Run agents. Review approvals. Monitor catalog health.
The Agent Control Center is the PIM admin hub for Agentic PIM — where teams enable specialist agents, triage insights, review approval proposals, and track catalog quality from agent findings (not attribute fill rates alone).
Open it from the PIM sidebar → Agent Control Center (requires ai_mgmt permissions).
The governed loop
Specialist agents propose — humans approve — approved changes apply to PIM — syndication publishes channel-ready output. Rejections and channel feedback map back as new findings.
| Step | What happens | Where in ACC |
|---|---|---|
| 1. Run agents | Manual, scheduled, event-triggered, or multi-step workflow runs | Agents · Workflows |
| 2. Review findings | SEO gaps, taxonomy conflicts, compliance issues surface as insights | Insights |
| 3. Approve writes | Every proposed attribute or content change waits in the queue | Approvals |
| 4. Monitor quality | Rollup health score from open insights + pending approvals | Catalog quality |
Catalog Quality measures open agent insights and pending approvals — not PIM attribute fill rate. Use completeness as a publish gate; use ACC for operational catalog intelligence.
Information architecture
The Control Center is organized into four areas in the PIM sidebar:
Page guide
In PIM, Agent Control Center pages live under /agent-center/…:
| Page | PIM path | What you do here |
|---|---|---|
| Agents | /agent-center/agents | Enable agents, configure policies, view per-agent metrics |
| Approvals | /agent-center/approvals | Review diffs, approve/reject/bulk approve proposed writes |
| Governance | /agent-center/policies | Toggle platform governance policies, audit runs |
| Activity | /agent-center/activity | Filter runs by agent/status; open live step timeline |
| Insights | /agent-center/insights | Triage findings by severity/type; bulk resolve |
| Catalog Quality | /agent-center/catalog-quality | Headline health + eight quality dimensions |
| Workflows | /agent-center/workflows | Create and monitor multi-step agent workflows |
| Orchestration | /agent-center/orchestration | Active workflows and run boards |
| Observability | /agent-center/observability | Aggregates: workflows, failures, token usage |
| Cost & Usage | /agent-center/cost-usage | Platform AI credits and model attribution |
| Business Impact | /agent-center/business-impact | Time-saved estimates from approval outcomes |
Legacy URLs (e.g. /agent-center/runs, /agent-center/missions) redirect to the canonical paths above.
Agents page
The Agents page lists all ten specialist agent types with enable/disable controls, run metrics, and a detail drawer:
| Drawer tab | Purpose |
|---|---|
| Overview | Recent activity, last run, links to Activity |
| Workflows | Prebuilt templates, custom goals, batch scope, parallel steps |
| Configuration | Auto-apply rules, confidence thresholds, blocked attributes |
| Permissions | Effective IAM grants (read-only) |
| Schedules | Cron-based recurring runs |
| Activity | Run history for this agent type |
Tenant-wide defaults and overrides are managed per agent. See the full catalog in Specialist agents.
Typical first actions
- Confirm default agents are enabled for your catalog (enrichment, SEO, taxonomy, compliance).
- Leave auto-apply off until your team trusts agent output.
- Trigger a manual run on a pilot category or product set.
Approvals
Every agent-proposed catalog write lands in Approvals before it applies.
- List view — server-paginated table with filters by agent, status, and product
- Detail view — proposed diff, approve, reject, or retry apply after failure
- Bulk actions — approve or reject many proposals in one session (bulk approve applies the same governed writes as single approve)
Approval types include attribute updates, category changes, content generation, translations, bulk actions, and status changes.
Auto-apply is available only for select agent types (e.g. duplicate detection) when tenant policy explicitly allows it. Default posture is review first.
Insights
The Insights dashboard surfaces open findings from agent runs:
- Priority / by-type / all tabs for triage
- Filter by severity (
critical,warning,info), insight type, product, or search - Bulk resolve when issues are fixed outside the agent loop
Common insight types: seo_weakness, taxonomy_conflict, compliance_gap, duplicate_detected, completeness_gap, marketplace_readiness, translation_gap, and more.
Dimension cards on Catalog Quality deep-link here with the right insight_type filter when a dimension maps to a single type.
Catalog Quality
Catalog Quality rolls up open insights and pending approvals into a headline score and eight quality dimensions. It is a monitoring layer — not a separate AI agent.
| Measures | Does not measure |
|---|---|
| Count/severity of open insights from specialist agents | PIM attribute fill rate |
| Pending approvals (headline factor) | Whether every SKU was scanned |
| Recent failed runs (headline factor) | Channel export success rate |
Scoring caveats
- No open insights → dimensions may show 100% — meaning no known problems, not catalog is complete.
- Headline and dimension scores use different formulas; failed runs can lower the headline while dimensions stay at 100%.
Full explainer → Catalog quality — findings not fill rates.
Activity & observability
| Page | What you see |
|---|---|
| Activity | Paginated run list; run detail with live SSE step stream during execution |
| Observability | Active workflows, failed runs (30d), token aggregates |
| Cost & Usage | Platform AI credit balance and per-model usage estimates (when using platform-billed models) |
| Business Impact | Approval-outcome time-saved model — estimates only, with methodology banners in UI |
When platform AI credits are unallocated or exhausted, agent runs on platform-billed models block with a clear message — contact your administrator to allocate credits in Backoffice.
Governance
Governance loads platform policies with active/inactive toggles per policy_key. Metrics include active policy count, failed runs (30d), pending approvals, and active risk controls.
- Per-agent enablement, auto-apply, and blocked attributes remain under Agents → Configuration
- Audit tab shows completed and failed runs for governance review
Workflows & orchestration
| Page | Purpose |
|---|---|
| Workflows | Create and run multi-step agent workflows on a product, category, family, or catalog scope — with a live execution timeline |
| Orchestration | Live board of active workflows and agent runs for batch catalog operations |
CataZenta ships ten prebuilt workflows (enrichment pipelines, marketplace launch, import onboarding, localization, and more). Pick a template when you create a workflow, or describe a custom goal and let CataZenta plan the steps.
→ Full catalog with step-by-step agent order: Prebuilt agent workflows
Workflows can run steps in sequence or in parallel, and some templates skip or add a step based on findings (for example, compliance only when enrichment reports critical issues). When a step proposes catalog changes, the workflow pauses for approval and resumes after your team approves in the queue.
Permissions
Agent Control Center requires the ai_mgmt module:
| Action | Permission |
|---|---|
| View runs, insights, approvals | ai_mgmt:prompts:view |
| Trigger runs, approve/reject | ai_mgmt:prompts:create |
| View / edit agent policies | ai_mgmt:configs:view / ai_mgmt:configs:update |
Product workspace embeds agent triggers and insight cards on product detail pages — with links back into the Control Center.
Zen AI vs specialist agents
| Zen AI | Specialist agents (ACC) | |
|---|---|---|
| Interaction | Conversational chat in PIM | Autonomous runs, schedules, workflows |
| Output | Inline suggestions in chat | Structured insights + approval proposals |
| Writes | Manual apply or workflow gates | Approval queue → apply on approve |
| Best for | Ad-hoc questions, one-off edits | Continuous catalog operations at scale |
Both operate on the same governed catalog. See Zen AI and Agentic PIM overview.
MCP and the Control Center
External MCP clients (Claude, ChatGPT, Cursor) call documented catalog tools on your tenant — search, enrich, publish, analyze completeness.
In-platform agents add what MCP alone cannot persist:
- Insights with severity, type, and bulk resolve
- Approval queue with diffs and bulk approve
- Catalog Quality rollup across dimensions
→ MCP overview · MCP quickstart · AI agents (MCP clients)
Getting started in PIM
- Open Agent Control Center from the PIM sidebar (confirm
ai_mgmtaccess). - Review enabled agents on the Agents page — start with enrichment, SEO, taxonomy, compliance.
- Run a pilot — manual run on one category or import batch; watch Activity for live steps.
- Triage Insights — resolve or convert findings into approval proposals.
- Approve proposals in Approvals before any write goes live.
- Check Catalog Quality — monitor headline score and dimensions weekly.
- Wire MCP (optional) — same catalog for conversational ops in Claude or Cursor.
Need API access? → Quick start · First API call
GO DEEPER