# eCommerce AI Implementation Failure Modes

Companion to **Why most eCommerce AI projects fail before the AI model is even involved**.

The model is not the system. Score each failure mode as **Present / Partial / Absent** on **your** stack. This is a diagnostic, not a client post-mortem.

```text
The AI Model  ≠  The Entire AI System

Production agent =
  Business Process
  + Business Data
  + Tools
  + Integration
  + Evaluation
  + Governance
```

## Failure modes (before the model)

| Mode | What it looks like | Why the model cannot save it | Fix |
| ---- | ------------------ | ---------------------------- | --- |
| **Fragmented data** | OMS order id ≠ Shopify id ≠ WMS id | Evidence cannot be joined | Canonical map; refuse “unknown” |
| **Poor product data** | Attributes in HTML; duplicate GTINs | Recommendations invent specs | Catalog contract ([`ai-ready-catalog-contract.md`](ai-ready-catalog-contract.md)) |
| **Disconnected systems** | Support cannot see ATP; finance cannot see ads | Agent answers from one silo | Named tools per system, one session |
| **Missing APIs** | CSV export, email to the 3PL | No tool = scrape or guess | OpenAPI/MCP; no HTML-as-catalog |
| **Broken workflows** | Refunds already fail in the OMS | Agent wraps a broken path | Repair the path; then wrap |
| **Unclear ownership** | IT owns the bot; ops owns the refund | Nobody owns evals or blast radius | Named business + eng owners |
| **No evaluation strategy** | Demo on three happy tickets | Drift after week two | Goldens + must-not-write |
| **No escalation** | Prompt: “ask a human if unsure” | Model will not be unsure | HITL queue ([`hitl-approval-architecture.md`](hitl-approval-architecture.md)) |
| **Weak governance** | Admin API key in the chatbot | Irreversible writes | Gateway + Cedar; Browser off |

## Pre-model gate (must all be true)

- [ ] First workflow named in one sentence
- [ ] Read tools named (not `executeQuery`)
- [ ] IDs for that workflow can be joined
- [ ] Owner named for evals + cost
- [ ] HITL path exists for any write
- [ ] Agents Classic is **not** the net-new host (after 30 Jul 2026)

If any box is empty, do not pick a foundation model this week.

## Related

- Readiness: [`ai-agent-readiness-checklist.md`](ai-agent-readiness-checklist.md)
- Agent vs workflow: [`hybrid-automation-decision-matrix.md`](hybrid-automation-decision-matrix.md)
- Security: [`ecommerce-agent-security-checklist.md`](ecommerce-agent-security-checklist.md)