AI automation for VTEX and Shopify: workflows that actually work

Most AI automation advice for e-commerce is either too theoretical (AI will revolutionize everything) or too trivial (use ChatGPT to write emails). This guide covers specific, implementable workflows that save real hours every week on VTEX and Shopify operations. Each workflow includes the trigger, the AI processing step, the output action, the tools involved and the actual cost. These are workflows tested in production, not ideas from a brainstorming session.

Workflow 1: bulk product descriptions via Claude API

Problem: 500 new products added monthly from suppliers with only technical specs, no customer-facing descriptions. Manual writing takes 15-20 minutes per product. That is 125-167 hours per month. Trigger: new product created in VTEX (via webhook on catalog change) or Shopify (via webhook on products/create). A Cloudflare Worker receives the webhook payload with product ID. Processing: the Worker fetches full product data from the platform API (title, category, attributes, images). It calls the Claude API with a prompt template specific to the product category. The prompt includes brand voice guidelines, SEO keywords for the category and 2-3 example descriptions as few-shot examples. Claude returns a JSON object with short_description, long_description, meta_title, meta_description and 5 bullet points. Output: for VTEX, the Worker calls the Catalog API to update the product. For Shopify, it calls the Admin API to update the product body_html and metafields. Descriptions are saved as drafts (unpublished) for human review. A Slack notification alerts the content team that new descriptions are ready for review. Cost: Claude API approximately $0.01-0.05 per product. For 500 products/month: $5-25. Cloudflare Workers free tier. Time saved: 100-150 hours per month. The content team shifts from writing to reviewing, which is 5-10x faster.

Workflow 2: auto-categorization with Vision models

Problem: suppliers send product images without consistent categorization. Manual categorization of 200 products per week takes 8-10 hours. Misclassification hurts search, recommendations and marketplace feeds. Trigger: batch processing on a weekly schedule via Cloudflare Workers Cron Trigger. The Worker queries Supabase for products flagged as uncategorized or newly imported. Processing: for each product, the Worker sends the product image and basic metadata to Claude with vision. The prompt includes your category taxonomy (hierarchical list of all categories and subcategories) and asks Claude to: assign primary and secondary categories, suggest relevant tags, flag if the image quality is insufficient for the storefront. Claude returns structured JSON with category IDs matching your existing taxonomy. Output: categories and tags are applied via platform API (VTEX Catalog API or Shopify GraphQL Admin API). Products with low-confidence classifications (Claude's stated confidence below 80%) are routed to a human review queue in Supabase. A weekly summary report is sent to the catalog team. Cost: Claude with vision approximately $0.01-0.03 per image. For 200 products/week: $8-24/month. Accuracy: typically 85-95% for well-defined taxonomies. The remaining 5-15% are edge cases that benefit from human judgment. Net time saved: 6-8 hours per week.

Workflow 3: dynamic email copy with Klaviyo

Problem: email campaigns use the same generic copy for all segments. Personalization is limited to first name merge tags. Writing segment-specific copy for 8 email flows times 4 segments equals 32 variants manually is not feasible. Trigger: Klaviyo flow triggers (abandoned cart, post-purchase, browse abandonment, win-back, review request). Instead of static templates, the email template calls an API endpoint for dynamic content. Processing: a Cloudflare Worker receives the request with customer data (segment, purchase history, browsed products, cart contents). It constructs a prompt for Claude that includes the customer context, email type and brand voice guidelines. Claude generates: subject line (2 variants for A/B testing), preview text, hero copy, product recommendation reasoning and CTA text. The copy is tailored to the segment: a first-time buyer gets educational content, a repeat customer gets loyalty messaging, an abandoned cart gets urgency appropriate to the product price point. Output: the Worker returns the generated copy to Klaviyo via API, which populates the email template and sends. All generated copy is logged in Supabase for performance tracking. After 30 days, the system analyzes which AI-generated copy variations performed best (open rate, click rate, conversion) and feeds this back into the prompt for continuous improvement. Cost: Claude API approximately $0.005-0.02 per email. At 10,000 emails/month: $50-200. Revenue impact: segment-specific copy typically improves email revenue by 15-30% compared to generic templates. Tools: Klaviyo (plans from $20/month), Claude API, Cloudflare Workers, Supabase for logging.

Workflow 4: competitor monitoring with scraping + LLM + Supabase

Problem: manually checking competitor prices, product launches and promotions takes 5-10 hours per week and the information is outdated by the time it is compiled. Trigger: daily Cloudflare Workers Cron Trigger runs at 6 AM. Processing: the Worker fetches competitor product pages using a scraping API (ScrapingBee, Bright Data or a custom scraper). Raw HTML is sent to Claude with instructions to extract: product prices, promotional badges, shipping offers, out-of-stock indicators and any new products since last check. Claude returns structured JSON. Data is stored in Supabase tables: competitor_prices (competitor, product_sku, price, promotional_price, scraped_at), competitor_products (new products detected), competitor_promotions (active promotions and end dates). Analysis: a separate Worker runs after scraping completes. It queries Supabase for price comparisons: your products versus competitors. It calls Claude with the comparison data and asks for: products where you are more than 10% more expensive (pricing opportunity), products where competitors are out of stock (advertising opportunity), new competitor products that you do not carry (assortment gap), and competitor promotions that you should counter. Output: a daily Slack digest with prioritized insights and recommended actions. Product managers can click through to take action. Historical data in Supabase enables trend analysis: is a competitor consistently lowering prices in a category? Are they expanding into a new vertical? Cost: scraping API $50-100/month (depends on volume and anti-bot difficulty). Claude API approximately $30-50/month for analysis. Supabase free tier. Total: $80-150/month for intelligence that would require a dedicated analyst.

Workflow 5: automated content QA

Problem: product pages go live with missing descriptions, broken images, incorrect prices, missing SEO metadata and policy violations. Manual QA of new content is slow and inconsistent. Trigger: webhook on product publish (VTEX or Shopify) or scheduled daily scan of recently updated products. Processing: the Worker fetches the full product page data from the platform API and runs it through a multi-step QA pipeline. Step 1 (data completeness): check for missing fields (description, images, price, weight, category, meta title, meta description). Flag any product missing required fields. Step 2 (content quality): send description and meta content to Claude with QA instructions: check for grammar errors, brand voice consistency, SEO best practices (title length, description length, keyword presence), prohibited claims (health claims, unsubstantiated superlatives) and duplicate content (compare against recent descriptions stored in Supabase using text similarity). Step 3 (image validation): send product images to Claude with vision to check for: minimum resolution, white background, no watermarks, correct aspect ratio, matching product title (the image should show what the title describes). Step 4 (pricing sanity check): compare price against category averages stored in Supabase. Flag prices that are more than 2 standard deviations from the category mean (likely a decimal error or wrong product). Output: QA results are stored in Supabase with severity levels (critical, warning, suggestion). Critical issues (missing price, no images) trigger immediate Slack alerts. Warnings (short description, missing meta) go into a daily digest. A Next.js dashboard shows QA status across the catalog with filters by severity, category and date. Cost: Claude API approximately $0.02-0.05 per product QA check. For 100 new/updated products per week: $8-20/month. Time saved: eliminates 10-15 hours per week of manual QA and catches issues that human reviewers miss.

Costs per workflow and total investment

Summary of monthly costs for a mid-size store (1,000-10,000 SKUs, 500 new products/month): Workflow 1 (descriptions): $5-25 Claude API. Workflow 2 (categorization): $8-24 Claude API. Workflow 3 (email copy): $50-200 Claude API plus $20+ Klaviyo. Workflow 4 (competitor monitoring): $80-150 scraping plus Claude API. Workflow 5 (content QA): $8-20 Claude API. Infrastructure: Cloudflare Workers free tier plus Supabase free tier. Total: $170-440/month for all five workflows. Development investment: 60-120 hours for all workflows (a developer familiar with the stack can build one workflow per week). Monthly maintenance: 4-8 hours for prompt tuning, error handling and new edge cases. ROI summary: these workflows replace approximately 200-300 hours of monthly manual work. At an average operator cost of R$25-40/hour, that is R$5,000-12,000/month in labor savings. The automation cost is R$850-2,200/month. Net saving: R$4,000-10,000/month.

Prioritization: where to start

Not all workflows deliver equal value. Start with the one that addresses your biggest pain point. If your catalog has sparse descriptions and you are losing SEO rankings: start with Workflow 1 (bulk descriptions). Impact is immediate and measurable via organic traffic growth within 4-8 weeks. If your team spends hours on manual QA and bad content reaches production: start with Workflow 5 (automated QA). This prevents revenue-impacting errors and reduces rework. If you compete on price and lack market intelligence: start with Workflow 4 (competitor monitoring). The insights drive pricing decisions that directly impact margins. If your email program underperforms: start with Workflow 3 (dynamic email copy). Revenue impact is measurable within the first campaign cycle. Build one workflow at a time. Get it stable in production for 2-3 weeks before starting the next. Each workflow shares infrastructure (Cloudflare Workers, Supabase, Claude API) so the second and third workflows are faster to build than the first. The common mistake is trying to automate everything at once, ending up with five half-built workflows instead of one fully operational one.