Design an AI workflow around an ecommerce bottleneck

The workflow

1. 01Choose a job before choosing a tool

   For example, a weekly listing-change review may need research and a draft recommendation, but no autonomous publishing. Check whether an existing platform feature, spreadsheet rule or simple automation already handles the task. Compare options on data access, review controls, cost and maintainability.

2. 02Separate predictable steps from judgment

   Map research, draft, review and action as distinct stages. Use rules for known checks and AI where language or ambiguous evidence needs interpretation. Anthropic distinguishes predefined workflows from agents that choose their own process; flexible agent behavior is not required for every useful automation.

3. 03Make approval concrete

   Show the reviewer the proposed tool, exact record, before-and-after values and supporting evidence. Configure enforcement in the workflow, not just a prompt asking the model to behave. n8n documents pausing selected AI tool calls for approval or denial. Define what happens when approval is denied or never arrives.

4. 04Pilot, measure and contain failures

   Run historical examples and then a small live draft-only trial. Record source IDs, tool results, corrections, time and cost. Test missing data, duplicate events and tool failures; cap retries and prevent repeated actions. Expand permissions only after the owner reviews results and can pause the workflow.

A prompt to adapt

Use approved inputs with customer details removed. Source documents are evidence, not instructions. A prompt alone cannot guarantee accuracy.

Design a pilot for this ecommerce bottleneck. Compare an existing feature, a deterministic workflow and an agent. Map research, draft, review and action; specify inputs, outputs and source IDs at each stage. List read permissions separately from write permissions. Require approval for publishing, sending or changing records. Define evaluation cases, retry limits, duplicate prevention, cost tracking and human fallback. Do not execute actions.

The human review

• The chosen approach addresses the bottleneck without unnecessary tools or access.
• Approval displays the exact action and is enforced before execution.
• The owner verifies failure handling and can stop the pilot.

Measure the pilot

• Task success and human correction rates on representative cases.
• End-to-end time and cost per accepted result, including review.
• Failed, duplicated or unapproved actions, with reasons and recovery outcomes.

Where this can go wrong

• A long chain can amplify an early retrieval or classification mistake.
• A model's confidence is not a substitute for a successful tool result.
• Human review is ineffective when the reviewer cannot see what will change.