprise APIs and systems of record are the transaction plane.

Architecture at a glance

Consumer AI Platforms

ChatGPT, Claude, Gemini, Copilot, future agents, mobile and voice assistants.

AI-Agent Orchestration Hub

MCP servers, vendor adapters, identity broker, consent, policy engine, RAG, tool registry, workflow router, audit, observability.

Enterprise Backend

CRM, OMS, payments, loyalty, billing, fulfillment, claims, case management, knowledge bases, data lake.

What is inside the orchestration hub?

Component	Purpose	Typical implementation
Channel / vendor adapter	Handles vendor-specific connection models, auth, payloads, tool schemas, and response formatting.	OpenAI Apps SDK / MCP adapter, Claude remote MCP, Gemini function-calling, Copilot connector.
MCP server layer	Standardized tool and resource exposure for AI clients that support MCP.	Remote MCP servers exposing tools like `get_order_status`, `start_refund`, `create_case`.
API gateway	Secure ingress, rate limiting, request validation, token validation, traffic logging, routing.	Kong, Apigee, Azure API Management, AWS API Gateway, custom gateway.
Identity broker	Authenticates the customer and binds the AI-agent session to an enterprise identity.	OIDC / OAuth, passkeys, magic link, step-up MFA, delegated authorization.
Consent & delegation	Captures explicit permission for an AI assistant to act on the customer's behalf.	Consent receipt, scope, duration, revocation, transaction-specific approvals.
Policy & authorization engine	Determines what the agent can read or do for this customer, intent, channel, and risk level.	OPA / Rego, Cedar, custom rules engine, fraud / risk scoring.
Knowledge fabric / RAG	Provides accurate, approved responses from enterprise content and operational knowledge.	Vector DB, search index, knowledge graph, CMS ingestion, policy versioning.
Tool / action registry	Defines tools, schemas, preconditions, required permissions, and execution behavior.	Tool catalog with JSON schemas, MCP tool definitions, OpenAPI specs.
Workflow orchestrator	Coordinates multi-step service processes across systems.	Temporal, Camunda, Durable Functions, Step Functions, Logic Apps.
Transaction execution	Calls backend APIs with least-privilege credentials and consistent error handling.	Microservices, API wrappers, ERP / CRM connectors, event-driven queues.
Audit & observability	Records each request, retrieval, answer, tool call, approval, transaction, and exception.	SIEM, log lake, OpenTelemetry, dashboards, QA review, compliance reports.

Vendor integration patterns

Vendor / channel	Best current integration pattern	Notes
OpenAI / ChatGPT	ChatGPT app or API integration connecting to the hub through MCP / tools; expose approved tools via a remote MCP server.	Good fit for customer-facing app experiences and structured tool calls. Needs careful schema, auth, consent, and review.
Anthropic / Claude	Expose enterprise capabilities through remote MCP servers or Claude tool-use patterns.	MCP is especially relevant — Anthropic originated and supports the protocol ecosystem.
Google / Gemini	Use Gemini function calling to translate natural language into structured calls; Vertex / Agent Platform for enterprise.	Strong where the enterprise controls the Gemini runtime and backend tool execution.
Microsoft Copilot	Copilot Studio agents, Power Platform connectors, M365 Agents SDK, and / or MCP where supported.	Strong fit for Microsoft-centric enterprises and authenticated agent workflows.
Future AI agents	Abstract vendor differences behind the hub and expose stable tool / resource contracts.	Avoid one-off vendor plumbing wherever possible.

Transaction flow: a refund request

1Customer asks their AI assistant: “Get me a refund for my damaged order.”
2AI vendor routes the request to the enterprise channel endpoint / app / MCP tool.
3Orchestration hub classifies intent as refund_request.
4Identity broker authenticates the customer and confirms account ownership.
5Consent service captures scope: “Allow this AI to request a refund for order #12345.”
6Policy engine checks refund rules, order state, amount, fraud signals, and eligibility.
7Tool registry selects inspect_order, validate_refund_eligibility, and create_refund.
8Workflow orchestrator calls OMS / payment APIs through the transaction execution service.
9Audit service logs request, policy version, tool calls, data accessed, approvals, and results.

API examples

01 GET  /customer/v1/orders/{orderId}/status
02 POST /customer/v1/refunds/eligibility
03 POST /customer/v1/refunds
04 POST /customer/v1/cases
05 POST /customer/v1/delivery/reschedule
06 POST /customer/v1/claims
07 POST /customer/v1/loyalty/adjustment

Want this mapped to your stack?

We translate this reference architecture into your CRM, OMS, payments, loyalty, and case systems — with the right vendor adapters and controls.

Start the discussion →