AI rules vs manual rules: when to use each in your feed management

Feed management has two flavours of automation: manual rules (deterministic transformations based on conditions) and AI rules (generative transformations using large language models). Both are valuable. Neither replaces the other. The skill is knowing which to reach for.

Manual rules: fast, free, deterministic

Manual rules work like Excel formulas applied to your entire catalogue on every feed run. They're deterministic — the same input always produces the same output. They're free to run. They're instant.

Manual rules are the right tool for:

• Field concatenation: Combine brand + title + colour + size into a single optimised title string
• Value substitution: Replace "in_stock" with "In Stock" for display compatibility
• Conditional logic: If price > 100, set custom_label_0 = "premium"
• Text cleaning: Remove specific HTML tags, strip certain phrases ("Sale:", "New:")
• Category mapping: If product_type = "Jackets", set google_product_category = "1604"
• Default values: If brand is empty, set brand = "Own Brand"
• Filtering: Exclude products where availability = "out_of_stock" and price < 5

These transformations are unambiguous. You know exactly what they'll produce. They run in milliseconds. There's no cost per product. For 80% of feed management tasks, manual rules are the right tool.

AI rules: slower, cost-per-product, generative

AI rules use a large language model (Trajekt uses Claude) to generate new content or make nuanced decisions that pattern matching can't handle. They're generative — the same input might produce slightly different outputs, and the output requires human review to validate initially.

AI rules are the right tool for:

• Description generation: Rewriting thin, HTML-laden, or marketing-speak descriptions into clean, factual, keyword-rich copy
• Attribute extraction: Identifying colour, material, or style from unstructured product names and descriptions when these fields aren't in your source data
• Category disambiguation: When your product falls into multiple potential Google taxonomy categories and you need intelligent selection
• Channel-specific copy: Writing different descriptions for Google Shopping (keyword-focused) vs ChatGPT (conversational, factual) from the same source product
• Missing data inference: Intelligently filling in fields where the information is implied by other fields but not explicitly stated

These tasks require understanding context, making judgements, and generating original text. No manual rule can do this reliably at scale.

Cost and caching

AI rules cost money — typically around £0.001 per product per rule. For a 10,000-product catalogue, a full description rewrite costs approximately £10. For most merchants, this is excellent value compared to paying a copywriter to rewrite descriptions manually.

The key economics tool is caching. Trajekt caches AI rule results against a hash of the product data. If a product hasn't changed between feed runs, the cached result is used and no credits are consumed. In practice, after the initial processing run, ongoing AI costs are minimal — only new or changed products consume credits.

The right combination

The most effective feed management combines both types in sequence:

1. Manual rules first: Clean the source data, extract fields that can be extracted mechanically, set up title structure, apply custom labels. This is fast, free, and handles the majority of transformation work.
2. AI rules after: Apply AI enrichment to the fields that benefit from it — description quality, attribute extraction for products with unstructured data, channel-specific copy generation. AI rules run on the output of your manual rules, so the AI is working with clean, structured input.

This staged approach maximises the value of AI credits and minimises unnecessary processing.

Targeting AI rules wisely

Not all products benefit equally from AI enrichment. Target your AI rules at:

• Products with descriptions under 100 characters
• Products with HTML-only descriptions
• Products where description equals title (a common data quality issue)
• New products (added to the catalogue in the last 7 days)
• High-value products where the ROI on enrichment is clearest

Excluding products that don't need enrichment from your AI rules saves credits and focuses the AI on work that will have measurable impact.

Validating AI output

The most important practice with AI rules: validate a sample of outputs before applying at full scale. Review 20-30 AI-generated descriptions across different product categories to confirm:

• The AI is using only information from the source product data (no hallucination)
• The output matches your brand's tone and style
• The descriptions are factual and specific, not vague or generic
• No legal or compliance issues in the generated content

Trajekt's AI rules are constrained to only use data present in your product record — this significantly reduces hallucination risk. But human review of initial output is always best practice before rolling out to your full catalogue.