A Multi-Agent, AI-Powered Content Processing Workflow: A Conceptual Overview

The Problem: Manual Updates Are Expensive

Product descriptions are essential to e-commerce success. They often follow a consistent structure, whereby businesses [typically] reuse wording/style/etc across their inventory. Poor quality descriptions, whether due to outdated formatting, missing information, broken HTML or inconsistent formatting, directly impact conversion rates and customer satisfaction.

Essential questions that reveal content quality issues include:

• On a scale of 1 to 5, rate how seamless your experience was with our product.
• Could you find the information you were looking for quickly?

The lower the responses, the lower quality experience customers have and the less likely they are to return (e.g., lower retention, conversion, repeat customers, referrals and, often, higher bounce rates). If businesses already use structure wording for descriptions and many of these details are imbedded in tables, is there a way to curate/modify this information? How would a manual approach compare to one that is semi-automated?

A Hypothetical Manual Breakdown

For a skilled employee earning $31.75/hour, the [hypothetical] manual process may look like the following:

• Locate & Review: 5-10 minutes to navigate to and read a single product page
• Research: 15-60 minutes researching similar products (depending on complexity/recency of update)
• Revisions: 10-90 minutes updating the description text and associated HTML code

This translates to 30-160 minutes per product. With 1,000 products to update, you’re looking at 500-2,666 hours of work.

Why Simple AI Solutions Fall Short

In 2025, the typical AI-assisted approach still involves significant manual overhead:

1. Navigate to each product page
2. Copy the description
3. Navigate to an AI model website
4. Copy/paste prompt and product description
5. Wait for output, copy the output
6. Return to product page and paste results

This process would still require 100-150 hours of manual effort, costing $3,175-$4,763. While superior to fully manual updates (reducing costs by 80% to 90%), this approach remains inefficient and error-prone.

Solution: A Two-Stage AI Workflow

A dual-validation approach guards against several failures of genAI models:

• Hallucinations: Using a multi-step validation workflow increases chance of catching fabricated information
• Formatting: Second pass ensures proper HTML/markdown standards
• Content preservation: Quality metrics prevent content drift
• Error Recovery: Stage 2 can fix Stage 1 mistakes without starting over

Stage 1: Creation

1a: Content Generation

• Specify user and system prompts tailored to content length and modification requests
• Ingest original description/HTML from products and description metafields
• Instruct AI model to make explicit revisions (cleanup, format enhance)
• Use different prompt templates for different product categories

1b: Quality Checks

• Track word/sentence structure changes and content preservation
• Flag broken images/videos and processing errors

Stage 2: Validation

2a: Validation-Focused Processing

• Specify validation-focused user and system prompts
• Ingest LLM-generated HTML from Stage 1
• Use a different model family to evaluate and refine content
• Different model family reduces systematic bias and catches model-specific “blind spots”

2b: Comprehensive Quality Assessment

• Compare original, Stage 1, and Stage 2 outputs
• Track changes made and validation decisions for process improvement

Alternative Options

High-Volume Generator/Evaluation

For high-value products or when conversion rate optimization is critical:

• Generator Model: Produces 10+ candidate outputs for a given input
• Evaluator Model: Scores/ranks all outputs and selects the best one(s)
• Ensemble Voting: Multiple models generate content; majority consensus wins
• Human-in-the-Loop: AI handles bulk processing, humans review flagged items (e.g. content preservation < .50)
• KPI Feedback Loops: Track performance metrics to continuously improve prompts

Cost-Benefit Analysis In Toy Example

When to Use Each Approach

• Single-stage: Bulk cleanup of well-structured content with clear specifications available in tables (for example, concise metadata maintenance)
• Two-stage: valuable products, inconsistent source data or quality-critical applications (for example, careful refinement of product information)
• Generator/Evaluator: Same as former, but conversion optimization is paramount and accuracy more valuable than cost.

Conclusion

Processing costs can be extremely prohibitive for some businesses. However, quality and freshness is valuable to conversion and growth in e-commerce. Investing in AI workflows can reduce costs, opening the doors to iteratively updating content and descriptions with quality information that customers will appreciate.

this example focuses on product descriptions, the same workflow is broadly applicable in other domains, especially in e-commerce and beyond:

• Category and Collection Pages
• SEO and Landing Pages
• Updating documentation, customer information that may appear on different pages.
• Automated Q/As and testimonials based on collection of emails.
• Refining blogs with updated details.
• Modifying and creating images using diffusion models

In other businesses:

• Property listings (e.g., MLS, rental ads) that require consistent formatting, accurate details, and engaging language.
• Contracts, terms of service, privacy policies and legal templates
• Analyst reports, market summaries and portfolio updates
• Material requirements, construction standards and project scope documents
• Pull relevant case details from your system and populate standardized forms, reducing manual writing time by 80-90% (keep human-in-the-loop for quick review)
• Management systems: When tasks come due, the system automatically generates draft documents based on available details and templates
• Case filings that are boilerplate and necessary but are rarely considered in proceedings.