Canva's Build-Time AI Expansion Pattern
TRIGGER
Classification systems using keyword matching achieve high coverage on common cases but fail on a long tail of edge cases where keywords don't directly align—yet using AI for all classification is expensive or slow at scale.
APPROACH
For their DesignDNA year-in-review campaign, Canva's Personalisation & Engagement team matched users to one of 7 design trends based on template usage. Input: style and theme keywords from templates each user interacted with, plus trend definitions from the Content Creative team. Output: each user scored and assigned to their highest-matching trend. Keyword matching covered 95% of users. For the remaining 5%, they curated a list of commonly-appearing template keywords that didn't match any trend, then used generative AI to expand each trend's keyword set with contextually relevant terms from this curated list—a one-time build step rather than per-user inference. This raised coverage from 95% to 99%, with the remaining 1% having insufficient template usage data. The campaign generated 95 million unique personalized DesignDNAs.
PATTERN
“Rules handle the head at zero marginal cost; AI expands them once at build time to catch the tail. The trap is choosing between "rules for everything" or "AI for everything" when the hybrid wins.”
✓ WORKS WHEN
- Rule-based approach achieves >90% coverage—edge cases are a minority
- Edge case vocabulary is bounded and can be curated upfront
- AI expansion can happen at build time rather than query time
- Categories are stable enough that expanded rules don't need frequent updates
✗ FAILS WHEN
- Rule-based approach covers <80% of cases—AI is needed for the majority
- Edge cases require real-time context that can't be captured in static keyword expansion
- Classification categories change frequently, requiring constant re-expansion
- The matching problem is semantic rather than lexical (keyword expansion won't help)