Intelligence Commerce Suite (ICS)

Overview

This release delivers the first production-capable version of the Intelligence Commerce Suite (ICS) — a multi-tenant AI platform that lets commerce businesses build, run, and manage AI agents that work directly with their commerce systems (Shopify, Salesforce, Equinox).

This release covers seven capability areas: Agent Studio, Knowledge, Prediction, Language, Connectors, Organisation & RBAC, and Cost Tracking.

Functional Changes – What’s New for Users

Commerce Agent Studio

Users can now create, configure, publish, and monitor AI agents entirely through the Agent Studio interface.

Agent Authoring

• Create a new agent from scratch or subscribe to a ready-made agent from the Agent Store with one click.
• Configure the agent’s language model, instructions (system prompt), and which business areas it can access (e.g., orders, products, pricing).
• Test the agent interactively in the Playground before publishing — uses real data, real connectors.
• Publish an immutable version when ready. Future updates create a new version without overwriting history.

Human-in-the-Loop (HITL) Approvals

• Agents can be configured to pause and wait for a supervisor’s approval before executing sensitive actions (e.g., processing a refund above a threshold, updating prices).
• The run report page shows an “Awaiting approval” banner with full context so the supervisor can approve or reject without leaving the platform.
• Once approved, the agent resumes exactly where it left off — no data is lost and no steps are re-executed.
• A paused run can wait for up to 7 days at zero cost before timing out.

Triggers – Automated Execution

• Schedule trigger: Set a cron schedule (e.g., every Monday at 8 AM) to run an agent automatically using the one-click “Schedule this agent” shortcut.
• Event trigger: Connect an agent to a commerce marketplace event (e.g., “a return request was submitted”) using the “Listen to events” shortcut. The agent starts automatically whenever that event fires — no manual intervention.
• Manual / Chat trigger: Run an agent on-demand or interact with it through a conversational chat interface.

Run Monitoring

• The run report shows a live step-by-step execution trace, including every tool call made (with full request and response detail), token usage, and the final output.
• Run lifecycle: pending → running → succeeded | failed | cancelled | interrupted (interrupted = waiting for human approval)

Agent Store & Sharing

• Platform admins publish out-of-the-box (OOB) agents available to all tenants.
• Tenant admins subscribe to OOB agents or clone them to create customised versions.
• Agents can be shared across tenants within the same organisation.

Knowledge

Users can ingest organisational documents and architecture content into the platform’s knowledge base, which agents and chat interfaces can then query.

• Upload documents (PDF, Word, PowerPoint, images, YAML, JSON, OpenAPI specs) — the platform extracts structured knowledge automatically.
• Human review inbox: Extracted content goes into a review queue. A knowledge admin can review, edit, and approve before it becomes queryable.
• Auto-approval for structured files (JSON, YAML, OpenAPI specs, architecture images) — these bypass the review inbox and are written directly to the knowledge graph.
• Chat with your knowledge base: Ask natural language questions and get answers grounded in the approved documents.
• Three specialised tracks:
  • Institutional Brain — organisational policies, decisions, and documents
  • Enterprise Architect — architecture diagrams, API specs, system documentation
  • Domain Expert — commerce-specialised AI models fine-tuned on your own domain documents
• Coverage benchmarking: Run a test suite to identify gaps in your knowledge base.
• Cost tracking: See how much LLM spend each knowledge capability is consuming, broken down by model and document type.

Prediction

Users can deploy pre-built ML models against their own commerce data or build custom models through a guided project workflow.

• 17 out-of-the-box models covering: demand forecasting, customer churn, customer lifetime value, next-purchase prediction, product lifecycle classification, return rate prediction, search ranking, customer segmentation, replenishment trigger, and more.
• Self-service project creation: Connect your data source, pick a model type, and the platform handles training, evaluation, and deployment.
• Monitor endpoint health: Dashboard showing the live status of every deployed model.
• Cost analytics: See prediction spend by project, model, run type (inference/training/batch), and dimension (user/business/store). Set a monthly org budget with alert thresholds.
• Idle endpoint reaper: The platform automatically detects and reclaims idle endpoints to reduce costs.

Connectors

Connectors let agents and workflows interact with external commerce platforms without exposing credentials or vendor-specific APIs to end users.

• Shopify: Read and update products, pricing, inventory, collections, customers, and orders through the Shopify Admin API.
• Salesforce: Read, create, update, and delete SObjects (products, orders, customers, pricing) through Salesforce hosted MCP servers.
• Equinox: Connect to the Equinox cloud admin MCP and subscribe to marketplace events for event-driven automation.
• Custom connectors: Add your own MCP server or REST/GraphQL API as a connector through a self-service registration flow.
• Credential security: Connector credentials (API tokens, OAuth secrets) are stored in a Vault and never exposed to agents, users, or logs. Agents interact with domain verbs (product.update, order.read) — the platform resolves which vendor and injects the credentials at runtime.
• OAuth keep-alive: A daily background job automatically refreshes OAuth tokens for connected platforms so scheduled and event-triggered agents don’t fail on expired credentials.

Organisation & User Management (RBAC)

Platform administrators can provision and manage organisations, tenants, and users through a structured hierarchy.

• Four user tiers: Instance Admin → Org Admin → Tenant Admin → Tenant User (day-to-day operator with domain-specific access).
• Invite users: A single UI action creates the user identity, user record, and role grant in one step.
• Action-level permissions: Finer-grained control over who can create, read, update, delete, execute, or chat with agents and workflows.
• Audit trail: Every significant action is recorded in a tamper-evident audit log with actor, timestamp, and target resource.

Cost Dashboard

• Platform-wide token cost view (instance admin): See LLM spend across all organisations and tenants.
• Prediction cost analytics (org admin): KPI strip, daily trends, breakdown by model/project/run type.
• Knowledge cost analytics (org admin): Breakdown by capability (Institutional Brain, Enterprise Architect, Domain Expert), model, and touchpoint.
• Budget controls: Set a monthly org budget and alert threshold. Receive alerts when spend approaches the limit.

API Changes

The following new and updated API endpoints are available in this release. All backend endpoints are prefixed with /intelligence-platform/<service>.

Agent Studio — /intelligence-platform/agent-studio

Connectors — /intelligence-platform/connectors

Tenancy — /intelligence-platform/tenancy

Knowledge — /intelligence-platform/knowledge

Technical Stack

Frontend

Backend

Deployment

Upgrade Notes

• Existing tenant roles do not automatically receive the new action-level permissions introduced in this release. Tenant admin roles for existing tenants must be updated manually through the role grant API to enable fine-grained access control for agents and workflows.
• Connector clients using X-Org-Id or X-User-Id headers for local testing must be updated to use a valid Cognito id_token bearer token — these headers are no longer accepted by the Connector service.
• Agent definition staging: PUT /v1/resources/{id}/agent-definition validates model handle, connector domain bindings, and tool bindings against what is provisioned for the requesting tenant before staging. Staging a workflow-type resource or a non-draft resource returns 422.
• Publish lifecycle: POST /v1/resources/{id}:publish freezes an immutable version. Publishing an already-published resource with an open working draft at the next version number is now supported.
• Clone: POST /v1/resources/{id}:clone clones a published resource into a new draft with lineage recorded. Cloning an unpublished resource returns 422.
• Org-wide resource list: GET /v1/resources?org_wide=true allows org admin callers to enumerate all tenant resources in their org. Requires org admin role.
• Delete with adoption guard: DELETE /v1/resources/{id} hard-deletes a resource. A published resource with active adoptions returns 409 with the remaining adoptions count — revoke adoptions first.

Revision History
2026-08-12 | JP – Created the page and added the content.